klobucnik, jan / miersch, david / sievers, soenkejan klobucnik[a], david miersch[b], soenke sievers...
TRANSCRIPT
No. 13 / November 2015 Klobucnik, Jan / Miersch, David / Sievers, Soenke Predicting Early Warning Signals of Financial Distress: The Role of Accounting Volatility Measures
Predicting Early Warning Signals of Financial Distress: The Role of Accounting
Volatility Measures*
Jan Klobucnik[a], David Miersch[b], Soenke Sievers[c]
First version: March 1, 2013 This draft: November 18, 2015
Abstract This study proposes a simple accounting-based framework that allows to assess financial distress up to five years in advance for US listed firms over the period of 1980-2010. We jointly model financial distress by two of its key driving forces, (1) a declining profitability, modeled by independent stochastic processes, and (2) insufficient liquid assets, viewed as a minimum level of operating liquidity. The financial distress model incorporates important underlying (i) time-series (ii) industry-sector (iii) firm-level growth, and (iv) accounting volatility developments. Cross-sectional and longitudinal analyses show improvements in the discriminatory power and demonstrate incremental information content beyond state of the art accounting-based models. For example, we find a 11 percent (4 percent) higher ranking performance documented by the area under the receiver operating characteristic curve (AUROC) relative to the standard Altmanβs z-score (Ohlsonβs o-score) models. Consequently, this study might provide important ex ante warning signals for investors, regulators and practitioners. JEL classification: C63, C52, C53, G33, M41 Keywords: Financial distress prediction, probability of default, accounting information, stochastic processes, simulation Data Availability: Data used in this study are available from public sources identified in the paper. * Acknowledgments: The authors are grateful to Wayne Landsman, Gilad Livne, Mingyi Hung, Kirill Novoselov, Igor
Goncharov, Ken Peasnell, Steve Stubben, Florin Vasvari and Jon Tucker for their valuable comments. This paper has also benefited from the comments of seminar participants at the Hong Kong University of Science and Technology (2013), the Eighth Accounting Research Seminar in Basel (2013), the Frankfurt School of Finance & Management (2013), the Seventeenth Financial Reporting and Business Communication Conference in Bristol (2013) and the Accounting and Audit Convention in Cluj (2015). An earlier version of this research project circulated under the title βBankruptcy Prediction Based on Stochastic Processes: A New Model Class to Predict Corporate Bankruptcies?β and is part of the first author's Ph.D. thesis (Klobucnik 2013). Any errors are our own.
[a] Cologne Graduate School in Management, Economics and Social Sciences (CGS), Albertus Magnus Platz, 50931 Cologne, Germany, e-mail: [email protected].
[b] (Corresponding Author) Cologne Graduate School in Management, Economics and Social Sciences (CGS), Albertus Magnus Platz, 50931 Cologne, Germany, e-mail: [email protected].
[c] Chair of International Accounting, University of Paderborn, Warburger Str. 100, 33098 Paderborn, Germany, e-mail: [email protected].
1
INTRODUCTION
This paper investigates a simple accounting-based framework that allows us to predict financial
distress up to five years in advance using a broad sample of US listed firms. Thus, it adds to the debate
of accounting-based vs. market-based bankruptcy prediction models, that centers on the question
whether market-based models outperform accounting-based models or vice (references include:
Agarwal and Taffler 2008; Das et al. 2009; Reisz and Perlich 2007; Xu and Zhang 2009).
Attributable to its academic and practical relevance, the two most widely applied classes to predict
financial distress1 are mainly driven by either employing accounting-information (references include:
Altman 1968; Beaver 1966, 1968b, 1968a; Ohlson 1980; Zmijewski 1984), or relying mostly on
market-information (references include: Bharath and Shumway 2008; Campbell et al. 2008; Hillegeist
et al. 2004; Vassalou and Xing 2004). Accounting-based prediction models made a major progress
since the 1960s given the multivariate approach by Altman (1968) and conditional probability models
as in Ohlson (1980), which are both best-known in research and practice today. 2 However,
accounting-based models are often criticized for their inherent limitations. As Vassalou and Xing
(2004) or Hillegeist et al. (2004) argue, accounting measures, according to their backward-looking
perspective, are less value-relevant to predict imminent financial distress. In addition, accounting
models are mainly static, single-period, and purely statistical models (Mensah 1984). They generally
exhibit limited theoretical grounding as they summarize the default conditions in their estimated
factor loadings (Agarwal and Taffler 2008). Consequently, the estimation results might be unstable
and not generalizable to other samples. In addition, several studies find evidence of a declining
predictive ability using accounting ratios (Beaver et al. 2005; Francis and Schipper 1999).
1 The literature has used a number of different metrics to proxy for financial distress (Balcaen and Ooghe
2006).Technically, financial distress is an early stage of business failure or bankruptcy, where the firm is unable to meet the terms of a loan obligation (Whitaker 1999). Separating different levels of financial strength is in line with Duffie et al. (2007), who merge Moodyβs Default Risk Service database with CRSP/Compustat and other data sources and distinguish amongst others 1) bankruptcy, 2) default and 3) failure, where failure includes 1) bankruptcy, 2) default and adds any failures to meet exchange listing requirements. Jones and Hensher (2004) define their financial distress indicator in a similar three-stage classification model and matched financially distressed firms to a group that is unable to (i) pay listing fees, (ii) generate sufficient working capital, (iii) meet loan agreements (loan defaults) and (iv) to make loan repayments (with debt/equity restructurings). Given that our study uses CRSP and Compustat as data source (see section 4 for details) we measure financial distress broadly and our approach can be regarded as broader business failure model.
2 For exhaustive literature reviews we refer to Dimitras et al. (1996); Aziz and Dar (2006); Balcaen and Ooghe (2006); Ravi Kumar and Ravi (2007); Jackson and Wood (2013).
2
Most importantly, however, the standard accounting models do not explicitly incorporate volatility
measures such as asset volatility, which is a major driving force for bankruptcy and thus prominently
highlighted in many studies investigating the performance of market- and accounting-based failure
models (see e.g., Hillegeist et al. 2004; Shumway 2001; Das et al. 2009). Naturally, more volatile
businesses correspond to higher default risk and neglecting this relation may cause lower model
performance. Consistent with these findings, PΓ‘stor and Veronesi (2003) provide evidence that (i) the
market volatility (measured by volatility of returns) increases with metrics of accounting volatility
(measured by volatility of profitability), (ii) the firm age is negatively related to uncertainty in
accounting measures, and (iii) more volatile firms are more likely to fail. Finally, accounting-based
models do not utilize the benefits of non-financial statement information contained in the market
prices of equity and debt (Campbell et al. 2008).
We address the previous shortcomings of accounting-based models by combining the research on
forecasting financial statement information using the equity valuation literature (refernces include:
Duffie and Lando 2001; Duffie et al. 2007; Favara et al. 2012; Garlappi et al. 2008; Garlappi and Yan
2011; Leland and Toft 1996; PΓ‘stor and Veronesi 2003, 2006; Schwartz and Moon 2000, 2001;
Anderson and Carverhill 2012) 3 with the well-established bankruptcy prediction literature. This
allows us to leave the static, backward-looking perspective of statistical models and instead offers a
multi-period, stochastically grounded, generalizable and forward looking approach.4
In contrast to the prior literature financial distress is triggered by the exogenous boundary condition
to meet short term obligations (i.e., viewed as down-and-out-barrier) and take into account the relative
information content of accounting volatility measures. Specifically, we estimate two key stochastic
processes (i) sales and (ii) costs. The stochastic processes are calibrated to track a firmβs underlying
3 Another early approach using a binomial process to model the adjusted cash position in order to predict failure is
found in Wilcox (1971, 1973). Selecting explanatory variables for a multivariate model motivated by firm valuation considerations is also found in Aziz et al. (1988). Finally, Emery and Cogger (1982) employ a structural stochastic prediction model linking the liquidity position of a firm to the net cash flow to measure a firmβs liquidity and solvency status. Contrary to our study the primary focus of Emery and Cogger (1982) is to provide measures on liquidity, not measuring the probability of financial distress.
4 Shumway (2001) emphasizes that static prediction models ignore the changing characteristics of firms through time. Although we use backward looking accounting data to initialize the stochastic processes to model future firm development for the large and anonymous dataset, one could also use other sources as expert knowledge to estimate forward looking parameters as the growth rate. One source could be analyst forecasts for the growth rate; however they are only available for around 26% of our sample.
3
drift and uncertainty of future operating cash flow and liquidity position.5 Additionally, we include
industrial-, time- and firm-specific parameters to capture group and individual firm effects as
demonstrated to be important components (see e.g., Chava and Jarrow 2004; Opler and Titman
1994).6
We assess current state of the art accounting and market-based financial distress prediction models,
i.e., Altman (1968), Ohlson (1980), Bharath and Shumway (2008) and Campbell et al. (2008) by
comparing three different dimensions: (i) forecast accuracy, (ii) information content, and (iii)
additional statistical inference, where we use a large sample comprised of 41,664 observations
classified as financially distressed and 288,885 non-distressed firm quarter observations spanning
from 1980Q1-2010Q4.
The results confirm that a stochastically driven model improves the long-run discrimination ability,
measured in terms of the area under the receiver operating characteristic curve (AUROC), and carries
incremental information not covered by the original and re-estimated versions of the o-score and z-
score models. First, our stochastically approach is more accurate in distinguishing between non-
delisting and delisting firms than the prominent o-score or z-score models, even if re-estimated (e.g.,
AUROC=0.8122 compared to AUROC=0.7818 and AUROC=0.7327 for the o-score and z-score
models). Second, a combination of all three models has significantly higher explanatory power for
financial distress (e.g., the pseudo-RΒ² increases about 47% by adding our stochastically driven
financial distress measure to the original o-score or z-score measures using a 1-year ahead prediction
horizon).
This paper contributes to the literature by separating financially distressed firms from their healthy
counterparts, which is of particular interest in a variety of contexts.7 Balcaen and Ooghe (2006) point
5 Referring to the financial distress drivers, Hilscher and Wilson (2013, p. 9) stated that βFirms are more likely to fail if
they are less profitable, have higher leverage, lower and more volatile past returns, and lower cash holdings.β 6 While this study picks up the idea that volatility in accounting numbers might be helpful for bankruptcy prediction, in
contrast to the above mentioned studies we do not simply add volatility as an additional variable, but consider the volatility measures in historical accounting numbers as components of the aforementioned stochastic processes that drives a firmβs operating performance and liquidity. While many financial statement variables have been shown to be useful in failure prediction a recurring and often confirmed result is that liquidity variables measured from accounting data have predictive power for bankruptcies up to five years into the future (Beaver 1966, 1968a; Gentry et al. 1985, 1987; Merwin 1942; Pompe and Bilderbeek 2005).
7 For example, specific fields of application and research that are worth mentioning include: First, the ongoing debates in the research literature about the relation between equity returns and a firmβs default risk, i.e. the βdistress-puzzleβ (see e.g., Campbell et al. 2008; Dichev 1998; Friewald et al. 2014; Kapadia 2011; Kim 2012; Vassalou and Xing 2004) or the explanation of credit risk premiums (see e.g., Correia et al. 2012; Das et al. 2009; Subrahmanyam et al. 2014).
4
out that future research should put more weight on longer-term failure prediction. Most studies focus
on failure prediction over the next period, which can be, e.g. one month or one year (see e.g., Chava
and Jarrow 2004; Shumway 2001). However, it is important to predict potential failure earlier, since
the majority of loans agreements fail after their first year (Falkenstein et al. 2000). Campbell et al.
(2008, p. 2900) describe this point succinctly with the paraphrase that distress prediction β[β¦] may
not be very useful information if it is relevant only in the extremely short run, just as it would not be
useful to predict a heart attack by observing a person dropping to the floor clutching his chestβ.
Information that a firm is approaching a difficult financial condition also allows for early preventive
and corrective actions of both management and stakeholders. While we acknowledge that some
studies investigate longer forecast horizons (Campbell et al. 2008; Duffie et al. 2007; Reisz and
Perlich 2007), this paper adds to this research line by also investigating longer forecast horizons of
up to five years.
The rest of this study is structured as follows. In the next section we give a brief overview of the
financial distress prediction literature and introduce the benchmarks. Section 3 introduces the model.
The subsequent section 4 presents the data and the model implementation while section 5 offers the
results from the empirical tests. Finally, section 6 summaries principal findings and concludes.
PREVIOUS RESEARCH
The role of accounting information in modern credit risk theory dates back to the 1930s and numerous
metrics and statistical tools have been proposed to predict the financial condition of
distressed/defaulted firms.8 Ravi Kumar and Ravi (2007) report in their literature review that the
Second, studies show that CEO compensations and executives turnovers are influenced by the presence of ex ante financial distress risk (see e.g., Chang et al. 2015; Gilson 1989). Third, financial distress also affects the dividend policy (DeAngelo and DeAngelo 1990) and corporate liquidity policy (Sufi 2009). Fourth, the use of financial distress models has been fully established in applied areas, e.g., bank regulation, rating and investing processes and to assess the creditworthiness of bond issuer and credit agreements. Fifth, financially distressed firms lose economic values, e.g., market shares, customers, suppliers and human resources, suffer in form of paid covenants penalties and/or higher cost of debt etc. (see Opler and Titman 1994; Purnanandam 2008).
8 Referring to the early studies, the pioneer work of Beaver (1966, 1968a, 1968b) using univariate analysis established a new strand on distress/default prediction literature. The univariate approaches lead to statistical-grounded techniques concentrated on the multivariate framework (references for the multivariate discriminant analysis are Altman (1968), Altman et al. (1977), Deakin (1972), critical commented by Eisenbeis (1977), Dambolena and Khoury (1980), Pompe and Bilderbeek (2005)) and conditional probability models (references include Martin (1977), Ohlson (1980) and the probit model by Zmijewski (1984)). Jackson and Wood (2013) identify a total of 25 alternative approaches throughout their literature review of 350 academic papers.
5
majority of studies are based on accounting measures. To overcome many of the limitations of
accounting-based models the recent literature examines financial distress models including particular
information from the stock markets. Examples of prior research include the contingent claims
framework introduced by Black and Scholes (1973) and Merton (1974) and adapted, for example, by
Sobehart et al. (2000), Vassalou and Xing (2004), Hillegeist et al. (2004), Bharath and Shumway
(2008) or more recently by Charitou et al. (2013). As shown by Shumway (2001), Beaver et al. (2005),
Campbell et al. (2008), and Correia et al. (2012) combining market-related with accounting-related
variables can increase the predictive ability of the traditional ad-hoc accounting-based models.9 On
the other hand, the empirical findings of Agarwal and Taffler (2008), Reisz and Perlich (2007), Das
et al. (2009) and Xu and Zhang (2009) imply that the statistical performance between market-based
and accounting-based prediction models remains competing. While market-based variables reflect
additional, timelier information not captured by fundamental statements (Beaver et al. 2012), several
studies suggest that accounting measures are incrementally informative to predict bankruptcy (see
e.g., Hillegeist et al. 2004). Xu and Zhang (2009) summarize that βthe option pricing theory-based
bankruptcy measure is more successful than the accounting variable-based measures alone, but it does
not subsume the accounting measures.β Moreover, Batta and Wan (2014) show that accounting-based
default models are less sensitive to stock market misvaluations than market-based default prediction
models.
The question then arises whether the benefits of market-based drivers could be incorporated for the
vast majority (i.e., more than 99%) of companies in the United States having no shares listed at the
NYSE/NASDAQ, i.e. private firms?10 Due to the predominance of private companies in the economy
9 However, in a related study Fu (2008) focuses on modeling future working capital and cash flow. Regarding the
working capital model she links the development of future working capital over time to estimates of expected net income and volatility of net income based on, e.g., the normal distribution. Similarly, regarding the cash flow model, she treats revenues as a random variable and estimates the distribution for revenues (expected revenue and volatility of revenue). However, this study does not combine various stochastic processes at the same time as in Schwartz and Moon (2001) or PΓ‘stor and Veronesi (2006) to forecast future company performance.
10 According to the latest (July 2015) statistics the World Federation of Exchanges reports 5,292 listed companies at the NASDAQ/NYSE (www.world-exchanges.org/statistics) which are related to 23,735,915 nonemployer establishments or 5,726,160 enterprises reported by the Statistics of U.S. Businesses 2012 (SUBS) (http://www.census.gov/econ/susb). For the definitions see http://www.census.gov/econ/susb/definitions.html. Contrary, the U.S. Internal Revenue Service (IRS) accounts 5.8 million active corporate tax returns (http://www.irs.gov/uac/Tax-Stats-2) and CRSP covers 26,500 active and inactive stocks listed at the NASDAQ/NYSE. The statistic provided by NASDAQ group reports 6,751 listed companies at the NASDAQ/NYSE/AMEX (see www.nasdaq.com/screening/company-list.aspx). Davis et al. (2007) reports that private firms employ over two-thirds of the US workforce. A further discussion of the relative importance of private vs. public firms is offered by e.g., Chen et al. (2011) and Hope et al. (2013). For example, Hope et al. (2013) compare the reporting quality between public and private firms using a large dataset of U.S. private companies.
6
that are not listed on exchanges, market-based models cannot be employed for these firms as they
lack essential market data. At the same time, it is especially important to evaluate default risk for non-
listed firms, for example private firms and small borrowers, as they experience high risk of default.
Moreover, there is evidence that equity markets are not necessarily efficient. Prior studies such as
Sloan (1996) find that the market does not accurately reflect all information in the financial
statements. During times of volatile markets, as for example the dot-com bubble, relying on
accounting information rather than market data helps to assess the default probability (see e.g., Das
et al. 2009). Consequently and different from recent bankruptcy prediction models by Shumway
(2001), Chava and Jarrow (2004) and Campbell et al. (2008), we restrict our information set to
accounting data to potentially allow default prediction for non-listed firms. The aforementioned
studies present recent attempts to find the most accurate prediction model; however, they rely on
market data and therefore are not applicable for non-listed firms.
While Dimitras et al. (1996), Balcaen and Ooghe (2006), or Ravi Kumar and Ravi (2007) offer
extensive reviews of the last 70 years of bankruptcy prediction literature and related prediction
techniques, we focus on statistical accounting models, because they are still used in the accounting
practice. In particular, we compare our fundamental-based framework to Altman (1968) and Ohlson
(1980). While there are several studies, for example Begley et al. (1996) and Hillegeist et al. (2004),
which demonstrate that Ohlsonβs model outperforms the Altman z-score, Agarwal and Taffler (2008)
find that the z-score model outperforms other statistical models. Moreover, Altman and Saunders
(1998) report that multivariate discriminant analysis models are by far the most widely used statistical
models. Hence, we use both z-score and o-score, which dominate the literature on statistical
bankruptcy prediction models, as there is no clear (ex ante) superiority of one over the other.11 Finally,
several studies (see e.g., Agarwal and Taffler 2008; Dichev 1998) confirm that the two statistical
models also perform well on more recent data.
11 Both models were estimated on data more than 30 years ago, hence they are completely out-of-sample for this purpose.
In addition, we employ the re-estimated coefficients for z- and o-score to address possible structural breaks over the last decades in the bankruptcy process. However, the empirical results for the updated coefficients are weaker. One reason may be that they are developed on a narrower definition of bankruptcy. This is in line with Begley et al. (1996), who do not find an improvement in predictive performance for the updated z- and o-score models and Hillegeist et al. (2004), who confirm this finding for the z-score model. Further, Beaver et al. (2012) and Beaver et al. (2005) find evidence for a significant decline in the predicative power of fundamental information. Balcaen and Ooghe (2006); Mensah (1984) also notes that non-stationary and data instability causes (1) lower predictive power, (2) fundamentally unstable factor loadings and (3) a sample-specific arbitrary character of the estimation results.
7
2.1 Z-Score Model
The multivariate discriminant approach of Altman (1968) is a widely used accounting-benchmark
model scoring financial distress risk based on proxy variables capturing the financial liquidity, long-
term and short-term profitability of assets, the capital structure and the asset turnover of a firm. The
original z-score model based on a paired sample design of 66 firms from 1946-1965 and predicting
bankruptcy over the next year has the following form:
1.2 1.4 3.3 0.6 0.999WC RE EBIT MVE SALESz scoreTA TA TA BVL TA
β = β + β + β + β + β (1)
where WC, RE, EBIT, MVE, SALES, TA and BVL correspond to working capital, retained earnings,
earnings before interest and taxes, market value of equity, sales, total assets and book value of total
liabilities, respectively. While Altman (1968) finds 95% accuracy for the small initial sample one
year before failure, subsequent tests until 2000 demonstrate the accuracy of 80-90% in distinguishing
solvent from insolvent firms (Altman 2000). Moreover, the model's predictive ability drops off
considerably for out-of-sample predictions (79%) and for longer horizons. Hence, to allow for sample
specific coefficients we re-estimate the z-score model using a standard logic regression design in a
growing window approach (henceforth z-scoreu).12
2.2 O-Score Model
Although still popular and widely employed, Altmanβs z-score model is subject to several criticisms
because of the multiple discriminant analysisβ highly restrictive assumptions. These are, among
others, the multivariate normality of the independent variables and omitting of non-linear effects as
argued by Ohlson (1980). As response, Ohlson (1980) introduces an alternative model based on logit
analysis, which is commonly considered as less demanding (Balcaen and Ooghe 2006). He estimates
the conditional probability model on a dataset of 105 bankrupt and 1058 non-bankrupt firms for the
period 1970 to 1976 with the help of the non-linear maximum likelihood technique. As result, he
finds nine variables to be significant and defines the o-score model as:
12 The results and specification are given in the Appendix 1. We are aware of the revised zΒ΄-score for private firms (see
Altman 1983) that uses book value of equity and do not rely on market information. For the purposes of comparison and with reference to wide adaption in the prior bankruptcy literature, we include the original z-score model (1968) to avoid an unwanted look-ahead biases and generate the updated model as additional benchmark.
8
[ ] [ ]1 2 0 & 0
1
1
1.320 0.407 log 6.030 1.430 0.076 price level
1.720 2.370 1 .830 0.285
0.521 | |
t tBVL TA NI NI
t t
t t
TA BVL WC CLo scoreGNP TA TA CA
NI FFOI ITA BVL
NI NINI NI
β β> < <
β
β
β = β β β + β β β + β
β β + β β β + β
ββ β β
(2)
where TA, BVL, WC, CL, CA, NI and FFO correspond to total assets, book value of total liabilities,
working capital, current liabilities, current assets, net income and funds from operations, respectively,
and Ix represents the indicator function, which is equal to one if condition x is fulfilled and zero
otherwise. While a higher z-score corresponds to higher solvency, the relation is reversed for the o-
score. As with the z-score model we re-estimate the original o-score coefficients (henceforth o-scoreu)
on a quarterly basis, using a growing window and the logit regression approach, to capture economic
changes in the underlying informativeness of financial ratios.13
Next, in order of comparability we use a logistic transformation 11+ππππ
,ππ β (ππ,βππ) to transform the
original (re-estimated) z-score (z-scoreu) and o-score (o-scoreu) into financial distress probabilities z-
prob (z-probu) and o-prob (o-probu), which are used as benchmarks in the following.14
One of the major critiques of the class of statistical models is their limited theoretical underpinning.
Statistical models capture information from the default process in the estimated parameters. Hence,
the exact default mechanism remains hidden and the parameter estimates are confined to a specific
time and sample. The necessary re-estimation of the parameters consumes a lot of data. Chava and
Jarrow (2004), for example, consider the sample period 1962-1999, but have to use the period 1962
to 1990 for parameter estimation. Hence, they can only predict bankruptcies for the period 1991 to
1999. Moreover, Xu and Zhang (2009) point out that the predictive variables and the functional form
of the predictive relationship are not derived rigorously for statistical models. In this context, Platt
and Platt (1990) claim the discrepancy of within-sample and out-of-sample classification results.
Mensah (1984) notes that the structure of statistical models requires a stable economic environment
(e.g., such as the rate of inflation, interest rates) and shows that data instability reduces the
13 The results and specifications are given in the Appendix 1. 14 More exactly, the z-score would be transformed into the default probability z-prob with standard normal cumulative
distribution function Ξ¦(-Z). However, Hillegeist et al. (2004) argue that this approximation is sensible. Moreover, it does not change the order and therefore has no impact on sorting based performance measures.
9
discrimination power. In the next section we present a theoretical approach, which is based on the
current line of company valuation research to address these issues. The proposed framework explicitly
models the firmβs processes and the default mechanism using an intuitive and theoretically grounded
approach.
THE MODEL
In the following we provide a simple yet comprehensive non-linear model, where (1) the firmβs
operating performance and (2) the development of its operating liquidity are the major driving forces
for survival. The focus is on the firmβs ability to generate sufficient cash inflows from its operations,
which is critical in avoiding distress (Luoma and Laitinen 1991; Pompe and Bilderbeek 2005). It is
well established through strong evidence that cash flows and operating performance are important
variables for bankruptcy (references include: Altman 1968; Beaver 1966; Emery and Cogger 1982;
Jones and Hensher 2004; Shumway 2001). Moreover, Mossman et al. (1998) demonstrate that among
different models, the cash flow model discriminates most consistently two to three years before
bankruptcy. In order to model the operating cash flow, we build upon recent equity valuation models
(Duffie and Lando 2001; Garlappi et al. 2008; Garlappi and Yan 2011; PΓ‘stor and Veronesi 2003,
2006; Schwartz and Moon 2000, 2001; Anderson and Carverhill 2012). These studies model a firmβs
key profitability by stochastic processes to derive a value estimate for the firm similar to a discounted
cash flow setting. The most recent models used in Garlappi and Yan (2011) and PΓ‘stor and Veronesi
(2006) employ mean-reverting processes for the revenues and earnings, respectively. For example,
Garlappi and Yan (2011) build a simple equity valuation model to investigate the impact of
shareholder recovery and the risk structure of equity; however, they obtain their probability of default
estimates from Moodyβs KMV. PΓ‘stor and Veronesi (2006) seek to explain the high valuations of
technology firms during the dot-com bubble rationally with respect for uncertainty of the average
profitability. However, their model neglects the probability of default and does not include costs
explicitly. Anderson and Carverhill (2012) simulate a well-designed structural equity valuation model
with mean-reverting revenues. Their study is based on a single benchmark firm and focuses on the
determinants of the optimal cash saving policy.15 While all of these models are related to financial
15 Contrary to the reported default rates from the latest Standard & Poorβs (2014) default study, their model calibrations
produce unreasonably high estimates for the default probability. For example, the single benchmark firm fail in 10.86
10
distress, neither study benchmarks its implied distress probabilities to alternative bankruptcy
prediction models.
With reference to the prior literature, we apply mean-reverting (Ornstein-Uhlenbeck) processes for
the modeling of the cash-generating processes (Anderson and Carverhill 2012; Garlappi and Yan
2011; Klobucnik and Sievers 2013; PΓ‘stor and Veronesi 2006; Schwartz and Moon 2001). One major
advantage of the approach for this purpose is that it offers explicit modeling of volatility, both in the
sales and cost processes. It is well established in the literature as shown by Shumway (2001) or
Hillegeist et al. (2004) that volatility is a major driving force for bankruptcy. While structural models
(e.g., Black and Scholes 1973 and Merton 1974) are successful because they estimate asset volatility
from market data, which usually includes various firm relevant information, this study restricts to
accounting information. Consequently, the use and interaction of different stochastic processes
intends to mimic the market volatility by combining various sources of accounting volatility. In
particular, there are three independent sources of volatility: sales volatility, growth volatility, and cost
volatility. They correspond to variations in business and market characteristics, future prospects, and
production technology, respectively.16 The accounting volatility measures are discussed further in
section 3.2. Given the modeling of the operating performance and its volatility, the idea is to adopt
this framework to predict early warning signals through the firmβs cash flow from operations.
In order to determine a measure of financial distress, we follow the perspective of Beaver (1966)
viewing a firmβs liquid assets as a buffer against unexpected cash outflows. The buffer is
operationalized by the operating liquidity reserve (OLR), which is defined as the difference of current
assets (including cash and unused credit lines) and current liabilities and closely related to the net
working capital position.17 As stated above, the OLR level is increased (decreased) by the positive
(negative) stochastic cash flow from operations, i.e. sales minus costs. It has long been known
(Merwin 1942; Smith and Winakor 1935) and is well established (Altman 1968; Kahya 1997; Ohlson
1980) that the operating cash level is an important indicator for bankruptcy and is used, in terms of
percent with reference to a 5-year prediction horizon (Anderson and Carverhill 2012, Table 1). In comparison, Standard & Poorβs report a global five years cumulative default probability in the βBBBβ category of about 3.4 percent (1981-2014). The corresponding default rate using Fitch rating category is about 2.03 percent (1990-2014).
16 In a similar view, PΓ‘stor and Veronesi (2006) account for a firm-specific and an economic component (reflecting the economic cycle) as drivers of their stochastic earnings process.
17 In the literature the terms βnet working capitalβ (see e.g., Bates et al. 2009) and βworking capitalβ (see e.g. Altman 1968) are usually used as synonyms. However, in this study we speak of the βoperating liquidity reserveβ to stress the difference to the measurement of working capital.
11
the net working capital position, in both the z-score and o-score models. In his study Altman (1968)
finds that the working capital ratio is the only significant individual predictor for failure. Moreover,
there is evidence that the role of liquidity in the form of working capital has increased from the 1980s
on (Begley et al. 1996). In contrast Acharya et al. (2012) emphasis that the role of the cash savings
are associated with higher, not lower credit risk in the long-term. A positive OLR demonstrates the
firmβs ability to pay off its short-term liabilities. As a measure of available liquidity, this indicates
that suppliers and creditors could be satisfied from this amount. Anderson and Carverhill (2012) and
Sufi (2009) demonstrate that the corporate liquid assets are a function of the expected cash flow from
operation, whereby a higher profitability risk is accompanied with higher liquidity reserves. For the
model we assume that the liquid reserves evolves with the firmβs cash flow from operations (see e.g.,
Emery and Cogger 1982; Anderson and Carverhill 2012). In case the OLR has been exhausted and
drops below a certain level, the firm is technically in financial distress, analogous to structural models
in Black and Scholes (1973) and Merton (1974). But in contrast to the contingent claims framework
we trigger financial distress by the short term operating liquidity reserve.18 This allows firms to
continue operating with negative net worth (i.e., total assets minus total liabilities) as long as the
liquidity from the reserve remains, which is consistent with empirical findings (Reisz and Perlich
2007). The implicit assumption is that creditors are reluctant to provide additional funding when both
the firmβs current assets do not cover current liabilities and it loses money from its operations. Overall,
the larger the cash flows from operations and the operating liquidity reserve, the lower the probability
of financial distress predicted by the model, which is in line with the results in Kahya (1997) and
Anderson and Carverhill (2012).
In the following, we present a firm stochastic profitability process implemented in a simplified income
statement that is accumulated in the firm-specific operating liquidity reserve.19
3.1 Sales and Costs Processes
Banker and Chen (2006); Barth et al. (2001), among others, find improvements in the predictive
ability by disaggregating earnings into its major components. Accordingly, we model the operating
18 This is consistent with Wruck (1990, p. 421), who defines financial distress as βthe situation were the cash flow is
insufficient to cover current obligationsβ. 19 In a related study Fu (2009) looks at future working capital and cash flow. However, the study links the development
of future working capital over time to estimates of expected net income and volatility of net income based on, e.g., the normal distribution. Hence, the analysis does not combine various stochastic processes at the same time to forecast future company performance and limit the scope of her study to a dataset of 183 small U.S manufacturing firms.
12
performance process by decomposed sales and cost processes. It is a management accounting
perspective where operating earnings are sales net of costs, which mimics the major cash flows of the
firm. The sales dynamics (S) are given by the stochastic differential equation:
( )( ) ( ) ( ) ( )S S
dS tg t dt t dz t
S tΟ= + β (3)
where ππ(π‘π‘) is the sales growth rate, ππππ(π‘π‘) the volatility of sales, and π§π§π π represents unanticipated
changes in sales, which follow a Wiener process. Hence, the development of the sales is driven by
the growth rate. This modeling approach is well established in the literature (Anderson and Carverhill
2012; Garlappi and Yan 2011; PΓ‘stor and Veronesi 2006; Schwartz and Moon 2001). For the sales
growth rate we assume a mean-reverting process:
( ) ( )( ) ( ) ( ) g gdg t g g t dt t dz tΞΊ Ο = β + β (4)
which pictures best the convergence to a long-run equilibrium growth rate οΏ½Μ οΏ½π determined by market
specifications as competition. The volatility of sales growth ππππ(π‘π‘) and the unanticipated changes in
growth rates π§π§ππ describe the unpredictable components (i.e., systematic noise) of the growth rate. The
speed of convergence parameter π π captures the concept of adjusting to long-run equilibriums. With
the help of the term ln(2)/π π the speed of convergence can be expressed as an intuitive half-life of the
process.20
Beside sales, the cost component has to be modeled. Following the approach in Klobucnik and Sievers
(2013) we abstract for simplicity from the separation of fixed and variable costs and model the selling,
general and administrative expenses and the cost of goods sold together in the cost ratio c with a
mean-reverting process:
( ) ( )( ) ( ) ( ) c cdc t c c t dt t dz tΞΊ Ο = β + β (5)
which also convergences to the long-run cost ratio ππΜ with the speed parameter π π and takes the
volatility of the cost rate ππππ(π‘π‘) and the unanticipated changes in costs π§π§ππ into account.
20 PΓ‘stor and Veronesi (2003) interpret the speed of mean-reversion ΞΊ as component of learning mechanisms about the
long-term equilibrium.
13
3.2 The Accounting Volatility Measures
For the accounting volatility measures in the above processes we assume a deterministic behavior
consistent with the company valuation literature. Regarding the volatility of accounting information,
Dambolena and Khoury (1980) study the variability of financial ratios using a paired sample of 23
bankrupt and nonbankrupt firms in a multivariate discriminant approach. Their findings indicate that
bankrupt firms experience higher accounting volatilities even for longer forecast horizons of up to
five years. In contrast, Pompe and Bilderbeek (2005) show that the inclusion of accounting ratio
volatilities adds only limited additional power for bankruptcy prediction when simply included as a
factor in a linear statistical model using a sample including 1369 bankrupt firms. Analogous to
accounting volatility measures, Duffie and Lando (2001) show that accounting noise (i.e., uncertainty
of fundamental information) increase the default intensity in a structural valuation setting. Therefore,
besides the high importance of volatility measured by stock market data in contingent claims models,
there seems to be some indicative evidence that volatility inherent in accounting data might be
beneficial for bankruptcy prediction.21 It is often assumed that maturing firmsβ businesses stabilize
over time (Nissim and Penman 2001; PΓ‘stor and Veronesi 2003). This means, that the abnormal part
for the volatility measures is faded away, thus the risk in a firmβs profitability becomes βnormalβ. We
will use the following three accounting volatilities to proxy the idiosyncratic profitability risk of (i)
sales volatility (ii) sales growth volatility and (iii) cost volatility, which converge deterministically to
a long-term volatility level πποΏ½π π , πποΏ½ππ and πποΏ½ππ with π π as the speed of convergence:
( ) ( )S S Sd t t dtΟ ΞΊ Ο Ο = β β (6)
( ) ( ) ( ) g g g gd t t dt t dtΟ ΞΊ Ο Ο ΞΊ Ο = β β = β β (7)
( ) ( )c c cd t t dtΟ ΞΊ Ο Ο = β β (8)
The long-run sales volatility and long-run cost volatility is assumed to decline to a normal noise level,
πππποΏ½ and πποΏ½ππ , while the long-term growth rate volatility πποΏ½ππ converges to zero so that formula (7)
21 Additional evidence is provided by Altman et al. (1977) using βstability of earningsβ, which is measured as the
standard error of the estimate around a 10-year trend in return on assets. In particular, this study notes that the stability of earnings is the second most important variable among a set of seven total explanatory variables considered. Earlier, Blum (1974) includes accounting variability measures, such as the standard deviation of net income, in a cash-flow framework to explain company failure. Donelson and Resutek (2015) study about earnings qualities show a strong relevance between earnings uncertainty and earnings expectations. Similar, Dichev and Tang (2009) emphasize substantial predictive improvements of long-term earnings implied by the volatility of earnings.
14
simplifies. By considering three sources of accounting volatility, the model attempts to capture the
different origins of volatility in a firmβs operations.
3.3 The Change in Operating Liquidity Reserve (OLR)
From the processes above, the firmβs earnings before interests, taxes, depreciation and amortization
EBITDA(t) as basis for the operating cash flow is derived as:
( ) ( )( ) ( )1EBITDA t c t S t= β β (9)
which is the gross profit margin (1 β ππ(π‘π‘)) multiplied by sales. From the earnings before interests,
taxes, depreciation and amortization the firm has to service its debt. Therefore, the interest positions
also play an important role for the firmβs liquidity. The net interest (interest income less interest
expense) is modeled as:
( ) ( )f LTDInterest t r cash t r LTD= β β β (10)
where cash is the interest earning cash position and LTD the interest-bearing long-term debt. Cash
earns interest at the risk-free rate ππππ while the firm has to pay πππΏπΏπΏπΏπΏπΏ for its long-term debt. We explicitly
take the financial activities into account because recent research shows that firms significantly
increased this position, e.g., to buffer the risk against future cash shortfalls (Bates et al. 2009; Lins et
al. 2010; Sufi 2009; Anderson and Carverhill 2012; Opler et al. 1999). Despite the importance of
excess cash holdings research argues that (un)used lines of credit are a second instrumental
component of corporates liquidity.22 Therefore, we explicitly model the amount of unused line of
credit available as additional liquidity buffer using the unique dataset of Sufi (2009). 23 By
22 Sufi (2009) notes that about 85% in his random sample of US firms have a line of credit between 1996 and 2003,
which on average representing 16% of total assets (where ca. 10% is the unused portion on line on credit). Lins et al. (2010) report an equivalent median credit line about 15% of total assets in his survey analysis of international companies. Yun (2009) stated that about 63% have lines of credit in his sample of 2,533 firms. Acharya et al. (2013) report a trade-off between credit lines and cash holdings for firms having a growing aggregate risk exposure.
23 Sufi (2009) provides an indicator variable that equals one if a firm has a line of credit and zero otherwise for the period 1996-2003. Additionally, he reports the handcollected data for the unused and used amount of credit lines disclosed in the 10-K filings. Specifically, we use this information to approximate a firms annually unused line of credits by the fitted values from linear (OLS) regression for each Fama-French 10-industry classification code:
[ ] [ ] [ ]ln (totalassets ) /0, 1, 2, 3,lineunused cash flow cash M B cashi i i iΞ² Ξ² Ξ² Ξ²= + + β + β
where cash flow is the operating cash flow of the firm (oancf), (total assetβcash) are the total assets of the firm adjusted for the cash position (at-che) and (M/B-cash) is the decile rank for the market-to-book value adjusted for the cash position. To limit the influence of outliers negative predicted values were set to zero and the results were truncated at the 95th
percentile. There are 2,534 / 46,653 firms (firm quarter) observation with a predicted unused line of credit.
15
incorporating unused credit lines we address firms operating with low or even negative working
capital (e.g., Dell Inc., Boeing Co.) but high operating performance. As result, the net cash flow from
operations CFFO that drives the changes in the operating liquidity reserve equals:
( ) ( ) ( )( )
( )( )
1 ( ) ( ) ( ) ( ) 0 |
( ) ( ) ( ) ( ) 0tax EBITDA t Interest t if EBITDA t Interest t
CFFO tEBITDA t Interest t if EBITDA t Interest t
β β + + > = + + β€ (11)
where tax is the corporate tax rate. In order to arrive at the change in the OLR we assume no major
cash flows from investing or financing activities for simplicity. Consequently, there is no investing
in property, plant and equipment in the short-term and there is no change in the capital structure.24
The latter assumption could be extended in future research by explicitly looking at financing
constraints as for example in Anderson and Carverhill (2012). Finally, the change in OLR is driven
by the cash flow from operations:
( ) ( )OLR t CFFO tβ = (12)
In order to implement the model, two adjustments are necessary. Compared with the daily frequency
of market data in market-based structural models, the highest available frequency of accounting data
is quarterly data. Moreover, the popular Black and Scholes (1973) and Merton (1974) model is based
on one stochastic process while the model proposed in this study employs different processes to
capture various dimensions. Hence, first, we discretize the stochastic processes as in Schwartz and
Moon (2001) to fit the quarterly frequency of accounting data. Second, due to the interaction of these
processes there is no closed form solution. Instead, we have to apply Monte-Carlo techniques to
simulate the cash-generating sample paths. Given this study aims to assess the predictive power for
bankruptcies for up to five years into the future, it is sufficient to simulate the OLR paths up to the
next twenty quarters. Similar to the structural models mentioned above, this study defines financial
distress as state-dependent criterion, where the firmβs current liabilities exceed its current assets plus
an optional liquidity reserve from unused credit lines for any quarter within the subsequent twenty
We thank Amir Sufi for making his dataset publicly available: http://faculty.chicagobooth.edu/amir.sufi/chronology.html.
24 However, by using selling, general and administrative expenses (item: xsgaq) and cost of goods sold (item: cogsq) from Compustat, we do consider research and development costs and amortization of tools and dies where the useful life is two years or less, which are part of investing activities. Moreover, we do take interest payments as part of the cash flow from operations into account, which can be regarded as financing activities in US GAAP. Including dividend payments as cash flow from financing activities and tax loss carry forward balances did not improve predictive power, therefore we do not consider these items in this model.
16
quarters, i.e. the OLR falls below a certain level b and the sample paths is classified as failed and no
longer simulated:
{ }( ) 1OLR k bI < = if ( ) | ( 1) bOLR k b OLR k< β β₯ , where ( )1, 20k t tβ + + (13)
Consequently, the empirical probability of financial distress equals the inverse survival probability at
time T conditional on the simulated sample paths for the OLR:
{ }( )1 10,000
1( )N
OLR k bk N
P financial distress IN <
= =
= β (14)
where N=10,000 is the number of Monte Carlo simulations. Regarding the financial distress boundary
b we need to answer when a firm enters the financial distress stage. For example, Reisz and Perlich
(2007) and Davydenko (2012) discuss findings for firm-specific boundaries. However, their
implication refers to asset-implied values and not to operating liquidity reserve specific boundaries.
While the OLR of solvent firms significantly differs by industry and period, the level for financially
distressed firms naturally converges to zero. In line with the findings of prior literature (see e.g.,
Brockman and Turtle 2003; Reisz and Perlich 2007), we trigger a financially distress event at an early
stage, i.e., before the current liabilities exceed the operating liquidity reserve. Therefore, we allow the
barrier b to be positive and discuss further evidence in the empirical analysis (section 4.2). The next
section describes the data and discusses the empirical estimation results.
17
DATA AND MODEL IMPLEMENTATION
4.1 The Data
Our sample consist of quarterly accounting data from Compustat from 1980Q1 to 2010Q4 for the US
market. 25 Thereby the results are comparable with prior research as the majority of studies are
conducted for the time period 1980-2003 as presented in Ravi Kumar and Ravi (2007). At the same
time, they are more up to date because the sample additionally covers the recent financial crisis. There
are three reasons for starting in 1980 in the literature. First, Dichev (1998) demonstrates that the
Compustat sample contains substantially more firms after 1980 and that business failures have
significantly increased after 1980. Second, the Bankruptcy Reform Act of 1978 changed the
institutional setting and economic risk factors of financially distressed firms (Hackbarth et al. 2015;
Hillegeist et al. 2004). Third, since both the z-score and the o-score models were developed before
1980, the out-of-sample perspective is guaranteed. While most studies use annual accounting data,
this study offers results on a more favorable quarterly frequency. Chava and Jarrow (2004) and
Campbell et al. (2008) demonstrate that the higher frequency of quarterly data, which contain more
timely information, can improve forecasting abilities of the models. We exclude (1) firms with SIC-
codes 6,000 to 6,999 (financial firms) and (2) firm quarter observations where essential accounting
information were unavailable. The equity price data and delisting codes are taken from CRSP daily
and monthly databases for NYSE/AMEX/NASDAQ/ARCA common stocks (share codes 10/11). To
include delisting firm quarter observations not in the date range of the CRSP/Compustat merged
databases (CCM) we follow the methodology of Beaver et al. (2007) for an extended merge
procedure.26 The accounting data is lagged by two months to ensure that the data are observable prior
to a delisting, i.e., it is assumed that financial statements are available by the end of the second month
after the firmβs fiscal quarter-end (Campbell et al. 2008; Correia et al. 2012).
This leaves a large and anonymous dataset of 330,549 firm quarter observations for 10,687 non-
financial firms from Compustat. While many bankruptcy studies work with small datasets, large
samples as in Chava and Jarrow (2004) offer more convincing results. Consistent with prior literature,
25 The risk-free rate and industry-classification portfolios were taken from Kenneth Frenchβs library. 26 We would like to thank Richard Price for providing the associated SAS program on his website:
http://richardp.bus.usu.edu/research.
18
we winsorize the independent variables at the 1st and 99th percentile of their pooled distributions to
prevent extreme values from driving the results.27
There are numerous proxies to empirically define financial distress/failure in the literature, which
range from narrow definitions, e.g., actual filing for bankruptcy or credit spreads to a broader
definition relying on a subsequently deterioration of financial ratios such as EBITDA or cash flow
coverage.28 This study follows the recent literature and applies a broad empirical definition for three
main reasons. First, as defaults are rare events, using this broader (delisting) measure increases the
sample size and therefore the robustness of the findings. Second, certain difficult to predict cases such
as strategic defaults receive less weight. With this approach we concentrate on a severe financial
weakness and avoid firms that only declare bankruptcy for strategic reasons (relief of debt) or
unexpected events (natural disasters) from driving the results (Balcaen and Ooghe 2006). Finally, a
broader definition ensures to capture cases where firms have to counter serious financial problems,
but could avoid filing for bankruptcy (Hilscher and Wilson 2013) or willfully delay their filings. In
general, the broadly conceived financial distress definition may capture more firms than relying on
the legal definition of bankruptcy where the default occurs after a firm is already in a financial distress
situation.
Thus, this study adopts exchange-delisting codes from CRSP, which are common in the literature and
more reliable given the underlying delisting regulations and standards existing for US stock
exchanges. In addition, CRSP delisting information offers explicit dates and reasons for a delisting.
Regarding the economic impact after the delisting date, Macey et al. (2008) report that stocks delisted
from the NYSE and subsequently traded on the so-called βPink-Sheetsβ nearly double their mean
volatility of closing prices, triple their average percentage spreads and simultaneously lose about 50
percent of their share price and daily trading volume. There are different sets of delisting codes related
to financial distress/business failure in the literature depending on the broadness of the definition.
Most studies use the definition of performance-related delistings (see for instance: Campbell et al.
27 Note that the values of sales and operating liquidity reserve do not need to be winsorized or trimmed given these
measures serve as initial parameters of the sample paths. However, the results are robust to alternative filtrations. 28 Examples for a broader definition include e.g., Andrade and Kaplan (1998) who define financial distress by (1) the
ability to meet contractual obligations and (2) debt restructuring. Wruck (1990) distinguishes between (1) stock-based definitions (i.e., poor stock price performance), (2) flow-based definitions (i.e., ability to meet current cash obligations) and the (3) legal definition (i.e., filing for bankruptcy protection). Platt and Platt (2002) search financially distressed companies according to (1) several years of negative net operating income (2) suspended dividends and (3) restructuring or layoffs. More recently Kapadia (2011) rely on an aggregated measure of business failure provided by Dun & Bradstreet.
19
2011; Caskey et al. 2012; Dichev 1998; Reisz and Perlich 2007; Shumway 2001). Performance
delistings are associated with negative changes for the firm and cover the CRSP delisting codes 400
and 550-585. Besides bankruptcy, performance delistings additionally include, among others,
insufficient capital, market capitalization or market-makers and non-payment of exchange fees
(Campbell et al. 2008; Dichev 1998). Similar to Xu and Zhang (2009) we consider all these cases as
financially distressed. By using this broad measure of distress we obtain 41,664 firm quarter
observations and 3,435 firms that were delisted for performance-related reasons. This yields a
cumulative distress rate of 32%, which is substantially higher than for example the 14% in Reisz and
Perlich (2007), who work with a smaller sample and a shorter sample period. Table 1 offers an
overview over the reasons and distributions of delistings and clearly show that performance delistings
can be associated with financial distress or bankruptcy.
We note that delistings as a proxy for bankruptcy have the drawbacks that they are only available for
exchange-listed firms and that they are sometimes not accurate with respect to the timing of
bankruptcy. However, Vassalou and Xing (2004), for example, argue that in their sample firms delist
two to three years prior to actual default. On the other hand, Dichev (1998) argues that some firms
that enter bankruptcy do not delist or delist significantly later. Franzen et al. (2007) find qualitatively
similar results when using performance delistings instead of legal definition using bankruptcy filings.
Nevertheless, this measure is adequate to make statements about cross-sectional variation in the
probability of distress and it is widely used (Campbell et al. 2011; Caskey et al. 2012; Dichev 1998;
Reisz and Perlich 2007; Shumway 2001).
------------------------- Please insert Table 1 about here -------------------------
4.2 Parameter Estimation
The stochastic prediction model has in total 17 parameters, which stem from the firmsβ balance sheets
and income statements. However, the employed information set is the same as for the z-score and o-
score models. Remember that the two statistical models are based on EBIT and sales among others,
and in the same vein our model, for example, employs sales and costs to arrive at the EBITDA.
20
The estimation of the model initial input parameters is presented in Table 2. It employs intuitive and
straightforward estimations based on the firm- and industry-specific first two moments of the
preceding seven quarters to control for seasonal effects. Where there are less than seven preceding
quarters available, we restrict the estimation to the available information set (with a minimum of three
quarters) to keep as many observations as possible in the sample. This is another advantage of the
model compared to the statistical models, which demand more data to be initialized as argued above.
Hence, this model does not need first-stage regression to estimate coefficients (Hillegeist et al. 2004)
and no holdout sample approach to test its performance. Moreover, it is less sensitive to structural
breaks as it depends on (short-term) accounting data and estimates parameters firm-specifically.
The initial sales growth rate is determined as weighted average over the respective series of the
previous seven quarters. Similar to Campbell et al. (2008) we incorporate declining weights to
account for a decreasing information content of lagged accounting data.29 For the initial variable cost
as fraction of revenues we use the estimated coefficient from rolling regression of quarterly costs on
sales assuming no intercept. The initial volatility of sales for each firm is calculated as the standard
deviation of the previous seven quarters. For the initial volatility of sales growth and the initial
volatility of costs we use the standard deviation of residuals from regressions similar to the accrual
quality measure in Francis et al. (2005). For the long-term parameter of costs we estimate industry
medians over the Fama-French 48-Industry classification for two main reasons. First, prior studies
(e.g., Chava and Jarrow 2004; Hillegeist et al. 2004) clearly demonstrate that industry effects are
important for bankruptcy prediction (because of different levels of competition, etc.). Second,
industry-specific parameters provide more stable estimates and neutralize individual outliers in the
large dataset (Klobucnik and Sievers 2013). Both the long-term volatility of sales and the long-term
volatility of costs are set to 0.05, the degree of a long-run noise level.30
The estimation of the speed of convergence parameter deserves more explanation. In the economics
literature the standard approach to estimate convergences is to use the absolute convergence
29 In particular, the weights are reduced every quarter by about 1/3 and the estimated growth rate is reduced about 50
percent with respect to a declining weighted-average. Initializing the growth rate with an arithmetic mean calculation produces similar results but does not cover firms that rapidly lose sales growth (within 8 quarters) and become financially distressed.
30 Moreover, we confirm these values empirically by estimating growing window standard deviations over the whole sample period for maturing firms (age>10 years), which yields a median value of 0.14 (0.03) for the volatility of sales (long-term volatility of costs), respectively. However, the results are not sensitive to these parameters.
21
regression (Rogoff 1996). Similar to Altomonte and Pennings (2008) we use the following regression
to estimate the convergence: 31
,
1, ,1,
ln t ii i t i t i
t i
gg
gΞ± Ξ² Ξ΅β
β
= + β +
(15)
where πππ‘π‘,ππ is the corresponding sales growth rate for time t and firm i and π½π½ππ yields the firm-specific
convergence. These firm-specific values are then pooled for each industry (Fama-French 48-Industry
classification) to define the median as industry-specific convergence parameters for reasons
mentioned above. The balance sheet positions are initialized as presented in table 2.
A final key parameter in our stochastic model is the boundary condition b, i.e., the OLR barrier to
classify a firm as financially distressed. As discussed in section 3.3, technically, a firm is in financial
distress if the current obligations exceeds the liquid funds (including the cash position and unused
credit lines). Empirically, the financially distressed firms in our sample enter a downward spiral
before the operating liquidity reserve is exhausted. Instead of setting an arbitrary level, we model a
down-and-out barrier b that is strictly positive. For each quarter, we estimate the initial distress barrier
b as the median OLR for financially distressed firms five years prior a corresponding delisting event
(estimated in a 5-year rolling window). If the initial OLR is below the prespecified barrier b we fund
the firm with additional interest-bearing debt by the required amount (this adjustment is necessary for
85,113 firm quarters). Prior research mainly focuses on the market-value of assets to specify a certain
default boundary condition (Davydenko 2012; Reisz and Perlich 2007). For example, Reisz and
Perlich (2007) estimate the mean (median) implied barrier level at 30.53% (27.58%) of the market-
value of assets (i.e., non-zero). Analogously, Davydenko (2012) empirically findings indicate a mean
(median) barrier level that equals 66.0% (61.6%) of the face value of debt. Given the boundary
condition is a critical parameter we conduct additional robustness checks in section 5.5. Figure 1 plots
31 Additionally, we use the approach of Klobucnik and Sievers (2013) to calculate the speed of convergence, i.e. the
kappas. Solving οΏ½ π π π π π π πππ π ππβπ π π π π π πππ π ππβ1π π π π π π πππ π ππβ1
=π‘π‘β8
ππ=π‘π‘β5οΏ½οΏ½ π π π π π π πππ π ππβπ π π π π π πππ π ππβ1
π π π π π π πππ π ππβ1
π‘π‘β4
ππ=π‘π‘β1οΏ½ β ππβ4βπ π οΏ½ for kappa yields an alternative estimator,
which is then pooled to medians for the same industry (three digit SIC codes). The estimated default probabilities and the other findings remain qualitatively the same. However the approach in this study is well established in the economics literature and therefore preferable.
22
the estimated barrier b over time. As shown, the barrier level reveals a positive relation to periods
with greater uncertainty.
The debt interest rate is defined as the interest expense divided by the book value of current and long-
term debt as in Francis et al. (2005). The corporate tax rate is set to 35% consistent with, among
others, Garlappi et al. (2008) and PΓ‘stor and Veronesi (2006). For the risk-free rate we choose the
short-term risk free rate represented by the 3-Month US-Treasury Bill rate.
------------------------- Please insert Figure 1 about here -------------------------
------------------------- Please insert Table 2 about here -------------------------
The summary statistics for the individual parameters are presented in Table 3 separating solvent and
distressed firms and the cross-sectional statistics. One can see that the sample firms are highly
divergent with respect to size (parameter 1: quarterly sales ranging from less than one million dollars
to more than four billion dollars), growth (parameter 2: quarterly sales growth ranging from less than
-13% to more than 14%) and profitability (parameter 7: sales margin ranging from 44% to more than
138%). Moreover, the substantially lower sales median of around 38 million dollars compared with a
mean of 280 million dollars demonstrates that there are few very large firms while the majority are
medium-sized firms. Finally, the initial OLR position (parameter 15) for solvent and financially
distressed firms differs significantly (the median for solvent (distressed) firms is 42 (12) million
dollars), which also illustrates the diversity of firms in the sample. Most importantly, we find
significant differences between solvent and distressed firms along the volatility of sales, sales growth
rate and variable costs (parameters 4, 5, and 9). Figure 2 (Panel A, B and C) highlights the
deterioration of the model parameters approaching the performance-related delisting.
------------------------- Please insert Table 3 about here -------------------------
------------------------ Please insert Figure 2 about here -------------------------
23
EMPIRICAL ANALYSIS
In this section we evaluate the performance of the distress prediction based on stochastic processes
approach (henceforth s-score) along several dimensions and benchmark it against the two prominent
statistical models, z-score and o-score, which rely on a similar set of accounting information. First,
we examine the forecast accuracy and construct validity among the main financial distress indicator,
the OLR. Second, we compare our modelβs default rates with observed default rates with the help of
various summary statistics and investigate its early warning ability, i.e. the ability to indicate as early
as possible an increase in distress risk. Third, we evaluate the modelsβ ability to distinguish between
distressed and non-distressed firms by using the well-established concept of accuracy ratios. This
prediction-oriented test assesses the cross-sectional discriminatory power (Vassalou and Xing 2004).
Fourth, we employ information content tests and examine the ability to explain actual defaults with
logit and hazard models as in Hillegeist et al. (2004). Most standard approaches concentrate on short-
term default prediction for the next period, which is usually one year into the future. This study looks
at different forecasting horizons as Duffie et al. (2007) emphasize their importance. In particular, we
investigate longer forecast horizons of up to five years (=20 quarters) because, naturally, it is
favorable to predict potential bankruptcy as early as possible.
5.1 Operating Liquidity Reserve (OLR) Forecast Accuracy
One way to judge the construct validity of our model is to compare the predicted OLR with the
corresponding actual OLR boundary condition implied from the financial statements. Table 4 presents
the absolute and relative forecast accuracy for the s-score model. Panel A contains the accuracy of
the OLR predictions computed as the median of 10,000 sample paths for each of the forecast horizons
(i.e., 1-20 quarters ahead where ππππππππpredict,ππ β (π‘π‘ + 1, β¦ π‘π‘ + 20)). Similar to the methodology of
Weiss et al. (2008) we evaluate the modelβs predictive performance by benchmarking the predictions
against the actual OLR (πππππππ π πππ‘π‘πππ π π π ) and report the absolute and relative differences:
( ) predictab
actu lt ts
aOLR LRBias Oβ= (16)
predict actualt t
actualt
relOLR OLRBia
OLRs
=
β
(17)
24
Panel B provides the descriptive forecast accuracy for the distressed sample. The results show that
the median (mean) absolute one quarter ahead differences in the OLR are -0.16 (+1.91) million or
about -3% (+9%) for the distressed firm sample.32 As expected the bias increase with the forecast
horizon. First, the average estimated time to distress is about 4 quarters (i.e., the median time when
the sample paths violate the down-and-out-barrier). Accordingly, the model relies more on the early
prediction paths. Second, the OLR is clearly industry-specific. Figure 3 gives an unbiased view of the
model performance by plotting the historical industry-specific 1-, 2-, and 3-year ahead predictions for
randomly selected industries. The graphs emphasize the mimicking character of OLR predictions for
longer horizons.
Table 5 reports the expected deterioration of mean and median levels of the OLR for the quarter
preceding a delisting. As previously discussed, the OLR of a firm approaching a performance-related
delisting converges to zero preserving a positive liquidity reserve before delisting. Contrary, the mean
(median) OLR of solvent firms are about 12 (6) times greater.
------------------------- Please insert Table 4 about here -------------------------
------------------------ Please insert Figure 3 about here -------------------------
------------------------- Please insert Table 5 about here -------------------------
5.2 Summary Statistics and Correlations
In this section we present the summary statistics and correlations of the different measures. Table 6,
Panel A, presents the summary statistics for the three models and for Standard & Poorβs long-term
credit rating, which is used to provide an alternative distress risk measure.
------------------------- Please insert Table 6 about here -------------------------
32 Weiss et al. (2008) report for the accounting-based earnings prediction model following Abarbanell and Bushee (1997)
a pooled 1-year ahead median percentage forecast error of 0.329.
25
All three models show the ability to separate delisting firms from non-delisting firms as the average
estimated distress probabilities for delisting firms are significantly higher. Looking at the medians,
the s-prob model reveals an even clearer ability to distinguish non-delisting firms (0.03) from
delisting firms (0.70). Overall, the z-prob and s-prob measure yields the highest estimated
probabilities while the updated z-probu and o-probu presents the lowest. As noted in Hillegeist et al.
(2004) the results indicate that the re-estimated versions are not well calibrated and do not reflect
changes in the underlying accounting ratios. For example, Franzen et al. (2007) report a negative
trend in the relevant accounting ratios, which increase the probability of misclassification. Not
surprisingly, all measures experience misclassifications with solvent (delisting) firms having high
(low) probabilities in the 99% (1%) percentile. However, we acknowledge that the original o-prob
model performs better for the 99% percentile for solvent firms with a relatively low probability of
0.66 compared to 0.99 and 1.00 for the z-prob and the s-prob models. Table 6, Panel B, displays the
correlation measures for the different models. We focus on Spearmanβs correlation measure (below
the diagonal) as the relationships are supposed to be monotonic but not necessarily linear, which is
what Pearsonβs measure detects. While the original statistical models are strongly correlated (0.73),
which is consistent with the value of 0.65 reported in Hillegeist et al. (2004), they are only modestly
correlated with the s-prob model (0.48 and 0.53). This shows that the model proposed in this study
captures relevant aspects that are not included in the z-prob and o-prob models. The lower correlations
(0.37 and 0.25) with the updated statistical models confirm this finding. These promising descriptive
statistics results are to be confirmed in the following analyses.
------------------------- Please insert Figure 4 about here -------------------------
Figure 4 offers the median estimated distress probabilities for the different models according to the
Standard & Poorβs rating class. The predicted probabilities of all models increase with a deteriorated
credit rating, as expected. Compared with the historic default rates from Standard & Poorβs, the s-
prob model pictures the growing inherent risk of credit ratings with a steep increase from rating class
βBB-β onwards to high probabilities levels in the substantial risk categories. However, while ratings
26
provide a reasonably good measure for rough classification of firms according to their risk of failure,
they are a poor measure of actual default probabilities as recently demonstrated by Hilscher and
Wilson (2013). Consequently, we do not consider ratings for the further analyses.
------------------------- Please insert Figure 5 about here -------------------------
Figure 5 displays the development of the estimated default probabilities over the time period 1980Q1
to 2010Q4. In line with the samples in Chava and Jarrow (2004), Campbell et al. (2008) and Kapadia
(2011), the data displays a steep increase in the number of bankruptcies which occurred in the period
from 1984Q1-1992Q3. As demonstrated earlier, the s-prob model pictures the behavior of the actual
defaults with peaks in the mid-1980s and around 2000. The declining pattern in the median financial
distress probability after 1990 is also related to, among less subsequent delistings, higher cash
holdings and a downturn in the leverage ratio (Bates et al. 2009). The o-prob model also performs
well while the z-prob model does not display periods with low bankruptcy rates properly.
Interestingly, for the financial crisis 2007-2009 all models show a slight increase in the average
default probability. As seen in Table 6, the z-prob model yields the highest estimated probabilities
with on average more than 18% while the value is less than 8% for the o-prob model.33
------------------------- Please insert Figure 6 about here -------------------------
Figure 6 demonstrates the evolution of the default probabilities in the quarters that precede the
delisting. One can clearly see the early warning ability of the s-prob model. Compared with the two
statistical models, the s-prob modelβs estimated default probabilities start to increase substantially
already around five years before actual delisting compared with the three years for the z-prob and the
33 Carefully note that the z- and o-score models need a mapping function to transform the score into a distress probability
estimates (see section 4.2). Hence, the calibration of the models could be adjusted by including an empirical mapping function as in the ZETA score (Altman et al. 1977) to change the magnitude of default probabilities. However, the ZETA score is a proprietary product that is not publicly available. As the analyses in the following depend on the ranking by default probabilities, a monotone mapping function would not change the results as it does not influence the ranking.
27
two years for the o-prob. These early signals are economically highly relevant to avoid costly
financial distress (Altman 1984; Warner 1977). One reason for the better performance of the s-prob
model for longer horizons might be the incorporation of the accounting volatility measures, since,
e.g., Agarwal and Taffler (2008) also demonstrate that market-based models, which take volatility
measures into account, perform better for longer horizons than statistical models.
In sum, this section demonstrates how well the models are calibrated to describe observed default
rates. However, the absolute probability levels are less relevant for the following cross-sectional
discrimination tests, which therefore offer another dimension of model testing.
5.3 Accuracy
When evaluating credit risk models two types of errors influence a modelβs quality. First, bankrupt
firms can be classified as non-bankrupt, i.e., a type I error (false negative). Second, non-bankrupt
firms might be measured as bankrupt, thereby making a type II error (false positive). Type I errors
are associated with higher costs, which is why they are commonly considered as less desirable
(Agarwal and Taffler 2008; Altman et al. 1977) while type II errors cause hypothetical loss of profits,
interests and fees.34 As the stochastic model yields expected default frequency estimates, there is no
cutoff value for classifying distressed and non-distressed firms. In contrast, the methodology for the
original z-score and o-score models focus on a specific cutoff by minimizing the type I and/or type II
error. Instead, all three models in this study are viewed as continuous measures (see Hillegeist et al.
2004, Reisz and Perlich 2007). This is desirable because corporate distress is not a well-defined
dichotomy in reality. Loan officers, for example, make continuous decisions at what rate to lend
(Hillegeist et al. 2004). Hence instead of directly counting the number of misclassified firms for a
specific cutoff point we evaluate the discriminatory capacity using a prevalent analytic accuracy
measure, the area under the receiver operating curve (AUROC) (see for references: Agarwal and
Taffler 2008; Caskey et al. 2012; Chava and Jarrow 2004; Reisz and Perlich 2007; Vassalou and Xing
2004). The ROC curve is constructed by varying the cutoff points and plotting the true positive rates
(correct classification of a financially distressed firm) versus the false positive rates (false
classification of a solvent firm). To compute the AUROC measure we follow the nonparametric
approach by Hanley and McNeil (1982) and DeLong et al. (1988) based on sorting the firms by their
34 For a discussion of the relative importance of a type I or type II error see for instance Altman et al. (1977), Altman
(1984), Andrade and Kaplan (1998) and BlΓΆchlinger and Leippold (2006).
28
probability of default estimates (from high to low) and assessing the number of actual defaults in the
highest k-percentiles (k=1,β¦,100).35 The AUROC measure, which ranges from 0.5 (random model)
to a maximum of 1.0 (perfect discrimination model) answers the question of how accurate the model
is in predicting actual defaults and determining cross-sectional distress risk. The higher the curve rises
towards the top left corner point, the higher the area under the curve (AUROC) and the better the
discrimination power.
To obtain fundamental assessments of the model performances table 7 compares the AUROC results
inferred from predicting actual defaults over a 5-year horizon (i) by longitudinal analysis over years
(Panel A), (ii) by cross-sectional and bootstrapped analysis over the sample period 1980-2010 (Panel
B and C), and (iii) by the pairwise statistical comparison for the calculated AUROC across the
investigated models.
------------------------- Please insert Table 7 about here -------------------------
Overall, the s-prob model experiences the highest accuracy (AUROC=0.8122) followed by the o-
prob model (AUROC=0.7818) and the z-prob model (AUROC=0.7327). The most striking result is
the declining accuracy of the statistical models. While the o-prob model performed remarkably well
to the beginning of 1990βs, the accuracy of the z-prob, o-prob and updated z-probu and o-probu models
subsequently decline. One reason for this finding might be due to varying number of bankruptcies
and delistings over time, where a higher number corresponds to a lower accuracy ratio (see Table 1).
The findings also imply deterioration in the predictive power of accounting variables indicating
unstable coefficients and miscalibration, which is consistent with the findings of Beaver et al. (2005),
Hillegeist et al. (2004) and Begley et al. (1996) and explain the lower cross-sectional discrimination
power of the re-estimated models. However, the o-prob model clearly outperforms the z-prob model
in each dimension. Given these results, it is not surprising that the AUROC measures for the z-prob
model and the o-prob model are widespread and depend highly on the sample as well as the forecast
horizon in the literature. For example Agarwal and Taffler (2008, table 2) report an AUROC value of
35 For a detailed discussion of this measure please see Hanley and McNeil (1983) and Stein (2002). For an overview
of the estimation and implementation procedures using STATA see Pepe et al. (2009).
29
0.89 (1-year forecast horizon) compared to the AUROC of 0.7794 (1-year forecast horizon) and
0.6483 (5-year forecast horizon) reported by Reisz and Perlich (2007, Table 3). Chava and Jarrow
(2004, Table 2) estimate a bootstrapped median AUROC of 0.8662 (1-year forecast horizon) based
on yearly firm observations from 1991-1999 using a hazard model. Franzen et al. (2007, Table 7)
report a mean cumulative accuracy ratio of 0.465 (1-year forecast horizon) for the R&D-adjusted o-
prob model using the sample period 1980-2003, which yields an equivalent AUROC of 0.7325.36
Similar, Jackson and Wood (2013) find an area under the ROC curve of 0.7801 (1-year forecast
horizon) based on an annually sample for UK listed firms from 2000-2009. Vassalou and Xing (2004),
find a cumulative accuracy ratio of 0.59 (5-year forecast horizon) for the structural Merton model in
predicting actual defaults that equals an AUROC of 0.7950. However, their model requires market
data and therefore restricts the sample. For example, they exclude firms with negative book value of
equity and firms, which do not have stock market quotes available. Using quarterly data Das et al.
(2009) find accuracy ratios of around 0.60 for accounting and market-based models. Overall, the
reported accuracy ratios are comparable to the prior literature while this study uses a broader measure
of distress (more than 2,000 delisting firms over the whole sample period compared to around 300
for most studies as for example in Shumway 2001) and longer forecast horizons.
In Panel C the results are confirmed by employing 1,000 bootstrapped resamples and calculate the
descriptive statistics for the AUROC (Chava and Jarrow 2004; Reisz and Perlich 2007). In unreported
statistics we compare the AUROC from the s-prob model pairwise with the statistical models
following the U-statistic algorithm by DeLong et al. (1988). As expected by the bootstrapped results,
the performance of the s-prob model is significantly different better compared to the statistical models
(p-value < 0.0001). For the sake of completeness, we sort by (i) cash, (ii) quick ratio, and (iii) assets-
to-liability ratio, which are known to have predictive power for bankruptcy (Ravi Kumar and Ravi
2007), but find, as expected, lower accuracies ((i) AUROC=0.2665, (ii) AUROC=0.3683, (iii)
AUROC=0.2168).
------------------------- Please insert Figure 7 about here -------------------------
36 Agarwal and Taffler (2008) note that the Cumulative Accuracy Profile (CAP) or Accuracy Ratio and the area under
the ROC Curve (AUROC) based on the same information set and are linear related by: πΆπΆπΆπΆπΆπΆ = 2 β (πΆπΆπ΄π΄πππππΆπΆ β 0.5).
30
Figure 7 shows the AUROC of the models for different forecast horizons. It starts with six years
before delisting and shows the evolution of the estimated default probabilities until actual delisting at
time zero. While the s-prob and the o-prob model perform overall comparably, the z-prob and the re-
estimated o-probu and z-probu models are clearly inferior with regard to distinguishing delisting from
non-delisting firms. Looking at the figure in more detail, one can see that until up to the delisting
quarter, the s-prob model performs better than the o-prob model. This is consistent with Mossman et
al. (1998) and Reisz and Perlich (2007), who find that the accounting-based ratio models are the best
single model during the year immediately preceding bankruptcy. Generally, the accuracy naturally
increases the closer the firm is to actual performance-related delisting. As discussed above the values
of around 0.80 to 0.90 for one year before delisting are in line with prior studies.
------------------------- Please insert Figure 8 about here -------------------------
Figure 8 shows the accuracy of the models over time. As before, the s-prob model and the o-prob
model outperform the z-prob model with regard to distinguishing delisting from non-delisting firms
after the initial sample years. In contrast, the z-prob model increase its discrimination power in
economic downturns (2000-2002, 2008-2009). The s-prob model and the statistical models show a
similar behavior over time that is also reflected in the correlation pattern. Overall, the accuracy of all
three models is decreasing until mid-1990s and increasing thereafter again. Interestingly, all models
perform remarkably well during the financial crisis. Finally, with reference to the modelβs industry-
specific character and the findings of Chava and Jarrow (2004), table 8 provide the AUROC
performances across the Fama-French 10-Industry classifications. The results support the industry-
specific accuracy of the s-prob model. Being precise, the AUROC ranking performance shows that
10 out of 10 industry classifications are outperformed by the proposed stochastically framework.
------------------------- Please insert Table 8 about here -------------------------
31
In the next section we investigate the incremental information content of the individual and combined
distress probability estimates.
5.4 Test of Information Content
Following Campbell et al. (2008), Chava and Jarrow (2004) and Shumway (2001) we estimate a
multiperiod (or dynamic) logit model to evaluate the incremental explanatory power. We assume that
the marginal probability of bankruptcy or failure over the next period follows a logistic distribution
and is given by:
( ) ( ), 1, 1
111 exp i t
i t it PDP Y
Ξ± Ξ² ββ β β β
= =+
(18)
where Y is coded one if the firm delists in period t (and zero otherwise) and PD is the explanatory
variable, which is the default probability estimate (coded as logit score) for the different models. The
major criteria to compare the s-score model to the benchmark models is the pseudo-RΒ² as in Shumway
(2001). Additionally we report the Vuong test statistic for strictly non-nested model selection and the
likelihood ratio statistic to compare the nested models.37 To examine the value of the modelsβ default
probabilities for longer forecast horizons we vary them up to five years prior to delisting as in
Campbell et al. (2008).
Table 9, Panel A, shows the results for the dynamic logit model estimation to determine the
information content of the models for a 1-year ahead prediction. Considered individually, all models
have substantial explanatory power for actual delistings as the coefficients are significantly different
from zero and strictly positive. Moreover, the pseudo-RΒ²s range from 0.0450 to 0.0831 individually.
Overall, the magnitudes of the pseudo-RΒ²s are consistent with the findings in Xu and Zhang (2009)
ranging from 0.07 to 0.16 or in Hillegeist et al. (2004) from 0.07 to 0.12 where both include market
information. The s-score and re-estimated z-scoreu models are more preferable in the 1-year forecast
horizon (pseudo-R2=0.0831 and pseudo-R2=0.0745) and perform comparably well while the o-score
and z-score are inferior in terms of the conveyed information (pseudo-R2=0.0689 and pseudo-
R2=0.0450). The findings are in line with the two mentioned studies. In contrast to Hillegeist et al.
37 The Vuong statistic is an established methodology to evaluate competing strictly nonnested models (see for references
Agarwal and Taffler 2008; Banker and Chen 2006; Dechow 1994; Dichev and Tang 2009; Hillegeist et al. 2004). For the implementation of a logit version for the Vuong test we follow Lechner (1991, table 3) and confirm the results using the SAS implementation of the Vuong test statistic.
32
(2004) the re-estimated z-scoreu performs significantly better than the original z-score. As Xu and
Zhang (2009) suggest, Table 9 also sets out the combination of the three measures (i.e., original and
re-estimated Model 6 and Model 7) to compare the incremental relevance carried by the distress
prediction models. For the combined models, the z-score explanatory parameter loses its significance
(the z-statistic declines from 1.79 to 1.34 referring to the results in model 8 vs. model 6). Looking at
the differences in pseudo-R2 of Model 6 and Model 7 versus the pseudo-R2 from the combined
versions (i.e., Model 8 and Model 9 including only the z-score and o-score explanatory variables),
the s-score seems to explain information not contained in the traditional measures.
Overall, the combined pseudo-RΒ² of 0.1069 and 0.1164, which are significantly larger than the
pseudo-RΒ² of the separated models reveal that all measures convey exclusive information.
Table 9, Panel B, Panel C, extent the forecast horizon of the dynamic logit model to three and five
years, respectively. As one expects, the pseudo-RΒ²s decrease for longer forecast horizons. However,
the results demonstrate that the s-score model clearly outperforms the two statistical models for longer
horizons of three to five years. For a horizon of three years, for example, the s-score model displays
a pseudo-RΒ² of 0.0489 compared to 0.0348 (0.0139) for the o-prob (z-prob) model.38 This is in line
with the findings above that the s-score model is superior in early prediction of defaults, which can
be due to the inclusion of volatility (Campbell et al. 2008). In all panels of Table 9 the combined
models (model 6 and model 7) offer the highest explanatory power. Surprisingly, the re-estimated
models and the z-score lose part of their explanatory power compared to the original o-score model
and become insignificant in the combined model versions referring to a 5-year forecast horizon. The
high explanatory power of the s-score for long-term financial distress predictions is confirmed by the
pairwise Vuong-statistic offered in Panel D. Overall, the results suggest that s-score model is
preferable compared to the original and re-estimated o-score and z-score prediction models in case
of long-term forecast horizons.
------------------------- Please insert Table 9 about here -------------------------
38 Recall from above that this literature is concerned with pseudo-RΒ²s ranging from 0.07 to 0.12 for a single explanatory
variable for a one period ahead forecast (Hillegeist et al. 2004). Hence, a pseudo-RΒ² value of about 0.05 for a three years ahead forecast is economic significant.
33
In order to test an alternative dimension of the predictive ability we assess the homogeneity of the
evaluated models in Table 10. Homogeneity measures are derived by ranking the related distress
probabilities into deciles (10=high probability of financial distress β 1=low probability of financial
distress) and compare the predictive ability. This assessment is closely related to the accuracy ratio
and offers additional insights in the modelsβ performances. Panel A1 reports the decile-based
correctly classified frequency of failed firms (separate and cumulative). The results suggest, that the
s-prob, o-prob and z-probu models identify about one-third of the financially distressed firm quarter
observations in the highest-probability decile. Panel A2 presents the ranking for the longer term 2-5
years ahead prediction horizon. Panel B addresses the question of equal classified financially
distressed firm quarter observations. For each probability decile we report the number and frequency
of correspondingly classified observations in relation to the s-prob results. As shown in Table 9 the
s-prob model uses information not enclosed in related models. The equally identified observations in
the highest decile range between 52.6%-66.4% stepwise increasing with each decile. In other words,
about one-third of the financially distressed observations in the highest s-prob probability decile are
individually not identified by related accounting models.
------------------------- Please insert Table 10 about here -------------------------
5.5 Robustness and Extensions
In this section, we report the robustness of the results laid out in the earlier sections. The question
arises as to whether the accuracy and information content hold when we use alternative (i) (market-
based) benchmark measures, and (ii) parametrizations for the s-score model.
(i) Market-based distress prediction models
First, we conduct all our ROC and information content analyses from section 5.2 to 5.4 using two
complementary market-based prediction models. In particular, we operationalize the research designs
for the contingent claim framework (EDF) following the sequential-iterations algorithm specified by
Crosbie and Bohn (2003) and Bharath and Shumway (2008) and the comprehensive information
34
approach (original c-prob and re-estimated c-probu) of Campbell et al. (2008) that incorporates both
market- and accounting covariates.39 As discussed above, we expect that market-information allows
for additional explanatory power regarding the financial distress prediction models. The summary
statistics, accuracy analysis and information content tests are tabulated in the corresponding tables
11, 12 and 13. The additional data requirements to calculate three alternative market-based distress
measures, (i) EDF, (ii) c-prob , and (iii) c-probu reduce the sample size relative to the accounting-
based sample from 330,549 (10,687) to 280,234 (10,143) firm quarter observations (firms) with
observable probability estimations and ensures comparability between the default probability
measures (Correia et al. 2012).40
However, the most surprising fact revealed in the cross-sectional AUROC analysis (see table 12,
Panel A) is the low accuracy of the contingent claims prediction model using the complete sample
spanning from 1980-2010 (AUROC=0.7629), which has multiple reasons. First, we recognize a low
prediction accuracy during the 1990βs increasing stepwise after the dot-com bubble 2000/2001. This
is in line with the findings of Beaver et al. (2005) indicating a decline in the value-relevance of
accounting ratios, while market information becomes more important in explaining default risk.
Second, the contingent-claim model does not compensate inefficient market-information (Bharath
and Shumway 2008; Das et al. 2009), particularly within a sample period enveloping market
downturns. An unreported longitudinal analysis shows a reduced explanation power of the EDF
model, especially during financial crisis 2008-2009. 41 Contrary, the c-prob and c-probu model
perform remarkably well over the complete sample period and clearly outperform the EDF, while the
s-prob model (AUROC=0.8121) predicts financially distressed events comparable to the original c-
prob model (AUROC=0.8109) 5-years prior to a delisting. While we acknowledge that the
performance of the c-probu is slightly better (AUROC=0.8398), recall that the s-prob model uses
accounting information only and could be employed in a private firm setting. Prior findings conducted
by Bauer and Agarwal (2014) imply a analogous ranking performance.42
39 The results and specification of the models are given in the Appendix 2 and 3. 40 For example we require at least 60 trading days of stock price returns prior to the month the implied asset volatility is
calculated, which is not available in 65,717 firm quarters with corresponding accounting distress measures. Note that the limiting market information are not been necessary in previous analyses.
41 In contrast, Correia et al. (2012) find that the βpredictive power of the different models for future credit returns is greater for bust periods.β
42 Bauer and Agarwal (2014) analyze the receiver operating characteristic using the parametrization of Shumway (2001), Campbell et al. (2008) Model 1, the contingent claims framework following Bharath and Shumway (2008) and Tafflerβs z-score model. The conducted AUROC comparison implies that Campbell et al. (2008) β parametrization is
35
Table 13 contains the corresponding information content analysis. As shown in Table 13, Panel A-C
(Model 6 vs. Model 8) the coefficient on the EDF model loses statistically significance (3-year and
5-year) when the s-score estimates are added to the a combined specification using the c-scoreu while
the s-score coefficient remains significantly for the 1-year and 3-year ahead prediction horizons. In
addition, the the Vuong statistic in Panel D demonstrates, that the EDF measure loses its relevance
as the forecast horizon is extended to 5-years ahead, while the c-scoreu model has the most
information content due to its skillful combination of accounting and market data.
------------------------- Please insert Table 11 about here -------------------------
------------------------- Please insert Table 12 about here -------------------------
------------------------- Please insert Table 13 about here -------------------------
(ii) Alternative Prediction Horizon and Specifications
Finally, with respect to the long-term prediction horizon of this study we extent the forecast horizon
of the re-estimation procedure (z-probu and o-probu) to predict financial distress 5-years prior to a
performance-related delisting.43 Balcaen and Ooghe (2006) point out that the horizon to re-estimation
or re-development on observations should meet the horizon of the predictive classification statements
(financially distressed/solvent) with reference to the parametric nature of the statistical framework.
For the sake of brevity, we have not presented these comprehensive results here (available upon
request). The corresponding results for the alternative z-probu and o-probu are (AUROC=0.7273 and
AUROC=0.7294). The findings indicate that an alternative estimation horizon to estimate the related
coefficients does not influence the accuracy of the statistical models. In addition, we test several
alternative specifications for the z- and o-score models reported in the previous literature (Begley et
superior to the contingent claims framework, even if their sample size covers the period 1985β2009 with 22,217 observations (Bauer and Agarwal 2014, table 5).
43 Traditionally, most studies define a short-term prediction horizon (i.e., less than 16 month in the future) to calibrate their distress/default prediction models. For example the original z-score model by Altman (1968) was developed defining bankruptcy firms having an average lead time of the financial statements of 7 Β½ months. Ohlson (1980) report an average lead time between the date of the fiscal year of the last relevant report and bankruptcy of 13 months. According to Crosbie and Bohn (2003) the time horizon specification for the contingent-claims model is regularly set to T=1 year even if the outcomes are used in predict distress/default for different forecast horizon.
36
al. 1996; Hillegeist et al. 2004), but do not find significant improvements in the ability to correctly
classifying financially distressed firms compared to the updated model versions.
(iii) Alternative Distress Barrier
With reference to the asset/debt implied boundary condition we employ industry-specific and
leverage-specific boundary conditions and test Parisian option barriers (Reisz and Perlich 2007) to
indicate a firm as financially distressed, but do not find significantly improvements in the
discriminatory power. Finally, we conduct an practical asset-implied barrier that equals 30% of the
firmβs average OLR 2-years prior a performance-related delisting event and obtained slightly
improvements in the cross-sectional discrimination power for the s-prob model (AUROC=0.8130).
However, since this criterion is somewhat arbitrary, we do not enhance our model in this regard.
37
CONCLUSION
Projecting the financial health is a challenging and important task. Recent research by Shumway
(2001), Chava and Jarrow (2004) and Campbell et al. (2008) demonstrate the effort to develop more
refined models. However, these models mostly require market data, which are not available for the
majority of firms. In this study, we present a new approach to predict corporate bankruptcy based on
stochastic processes (s-score). The novel framework for bankruptcy prediction is based on recent
research in company valuation and has several advantages over standard statistical models. First, it is
theoretically well grounded and can address the problem of the backward-looking perspective of
accounting-based models. Second, it demands less data than the classical statistical models (such as
probit, logit and MDA approaches). Most importantly, however, it explicitly considers accounting
volatility, which is an important factor of early financial distress warning.
From a detailed examination we find the following results. First, the bankruptcy prediction based on
stochastic processes fits the distribution of historic default rates reasonably well and thus provides
early warning signals. Second, the bankruptcy prediction based on stochastic processes is more
accurate in distinguishing between non-delisting and delisting firms than the prominent z- or o-score
models. Third, this model significantly outperforms the two statistical models for longer time
horizons, which is probably due to the explicit incorporation of accounting volatility measures.
Fourth, while the prominent z- or o-score models perform comparably well, a combination of all three
models has a significantly greater explanatory power.
As a common critique to structural models, one possible drawback of our model is the strong
underlying assumptions (Xu and Zhang 2009). The assumptions about firmsβ cash generating
operating processes might be too simplified. However, the model successfully addresses several
drawbacks of the statistical bankruptcy models and offers solid results. Obviously, the financial
distress prediction based on stochastic processes has many degrees of freedom. It can therefore be
regarded as a novel model class whose flexibility is a strength and an adequate response to the
complex process of financial distress. Future research on this topic could help to model more detailed
financing and investing policies of the firm and incorporate analystsβ forecasts to improve the
initializing parametrization.
38
Appendix
In this section we present the construction and estimation results of the re-estimated accounting-based
(Appendix 1) and market-based financial distress prediction models (Appendix 2: expected default
frequency (EDF) model, Appendix 3: c-score model of Campbell et al. (2008)).
Appendix 1 Summary statistics: Original and re-estimated Altman (1968) and Ohlson (1980)
Panel A: Updated Z-Score Model (z-probu) (firm quarters from 1980Q1 β 2010Q4) N Mean Median Std.dev. Constant 124 (3.24) (3.03) 0.58 X1=WC/TA 124 (1.87) (1.81) 0.36 X2=RE/TA 124 (0.15) (0.17) 0.14 X3=EBIT/TA 124 (10.60) (9.22) 3.82 X4=MVE/TL 124 (0.10) (0.02) 0.27 X5=SA/TA 124 0.08 0.03 0.48 Pseudo R2 124 0.17 0.17 0.01
Panel B: Updated O-Score Model (o-probu) (firm quarters from 1980Q1 β 2010Q4) N Mean Median Std.dev. Constant 124 (0.33) (0.27) 0.16 X1=SIZE 124 (0.45) (0.45) 0.09 X2=TLTA 124 1.85 1.80 0.35 X3=WCTA 124 (1.13) (1.26) 0.65 X4=CLCA 124 (0.04) (0.10) 0.21 X5=NITA 124 (0.70) (0.86) 0.59 X6=FULT 124 (0.59) (0.42) 0.70 X7=INTWO 124 1.30 1.14 0.45 X8=OENEG 124 0.31 0.41 1.41 X9=CHIN 124 (1.41) (1.41) 0.34 Pseudo R2 124 0.23 0.24 0.03 Panel A reports summary statistics for the updated z-score model (estimated by a growing window logit regression model, starting in firm quarter 1978Q1). For example, we estimate the updated coefficients to calculate the z-score in 1980Q1 using firm quarter observations (X1-X5) from 1978Q1-1979Q3 and delisting information (CRSP delisting codes 400, 500-585) from 1979Q4 (out-of-sample-approach). We list the mean, median and standard deviation of the estimated coefficients by summarizing the main sample period of 1980Q1-2010Q4 (=124 firm quarters). The dependent variable in the logit regression model equals one if a delisting occurs within the next 4 quarters. All estimated coefficients have the expected signs. We winsorize all measures at the 1st and 99th percentile to avoid outliers driving the results. The z-scores are transformed into probabilities using the standard logistic function (z-probu=1/(1+exp(-z-scoreu)). Panel B reports summary statistics for the updated o-score model (estimated by a growing logit regression model, starting in firm quarter 1978Q1). As in Panel A we ensure all measures are observable at that quarter over which the updated probabilities are calculated (out-of-sample-approach). The dependent variable in the logit regression model equals one if a delisting occurs within the next 4 quarters. We winsorize all measures at the 1st and 99th percentile to avoid outliers driving the results. Consistent with Begley et al. (1996), Grice and Dugan (2003) and Hillegeist et al. (2004) we find that the coefficients are not stable across different time periods. The o-scores are transformed into probabilities using the standard logistic function (o-probu=1/(1+exp(-o-scoreu)). WC/TA = working capital/total assets; RE/TA = retained earnings/total assets; EBIT/TA = earnings before interest and taxes/total assets; MVE/TL = market capitalization/total liabilities; SA/TA = sales/total assets; SIZE = ln(total assets/ lagged GNP level (with base year 1968)); TLTA= total liabilities/total assets; CLCA= current liabilities/current assets; NITA= net income/total assets; FULT=pretax income/total liabilities; INTWO = 1 if the net income in the last two quarter were negative and zero otherwise; OENEG= 1 if total liabilities > total assets; CHIN=(net incomet β net incomet-1)/(|net incomet|+|net incomet-1|); Pseudo-R2= defined by McFaddenβs (1974)-Pseudo-R2: (1-L1/L0).
39
Appendix 2
Results of the re-estimated EDF Measure
The contingent claims-based framework is a widely used operationalization of a market-based
prediction model that requires (daily) market- and accounting observations of: (1) the total firm value
of assets (V=iteratively inferred), (2) the related volatility of the firm assets (ππππ=iteratively inferred),
(3) the face value of total debt maturing at time T (B=debt in current liabilities plus one-half of the
long-term interest bearing debt), (4) the value of equity (E=market value of equity from CRSP), (5)
the expected return on assets (ΞΌ=iteratively inferred), (6) the prediction horizon T, and (7) the risk-
free rate (r=1-year US-Treasury Bill Rate), as input parameters. The underlying assumptions of the
contingent claims framework are: (i) the total firm value of assets V follows a geometric Brownian
motion and (ii) the total debt, i.e., the total claims of a firm, maturing at time T.44 In order to receive
estimates for the firm value of assets V and volatility of assets ππππ, we operationalize the sequential-
iterations algorithm as described by Crosbie and Bohn (2003), Vassalou and Xing (2004), Bharath
and Shumway (2008) and more recently in Jessen and Lando (2015) with a maximum of 15 iterations.
The value of equity, viewed as a European call option is obtained from:
1 2E( ; ) ( ) ( )rTV T VN d Be N dβ= β (19)
2
1ln( / ) ( 0.5 )V
V
V B TdT
Β΅ ΟΟ+ +
= (20)
2
2 1ln( / ) ( 0.5 )V
vV
V B Td d TT
Β΅ ΟΟΟ+ β
= β = (21)
where we require, similar to Bharath and Shumway (2008), a forecast horizon of T=1 year and at least
60 trading days over the previous 12 months. Finally, the probability of default EDF is given by:
2 2
2
P(V ) 1 ( ) ( d )
ln( / ) ( 0.5 )T
V
V
EDF B N d N
V B TNT
Β΅ ΟΟ
= < = β = β
+ β= β
(22)
44 To be precise, the theoretical underpinning of the Black-Scholes model requires additional assumptions, for example
lognormally distributed returns, constant volatilities and other related market frictions etc. (Agarwal and Taffler 2008).
40
where N(β) is the cumulative standard normal distribution function. 45 We winsorize all model
parameters at the 1st and 99th percentile to avoid outliers driving the results.
45 Additionally, we compare our estimated EDF-measure with the aggregated default likelihood indicator presented in
Vassalou and Xing (2004). We thank the authors Maria Vassalou and Yuhang Xing to make their results available at link: http://maria-vassalou.com/research/data. Based on the merged 146,189 firm quarter observation with nonmissing information from 1980Q-1999Q4 we compare the AUROC of both measures. The findings indicate, as expected, that the performance of implied default probability following Bharath and Shumway (2008) and the default probability by Vassalou and Xing (2004) are not significantly different from each other (AUROCBharath/Shumway = 0.7650 vs. AUROCVassalou/Xing = 0.7651). The test is based on an indicator variable for financial distress that equals one if a delisting (indicated by CRSP delisting codes 400, 500-585) occurs within the next 20 quarters (i.e., a delisting firm quarter is determined if the firm experience a delisting within the next 20 quarters as defined in table 1).
41
Appendix 3
Results of the re-estimated C-Score Model No. 2 of Campbell et al. (2008)
Updated C-Score Coefficients (c-probu) (firm month from 1980|01 β 2010|12)
N Mean Median Std.dev. Constant 372 (14.86) (14.49) 2.77 X1=NIMTAAVG 372 (8.02) (6.51) 3.02 X2=TLMTA 372 1.49 1.44 0.75 X3=EXRETAVG 372 (0.58) (0.69) 1.56 X4=SIGMA 372 0.72 0.49 4.10 X5=RSIZE 372 (0.59) (0.59) 0.12 X6=CASHMTA 372 (1.58) (1.15) 0.93 X7=MB 372 0.07 0.07 0.03 X8=PRICE 372 (0.60) (0.60) 0.11 Pseudo R2 372 0.28 0.30 0.03 This table reports summary statistics for the re-estimated c-score model (estimated by a growing window logit regression model, starting in firm month 1976|01). For example, we estimate the updated coefficients to calculate the c-score in January of 1980 using firm quarter observations (X1-X8) as defined by the Appendix of Campbell et al. (2008) from January, 1976-November, 1979 and monthly delisting information (CRSP delisting codes 400, 500-585) from December, 1979 (out-of-sample-approach). We list the mean, median and standard deviation of the coefficients by summarizing the main sample period of 1980|01-2010|12 (=124 firm quarters). The dependent variable in the logit regression model equals one if a delisting occurs within the next month. Consistent with the alternative failure model (Model 1) of Campbell et al. (2008) and the re-estimated version of Bauer and Agarwal (2014) the coefficients have the expected signs. We winsorize all measures at the 1st and 99th percentile to avoid outliers driving the results. The c-scores are transformed into probabilities using the standard logistic function (c-probu=1/(1+exp(-c-scoreu)). Pseudo R2= defined by McFaddenβs (1974)-Pseudo R2: (1-L1/L0).
42
References
Abarbanell, J. S., and B. J. Bushee. 1997. Fundamental Analysis, Future Earnings, and Stock Prices.
Journal of Accounting Research 35 (1):1-24.
Acharya, V., S. A. Davydenko, and I. A. Strebulaev. 2012. Cash Holdings and Credit Risk. Review
of Financial Studies 25 (12):3572-3609.
Acharya, V. V., H. Almeida, and M. Campello. 2013. Aggregate Risk and the Choice between Cash
and Lines of Credit. The Journal of Finance 68 (5):2059-2116.
Agarwal, V., and R. Taffler. 2008. Comparing the Performance of Market-Based and Accounting-
Based Bankruptcy Prediction Models. Journal of Banking & Finance 32 (8):1541-1551.
Altman, E. I. 1968. Financial Ratios, Discriminant Analysis and the Prediction of Corporate
Bankruptcy. The Journal of Finance 23 (4):589-609.
Altman, E. I. 1983. Corporate Distress: A Complete Guide to Predicting, Avoiding and Dealing with
Bankruptcy. New York.
Altman, E. I. 1984. A Further Empirical Investigation of the Bankruptcy Cost Question. The Journal
of Finance 39 (4):1067-1089.
Altman, E. I. 2000. Predicting Financial Distress of Companies: Revisiting the Z-Score and Zeta
Models. Stern School of Business, New York University:9-12.
Altman, E. I., R. G. Haldeman, and P. Narayanan. 1977. Zeta Tm Analysis a New Model to Identify
Bankruptcy Risk of Corporations. Journal of Banking & Finance 1 (1):29-54.
Altman, E. I., and A. Saunders. 1998. Credit Risk Measurement: Developments over the Last 20
Years. Journal of Banking and Finance 21 (11/12).
Altomonte, C., and E. Pennings. 2008. Productivity Growth and the Speed of Convergence of
Domestic Firms.
Anderson, R. W., and A. Carverhill. 2012. Corporate Liquidity and Capital Structure. Review of
Financial Studies 25 (3):797-837.
Andrade, G., and S. N. Kaplan. 1998. How Costly Is Financial (Not Economic) Distress? Evidence
from Highly Leveraged Transactions That Became Distressed. The Journal of Finance 53
(5):1443-1493.
Aziz, M. A., and H. A. Dar. 2006. Predicting Corporate Bankruptcy: Where We Stand? Corporate
Governance: The international journal of business in society 6 (1):18-33.
43
Balcaen, S., and H. Ooghe. 2006. 35 Years of Studies on Business Failure: An Overview of the Classic
Statistical Methodologies and Their Related Problems. The British Accounting Review 38
(1):63-93.
Banker, R. D., and L. Chen. 2006. Predicting Earnings Using a Model Based on Cost Variability and
Cost Stickiness. The Accounting Review 81 (2):285-307.
Barth, M. E., D. P. Cram, and K. K. Nelson. 2001. Accruals and the Prediction of Future Cash Flows.
The Accounting Review 76 (1):27-58.
Bates, T. W., K. M. Kahle, and R. M. Stulz. 2009. Why Do Us Firms Hold So Much More Cash Than
They Used To? The Journal of Finance 64 (5):1985-2021.
Batta, G., and W. Wan. 2014. The Impact of Equity Misvaluation on Predictive Accuracy of
Bankruptcy Models. Journal of Fixed Income 24 (2):5-18-18.
Bauer, J., and V. Agarwal. 2014. Are Hazard Models Superior to Traditional Bankruptcy Prediction
Approaches? A Comprehensive Test. Journal of Banking & Finance 40:432-442.
Beaver, W., M. McNichols, and R. Price. 2007. Delisting Returns and Their Effect on Accounting-
Based Market Anomalies. Journal of Accounting and Economics 43 (2-3):341-368.
Beaver, W. H. 1966. Financial Ratios as Predictors of Failure. Journal of Accounting Research 4:71-
111.
Beaver, W. H. 1968a. Alternative Accounting Measures as Predictors of Failure. The Accounting
Review 43 (1):113-122.
Beaver, W. H. 1968b. Market Prices, Financial Ratios, and the Prediction of Failure. Journal of
Accounting Research 6 (2):179-192.
Beaver, W. H., M. Correia, and M. F. McNichols. 2012. Do Differences in Financial Reporting
Attributes Impair the Predictive Ability of Financial Ratios for Bankruptcy? Review of
Accounting Studies 17 (4):969-1010.
Beaver, W. H., M. F. McNichols, and J.-W. Rhie. 2005. Have Financial Statements Become Less
Informative? Evidence from the Ability of Financial Ratios to Predict Bankruptcy. Review of
Accounting Studies 10 (1):93-122.
Begley, J., J. Ming, and S. Watts. 1996. Bankruptcy Classification Errors in the 1980s: An Empirical
Analysis of Altman's and Ohlson's Models. Review of Accounting Studies 1 (4):267-284.
Bharath, S. T., and T. Shumway. 2008. Forecasting Default with the Merton Distance to Default
Model. Review of Financial Studies 21 (3):1339-1369.
44
Black, F., and M. Scholes. 1973. The Pricing of Options and Corporate Liabilities. The Journal of
Political Economy 81 (3):637-654.
BlΓΆchlinger, A., and M. Leippold. 2006. Economic Benefit of Powerful Credit Scoring. Journal of
Banking & Finance 30 (3):851-873.
Blum, M. 1974. Failing Company Discriminant Analysis. Journal of Accounting Research 12 (1):1-
25.
Brockman, P., and H. J. Turtle. 2003. A Barrier Option Framework for Corporate Security Valuation.
Journal of Financial Economics 67 (3):511-529.
Campbell, J. Y., J. Hilscher, and J. Szilagyi. 2008. In Search of Distress Risk. The Journal of Finance
63 (6):2899-2939.
Campbell, J. Y., J. D. Hilscher, and J. Szilagyi. 2011. Predicting Financial Distress and the
Performance of Distressed Stocks. Journal of Investment Management 9 (2):1-21.
Caskey, J., J. Hughes, and J. Liu. 2012. Leverage, Excess Leverage, and Future Returns. Review of
Accounting Studies 17 (2):443-471.
Charitou, A., D. Dionysiou, N. Lambertides, and L. Trigeorgis. 2013. Alternative Bankruptcy
Prediction Models Using Option-Pricing Theory. Journal of Banking & Finance 37 (7):2329-
2341.
Chava, S., and R. A. Jarrow. 2004. Bankruptcy Prediction with Industry Effects. Review of Finance
8 (4):537-569.
Chen, F., O.-K. Hope, Q. Li, and X. Wang. 2011. Financial Reporting Quality and Investment
Efficiency of Private Firms in Emerging Markets. The Accounting Review 86 (4):1255-1288.
Correia, M., S. Richardson, and Δ°. Tuna. 2012. Value Investing in Credit Markets. Review of
Accounting Studies 17 (3):572-609.
Crosbie, P., and J. Bohn. 2003. Modeling Default Risk. Modeling Methodology, Moodyβs KMV.
Dambolena, I. G., and S. J. Khoury. 1980. Ratio Stability and Corporate Failure. The Journal of
Finance 35 (4):1017-1026.
Das, S. R., P. Hanouna, and A. Sarin. 2009. Accounting-Based Versus Market-Based Cross-Sectional
Models of Cds Spreads. Journal of Banking & Finance 33 (4):719-730.
Davis, S. J., J. Haltiwanger, R. Jarmin, and J. Miranda. 2007. Volatility and Dispersion in Business
Growth Rates: Publicly Traded Versus Privately Held Firms. In Nber Macroeconomics
Annual 2006, Volume 21: MIT Press, 107-180.
45
Davydenko, S. A. 2012. When Do Firms Default? A Study of the Default Boundary. In A Study of
the Default Boundary (November 2012). EFA Moscow Meetings Paper.
Deakin, E. B. 1972. A Discriminant Analysis of Predictors of Business Failure. Journal of Accounting
Research 10 (1):167-179.
Dechow, P. M. 1994. Accounting Earnings and Cash Flows as Measures of Firm Performance: The
Role of Accounting Accruals. Journal of Accounting and Economics 18 (1):3-42.
DeLong, E. R., D. M. DeLong, and D. L. Clarke-Pearson. 1988. Comparing the Areas under Two or
More Correlated Receiver Operating Characteristic Curves: A Nonparametric Approach.
Biometrics 44 (3):837-845.
Dichev, I. D. 1998. Is the Risk of Bankruptcy a Systematic Risk? The Journal of Finance 53 (3):1131-
1147.
Dichev, I. D., and V. W. Tang. 2009. Earnings Volatility and Earnings Predictability. Journal of
Accounting and Economics 47 (1-2):160-181.
Dimitras, A. I., S. H. Zanakis, and C. Zopounidis. 1996. A Survey of Business Failures with an
Emphasis on Prediction Methods and Industrial Applications. European Journal of
Operational Research 90 (3):487-513.
Donelson, D. C., and R. J. Resutek. 2015. The Predictive Qualities of Earnings Volatility and Earnings
Uncertainty. Review of Accounting Studies 20 (1):470-500.
Duffie, D., and D. Lando. 2001. Term Structures of Credit Spreads with Incomplete Accounting
Information. Econometrica 69 (3):633-664.
Duffie, D., L. Saita, and K. Wang. 2007. Multi-Period Corporate Default Prediction with Stochastic
Covariates. Journal of Financial Economics 83 (3):635-665.
Eisenbeis, R. A. 1977. Pitfalls in the Application of Discriminant Analysis in Business, Finance, and
Economics. The Journal of Finance 32 (3):875-900.
Emery, G. W., and K. O. Cogger. 1982. The Measurement of Liquidity. Journal of Accounting
Research 20 (2):290-303.
Falkenstein, E. G., A. Boral, and L. V. Carty. 2000. Riskcalc for Private Companies: Moody's Default
Model. Moodyβs Investors Service, As published in Global Credit Research, May.
Favara, G., E. Schroth, and P. Valta. 2012. Strategic Default and Equity Risk across Countries. The
Journal of Finance 67 (6):2051-2095.
Francis, J., R. LaFond, P. Olsson, and K. Schipper. 2005. The Market Pricing of Accruals Quality.
Journal of Accounting and Economics 39 (2):295-327.
46
Francis, J., and K. Schipper. 1999. Have Financial Statements Lost Their Relevance? Journal of
Accounting Research 37 (2):319-352.
Franzen, L. A., K. J. Rodgers, and T. T. Simin. 2007. Measuring Distress Risk: The Effect of R&D
Intensity. The Journal of Finance 62 (6):2931-2967.
Fu, W. H.-W. 2009. Stochastic Modeling for Financial Distress and Bankruptcy in a Small Business,
University of Manitoba.
Garlappi, L., T. Shu, and H. Yan. 2008. Default Risk, Shareholder Advantage, and Stock Returns.
Review of Financial Studies 21 (6):2743-2778.
Garlappi, L., and H. Yan. 2011. Financial Distress and the CrossβSection of Equity Returns. The
Journal of Finance 66 (3):789-822.
Grice, J. S., and M. T. Dugan. 2003. Re-Estimations of the Zmijewski and Ohlson Bankruptcy
Prediction Models. Advances in Accounting 20:77-93.
Hackbarth, D., R. Haselmann, and D. Schoenherr. 2015. Financial Distress, Stock Returns, and the
1978 Bankruptcy Reform Act. Review of Financial Studies 28 (6):1810-1847.
Hanley, J. A., and B. J. McNeil. 1982. The Meaning and Use of the Area under a Receiver Operating
Characteristic (Roc) Curve. Radiology 143 (1):29-36.
Hanley, J. A., and B. J. McNeil. 1983. A Method of Comparing the Areas under Receiver Operating
Characteristic Curves Derived from the Same Cases. Radiology 148 (3):839-843.
Hillegeist, S. A., E. K. Keating, D. P. Cram, and K. G. Lundstedt. 2004. Assessing the Probability of
Bankruptcy. Review of Accounting Studies 9 (1):5-34.
Hilscher, J., and M. I. Wilson. 2013. Credit Ratings and Credit Risk: Is One Measure Enough? Paper
read at AFA 2013 San Diego Meetings Paper.
Hope, O.-K., W. B. Thomas, and D. Vyas. 2013. Financial Reporting Quality of U.S. Private and
Public Firms. The Accounting Review 88 (5):1715-1742.
Jackson, R. H. G., and A. Wood. 2013. The Performance of Insolvency Prediction and Credit Risk
Models in the Uk: A Comparative Study. The British Accounting Review 45 (3):183-202.
Jessen, C., and D. Lando. 2015. Robustness of Distance-to-Default. Journal of Banking & Finance
50:493-505.
Jones, S., and D. A. Hensher. 2004. Predicting Firm Financial Distress: A Mixed Logit Model. The
Accounting Review 79 (4):1011-1038.
Kahya, E. 1997. Prediction of Business Failure: A Funds Flow Approach. Managerial Finance 23
(3):64-71.
47
Kapadia, N. 2011. Tracking Down Distress Risk. Journal of Financial Economics 102 (1):167-182.
Klobucnik, J. 2013. Company Valuation and Bankruptcy Prediction Forecasting of Companies'
Future Perfomance through Analyst Assessments and Stochastic Modeling. Ph.D dissertation,
Accounting, University of Cologne.
Klobucnik, J., and S. Sievers. 2013. Valuing High Technology Growth Firms. Journal of Business
Economics 83 (9):947-984.
Lechner, M. 1991. Testing Logit Models in Practice. Empirical Economics 16 (2):177-198.
Leland, H. E., and K. B. Toft. 1996. Optimal Capital Structure, Endogenous Bankruptcy, and the
Term Structure of Credit Spreads. The Journal of Finance 51 (3):987-1019.
Lins, K. V., H. Servaes, and P. Tufano. 2010. What Drives Corporate Liquidity? An International
Survey of Cash Holdings and Lines of Credit. Journal of Financial Economics 98 (1):160-
176.
Luoma, M., and E. Laitinen. 1991. Survival Analysis as a Tool for Company Failure Prediction.
Omega 19 (6):673-678.
Macey, J., M. O'Hara, and D. Pompilio. 2008. Down and out in the Stock Market: The Law and
Economics of the Delisting Process. Journal of Law and Economics 51 (4):683-713.
Martin, D. 1977. Early Warning of Bank Failure: A Logit Regression Approach. Journal of Banking
& Finance 1 (3):249-276.
Mensah, Y. M. 1984. An Examination of the Stationarity of Multivariate Bankruptcy Prediction
Models: A Methodological Study. Journal of Accounting Research 22 (1):380-395.
Merton, R. C. 1974. On the Pricing of Corporate Debt: The Risk Structure of Interest Rates. The
Journal of Finance 29 (2):449-470.
Merwin, C. L. 1942. Financing Small Corporations in Five Manufacturing Industries, 1926-36. NBER
Books.
Mossman, C. E., G. G. Bell, L. M. Swartz, and H. Turtle. 1998. An Empirical Comparison of
Bankruptcy Models. Financial Review 33 (2):35-54.
Nissim, D., and S. Penman. 2001. Ratio Analysis and Equity Valuation: From Research to Practice.
Review of Accounting Studies 6 (1):109-154.
Ohlson, J. A. 1980. Financial Ratios and the Probabilistic Prediction of Bankruptcy. Journal of
Accounting Research 18 (1):109-131.
Opler, T., L. Pinkowitz, R. Stulz, and R. Williamson. 1999. The Determinants and Implications of
Corporate Cash Holdings. Journal of Financial Economics 52 (1):3-46.
48
Opler, T. C., and S. Titman. 1994. Financial Distress and Corporate Performance. The Journal of
Finance 49 (3):1015-1040.
PΓ‘stor, Δ½., and P. Veronesi. 2003. Stock Valuation and Learning About Profitability. The Journal of
Finance 58 (5):1749-1790.
PΓ‘stor, Δ½., and P. Veronesi. 2006. Was There a Nasdaq Bubble in the Late 1990s? Journal of
Financial Economics 81 (1):61-100.
Pepe, M., G. Longton, and H. Janes. 2009. Estimation and Comparison of Receiver Operating
Characteristic Curves. The Stata Journal 9 (1):1.
Platt, H. D., and M. B. Platt. 1990. Development of a Class of Stable Predictive Variables: The Case
of Bankruptcy Prediction. Journal of Business Finance & Accounting 17 (1):31-51.
Platt, H. D., and M. B. Platt. 2002. Predicting Corporate Financial Distress: Reflections on Choice-
Based Sample Bias. Journal of Economics and Finance 26 (2):184-199.
Pompe, P. P. M., and J. Bilderbeek. 2005. The Prediction of Bankruptcy of Small- and Medium-Sized
Industrial Firms. Journal of Business Venturing 20 (6):847-868.
Ravi Kumar, P., and V. Ravi. 2007. Bankruptcy Prediction in Banks and Firms Via Statistical and
Intelligent Techniques β a Review. European Journal of Operational Research 180 (1):1-28.
Reisz, A. S., and C. Perlich. 2007. A Market-Based Framework for Bankruptcy Prediction. Journal
of financial stability 3 (2):85-131.
Rogoff, K. 1996. The Purchasing Power Parity Puzzle. Journal of Economic Literature 34 (2):647-
668.
Schwartz, E. S., and M. Moon. 2000. Rational Pricing of Internet Companies. Financial Analysts
Journal 56 (3):62-75.
Schwartz, E. S., and M. Moon. 2001. Rational Pricing of Internet Companies Revisited. Financial
Review 36 (4):7-26.
Shumway, T. 1997. The Delisting Bias in Crsp Data. Journal of Finance:327-340.
Shumway, T. 2001. Forecasting Bankruptcy More Accurately: A Simple Hazard Model. The Journal
of Business 74 (1):101-124.
Sloan, R. 1996. Do Stock Prices Fully Reflect Information in Accruals and Cash Flows About Future
Earnings? The Accounting Review 71 (3):289-315.
Smith, R. F., and A. H. Winakor. 1935. Changes in the Financial Structure of Unsuccessful Industrial
Corporations: University of Illinois.
49
Sobehart, J. R., R. Stein, V. Mikityanskaya, and L. Li. 2000. Moodyβs Public Firm Risk Model: A
Hybrid Approach to Modeling Short Term Default Risk. Moodyβs Investors Service, Global
Credit Research, Rating Methodology, March.
Stein, R. M. 2002. Benchmarking Default Prediction Models: Pitfalls and Remedies in Model
Validation. Moodyβs KMV, New York 20305.
Sufi, A. 2009. Bank Lines of Credit in Corporate Finance: An Empirical Analysis. Review of
Financial Studies 22 (3):1057-1088.
Vassalou, M., and Y. Xing. 2004. Default Risk in Equity Returns. The Journal of Finance 59 (2):831-
868.
Warner, J. B. 1977. Bankruptcy Costs: Some Evidence. The Journal of Finance 32 (2):337-347.
Weiss, D., P. A. Naik, and C.-L. Tsai. 2008. Extracting Forward-Looking Information from Security
Prices: A New Approach. The Accounting Review 83 (4):1101-1124.
Whitaker, R. B. 1999. The Early Stages of Financial Distress. Journal of Economics and Finance 23
(2):123-132.
Wruck, K. H. 1990. Financial Distress, Reorganization, and Organizational Efficiency. Journal of
Financial Economics 27 (2):419-444.
Xu, M., and C. Zhang. 2009. Bankruptcy Prediction: The Case of Japanese Listed Companies. Review
of Accounting Studies 14 (4):534-558.
Yun, H. 2009. The Choice of Corporate Liquidity and Corporate Governance. Review of Financial
Studies 22 (4):1447-1475.
Zmijewski, M. E. 1984. Methodological Issues Related to the Estimation of Financial Distress
Prediction Models. Journal of Accounting Research 22:59-82.
50
Table 1: Summary Delisting Statistics
Panel A: Performance-related Delisting Codes and Frequency (N= 330,549 firm quarters from 1980Q1 to 2010Q4)
CRSP delisting code Delisting reasons Unique
firms Firm
quarters %
400 Issue stopped trading as result of company liquidation. 2 20 0.05% 550 Delisted by current exchange - insufficient number of market makers. 156 1,466 3.52% 551 Delisted by current exchange - insufficient number of shareholders. 62 778 1.87% 552 Delisted by current exchange - price fell below acceptable level. 573 7,225 17.34% 560 Delisted by current exchange - insufficient capital, surplus, and/or equity 635 7,317 17.56%
561 Delisted by current exchange - insufficient (or non-compliance with rules of) float or assets. 346 4,003 9.61%
570 Delisted by current exchange - company request (no reason given). 186 2,441 5.86%
573 Delisted by current exchange - company request, deregistration (gone private). 44 665 1.60%
574 Delisted by current exchange - bankruptcy, declared insolvent. 369 4,938 11.85%
575 Delisted by current exchange - company request, offer rescinded, issue withdrawn by underwriter. 2 24 0.06%
580 Delisted by current exchange - delinquent in filing, non-payment of fees. 397 4,162 9.99%
581 Delisted by current exchange - failure to register under 12G of Securities Exchange Act. 37 431 1.03%
582 Delisted by current exchange - failure to meet exception or equity requirements. 94 1,107 2.66%
583 Delisted by current exchange - denied temporary exception requirement. 1 8 0.02%
584 Delisted by current exchange - does not meet exchangeβs financial guidelines for continued listing. 456 6,179 14.83%
585 Delisted by current exchange - protection of investors and the public interest. 76 900 2.16%
400, 550-585 Delisting - all reasons 3,435 41,664 100.00%
This table lists the financial distress-related delisting reasons and their frequency and percentages of firm quarters in the sample. Generally, CRSP delisting codes 400-499 denote liquidations and 500-599 denote issues dropped from the stock exchange. For the purpose of this study performance-related delisting codes 400 and 550-585 are considered as described in section 4.1. In total there are 3,435 delisted firms (with 41,664 quarters) having distress-related information 20 quarters ahead of delisting. Note that after 1987 CRSP assigned the three digit delisting code β500β to the delisting category prior coded β5β (Shumway 1997).
(continued on next page)
51
(Table 1 continued)
Panel B: Distribution of delistingβs (N= 330,549 firm quarters from 1980Q1 to 2010Q4)
Year # Traded firm's
# Delisted firm's
(%) Delisting rate
# Firm quarters
# distressed firm
quarters
(%) Delisting firm
quarter rate 1980 1,652 3 0.2% 5,774 124 2.1% 1981 1,602 9 0.6% 5,519 141 2.6% 1982 2,238 6 0.3% 6,156 251 4.1% 1983 2,805 12 0.4% 8,908 752 8.4% 1984 2,954 55 1.9% 10,059 1,068 10.6% 1985 3,052 81 2.7% 10,303 1,172 11.4% 1986 3,080 125 4.1% 10,562 1,362 12.9% 1987 3,041 90 3.0% 10,410 1,544 14.8% 1988 3,148 89 2.8% 10,617 1,722 16.2% 1989 3,167 114 3.6% 11,083 1,823 16.4% 1990 3,122 141 4.5% 10,967 1,690 15.4% 1991 3,130 166 5.3% 11,035 1,573 14.3% 1992 3,101 197 6.4% 10,955 1,279 11.7% 1993 3,227 86 2.7% 11,363 1,362 12.0% 1994 3,415 105 3.1% 12,006 1,613 13.4% 1995 3,670 94 2.6% 12,739 1,802 14.1% 1996 3,805 105 2.8% 13,358 2,187 16.4% 1997 3,979 127 3.2% 13,771 2,542 18.5% 1998 4,023 208 5.2% 13,843 2,718 19.6% 1999 3,831 214 5.6% 13,306 2,459 18.5% 2000 3,606 179 5.0% 12,675 2,315 18.3% 2001 3,462 255 7.4% 11,905 1,769 14.9% 2002 3,327 204 6.1% 11,830 1,335 11.3% 2003 3,138 162 5.2% 11,355 974 8.6% 2004 3,000 55 1.8% 11,051 1,040 9.4% 2005 2,911 86 3.0% 10,541 964 9.1% 2006 2,802 37 1.3% 10,243 914 8.9% 2007 2,750 27 1.0% 9,882 1,030 10.4% 2008 2,640 93 3.5% 9,630 935 9.7% 2009 2,561 110 4.3% 9,476 680 7.2% 2010 2,487 63 2.5% 9,227 524 5.7%
This table provides the distribution and summary statistics for the sample period (1980-2010). Traded firms comprise all unique observable firms in the sample, the number of delisted firms is the total number of firms delisted in a specific year and the number of firm quarters equals the total number of observations (traded and financially distressed firms). A firm quarter is considered as delisting firm quarter if the company is delisted in the next 20 quarters ahead. Note that 137 firms are delisted after 2010 (e.g., 2011-2014). In total, the sample consists of 3,435 delisted firms with nonmissing data (i.e., delistings with CRSP delisting codes 400, 550-585) resulting in 41,664 financially distressed firm quarters.
52
Table 2: Estimation of Parameters
No. Label Description Measurement (abbreviations are Compustat mnemonics)
Sales dynamics 1 S = initial sales = quarterly firm sales
2 ππ0 = initial sales growth rate = estimated as smoothed weighted average using the 7 lags of quarterly sales
growth rates: ( )71, where 0.7 ln
107
t saleqt itg and git it saleqittΞ± Ξ±
=β β = =
β=
3 οΏ½Μ οΏ½π = long-term sales growth rate = 0.0075 [per quarter]
4 ππππ,0 = initial volatility of the sales growth rate
= estimated by the root mean squared error (πποΏ½ππ) of a recursive least squares growing (starting in 1978Q1) firm-by-firm regression model (using a smoothed series of moving-averaged filtered quarterly sales ( )MA
saleqt
observations):
( )
;1
( ) ( )
ln1
and
15 1 2 3 4
MA MAg git itit
with firms i across firm quarters t
MAsaleqMA itwhere git MAsaleqit
MAsaleq saleq saleq saleq saleq saleqit it it it it it
Ξ± Ξ² Ξ΅= + +β
=
β
= + + + +β β β β
5 ππππ,0 = initial sales volatility = 1
7
27ln 010
saleqt gsaleqtt
β β β =
6 πππποΏ½ = long-term volatility of sales = 0.05 [per quarter]
Cost dynamics
7 ππ0 = initial variable cost rate = estimated by the slope coefficient (π½π½) of a recursive least squares rolling firm-by-firm regression model (with 7 lagged firm quarters and assuming a zero intercept): ; ( ) andc S wherec cogsq xsgaq S saleqit it it it it it itΞ² Ξ΅= + = + =
8 ππΜ = industry median long-term variable cost
= estimated as industry-specific median of costs (Fama-French (2015) 48-industry classification):
, 1980, , 2010481970
median for tffiT cogs xsgat t
salett= β¦
+β=
9 ππππ,0 = initial volatility of variable costs
= estimated by the root mean squared error (πποΏ½ππ) of a recursive least squares rolling firm-by-firm regression model (with 7 lagged firm quarters):
;1cogsq xsgaqit itvk vk wherevkit it itit
saleqitΞ± Ξ² Ξ΅
+= + + =β
10 πππποΏ½ = long-term volatility of variable costs = 0.05 [per quarter]
(continued on next page)
53
(Table 2 continued)
Other parameters
11 ΞΊ = speed of adjustment
= estimated by the slope coefficient (π½π½) using an industry-specific (Fama-French (2015) 48-industry classification) least squares firm-by-industry regression model (re-estimated each year):
1
1 1 2ln ln1 2 2
2
1980, ..., 2010
sale saleit it
sale sale saleit it ititsale sale saleit it it
saleit
where t
Ξ± Ξ² Ξ΅
β β
ββ β β= + β +ββ β β
β
=
12 tax = tax rate = 0.35
13 ππππ = risk-free rate = 3-Month US-Treasury Bill Rate
14 πππΏπΏπΏπΏπΏπΏ = interest on debt = ( )xint
dlc dltt+
Balance sheet positions
15 OLR = operating liquidity reserve = actq+(unused credit line)a-lctq
16 cash = interest-bearing cash and cash equivalents = cheq
17 LTD = long-term interest-bearing liabilities = dlttq+(additional debt)
COMPUSTAT Quarterly data (q) Annual data (a) item number mnemonic description item
number mnemonic description
#1 xsgaq Selling, General, and Administrative Expenses
#12 sale Sales (Net)
#2 saleq Sales (Net) #15 xint Interest and Related Expense β Total
#30 cogsq Cost of Goods Sold #34 dlc Debt in Current Liabilities
#36 cheq Cash and Equivalents #41 cogs Cost of Goods Sold #37 rectq Receivables β Total #142 dltt Long-Term Debt β Total #38 invtq Inventories β Total #189 xsga Selling, General, and
Administrative Expenses
#39 acoq Current Assets β Other #40 actq Current Assets β Total (as sum of
cheq, acoq, invtq and rectq)
#45 dlcq Debt in Current Liabilities #46 apq Accounts Payable #49 lctq Current Liabilities β Total #51 dlttq Long-Term Debt β Total
(continued on next page)
54
(Table 2 continued) This table presents the estimators for the different model parameters on a quarterly (yearly) basis (using Compustat mnemonics for reference). aThe unused credit line is estimated by (Fama-French (2015) 10-industry classification using linear (OLS) regression:
[ ] [ ] [ ]ln (totalassets ) /0, 1, 2, 3,lineunused cash flow cash M B cashi i i iΞ² Ξ² Ξ² Ξ²= + + β + β as described in section 3.3.
55
Table 3: Initial Parameters
Univariate statistics (N=330,549 firm quarters from 1980Q1 to 2010Q4) No. Status N Mean Median Std.dev. 1% 99%
Sales dynamics 1 initial sales Solvent 288,885 311.85 46.89 1080.23 0.35 4738.00 Distressed 41,664 62.58 7.264 312.40 0.01 862.98 Full sample 330,549 280.43 38.04 1019.30 0.16 4379.00
2 initial sales growth rate Solvent 288,885 0.011 0.010 0.036 -0.104 0.126 Distressed 41,664 0.001 0.002 0.056 -0.154 0.148 Full sample 330,549 0.009 0.009 0.039 -0.126 0.138
3 long-term sales growth rate Solvent 288,885 0.008 0.008 0.000 0.008 0.008 Distressed 41,664 0.008 0.008 0.000 0.008 0.008 Full sample 330,549 0.008 0.008 0.000 0.008 0.008
4 initial sales volatility Solvent 288,885 0.199 0.144 0.187 0.031 1.054 Distressed 41,664 0.335 0.240 0.290 0.043 1.395 Full sample 330,549 0.216 0.153 0.208 0.031 1.207
5 initial volatility of the sales growth rate Solvent 288,885 0.041 0.028 0.045 0.005 0.263 Distressed 41,664 0.069 0.046 0.071 0.007 0.363 Full sample 330,549 0.045 0.030 0.050 0.005 0.318
6 long-term volatility of sales Solvent 288,885 0.050 0.050 0.000 0.050 0.050 Distressed 41,664 0.050 0.050 0.000 0.050 0.050 Full sample 330,549 0.050 0.050 0.000 0.050 0.050 Cost dynamics
7 initial variable cost rate Solvent 288,885 0.899 0.901 0.153 0.423 1.382 Distressed 41,664 1.020 0.978 0.186 0.545 1.382 Full sample 330,549 0.914 0.910 0.162 0.436 1.382
8 industry median long-term variable cost Solvent 288,885 0.903 0.904 0.034 0.791 0.956 Distressed 41,664 0.904 0.907 0.037 0.793 0.957 Full sample 330,549 0.903 0.904 0.035 0.791 0.956
9 initial volatility of variable costs Solvent 288,885 0.096 0.030 0.325 0.003 1.521 Distressed 41,664 0.318 0.078 0.714 0.005 3.775 Full sample 330,549 0.124 0.033 0.403 0.003 2.556
10 industry median long-term volatility of variable costs Solvent 288,885 0.050 0.050 0.000 0.050 0.050
Distressed 41,664 0.050 0.050 0.000 0.050 0.050 Full sample 330,549 0.050 0.050 0.000 0.050 0.050
(continued on next page)
56
(table 3 continued)
Other parameters 11 speed of adjustment Solvent 288,885 0.126 0.126 0.031 0.051 0.218 Distressed 41,664 0.124 0.124 0.030 0.054 0.205 Full sample 330,549 0.126 0.126 0.031 0.051 0.217
12 tax rate Solvent 288,885 0.350 0.350 0.000 0.350 0.350 Distressed 41,664 0.350 0.350 0.000 0.350 0.350 Full sample 330,549 0.350 0.350 0.000 0.350 0.350
13 risk-free rate Solvent 288,885 0.012 0.012 0.007 0.000 0.035 Distressed 41,664 0.012 0.012 0.006 0.000 0.025 Full sample 330,549 0.012 0.012 0.007 0.000 0.035
14 interest rate on debt Solvent 288,885 0.022 0.020 0.021 0.000 0.120 Distressed 41,664 0.028 0.023 0.024 0.000 0.120 Full sample 330,549 0.023 0.020 0.021 0.000 0.120 Balance sheet positions
15 operating liquidity reserve Solvent 288,885 190.123 41.885 620.028 1.943 2315.000 Distressed 41,664 39.148 12.068 173.580 1.681 498.370 Full sample 330,549 171.094 32.947 585.053 1.814 2153.375
16 interest-bearing cash and cash equivalents Solvent 288,885 91.070 11.039 288.082 0.005 1710.519 Distressed 41,664 17.509 1.325 103.798 0.000 285.000 Full sample 330,549 81.798 8.572 272.919 0.001 1562.591
17 interest-bearing liabilities Solvent 288,885 289.374 22.464 916.087 0.000 5469.689 Distressed 41,664 107.901 13.086 463.166 0.000 1850.109 Full sample 330,549 266.500 20.186 874.130 0.000 5098.000
18 unused credit line Solvent 288,885 17.500 0.000 69.398 0.000 360.704 Distressed 41,664 4.434 0.000 87.964 0.000 109.664 Full sample 330,549 15.853 0.000 72.132 0.000 338.400
This table provides the summary statistics for the main s-score model variables (see Table 2 for the calculation). All rates are quarterly growth rates and the balance sheet positions are in million dollars. There are 330,549 firm quarter observations over the period 1980Q1 to 2010Q4 (288,885 solvent firm quarters and 41,664 financially distressed firm quarters). Financially distressed firm quarters are indicated if the firm was delisted 20 quarter ahead. For the interest rate on debt 14,799 observations were set to zero as no debt expense was recorded. For the operating liquidity reserve 85,113 values below the barrier b were funded by additional external debt to allow the stochastic processes to be initialized as described in section 4.2. All differences are statistically significant (at the 1%-level) based on a t-test of the means between the solvent and financial distressed sample (we confirm the results by testing the differences of the medians with the Wilcoxon rank-sum test yielding the same results).
57
Figure 1: Distress Barrier Level Over Time
This figure plots the distress barrier b using the median operating liquidity reserve (OLR) in million USD five years prior to a corresponding delisting event, recursively estimated in a 5-year rolling window approach to avoid look-ahead bias.
58
Figure 2: Comparison of Initial Parameters
Panel A:
Panel B:
(continued on next page)
59
(figure 2 continued)
Panel C:
This figures shows the mean accounting volatility measures (see No. 9, No. 4, No. 5 in table 3), mean initial sales growth rate (No. 2), median income statement and balance sheet positions (No. 1, No. 15, No. 16, No. 17) and the other model parameters (No. 7, No. 11, No. 14) along with the trend line for the quarters before a delisting (N=41,664 financially distressed firm quarters). Note that the balance sheet positions of the OLR (OLR=actq+(unused credit line)-lctq) is neither winsorized nor adjusted for initialization of the s-score model to provide an unbiased view.
60
Table 4: S-prob model Forecast Accuracy Panel A: Operating liquidity reserve - forecast accuracy for complete sample (288,885 solvent and 41,664 financially distressed firm quarters from 1980Q1-2010Q4)
Absolute OLR forecast inaccuracy (in Millions of Dollars) (OLRpredict β OLRactual)
Relative OLR forecast inacurracy (OLRpredict - OLRactual)/ OLRactual
absolute relative Qtr ahead N Mean Median Std.dev. N Mean Median Std.dev.
1 310,934 17.24 0.64 138.31 310,934 0.58 0.03 7.84
2 295,991 34.45 1.81 213.97 295,991 1.19 0.08 14.41
3 284,136 50.71 3.04 280.85 284,136 1.74 0.12 20.62
4 276,218 66.65 4.34 339.59 276,218 2.27 0.17 26.83
5 263,455 81.18 5.58 395.65 263,455 2.79 0.21 33.40
6 254,291 95.67 6.97 445.01 254,291 3.28 0.25 39.39
7 245,568 110.30 8.36 496.55 245,568 3.74 0.29 44.31
8 239,556 126.37 10.00 550.37 239,556 4.21 0.34 49.74
9 229,670 140.45 11.29 602.29 229,670 4.65 0.38 55.39
10 222,391 154.74 12.73 651.64 222,391 5.11 0.41 62.35
11 215,291 168.93 14.18 704.12 215,291 5.58 0.45 69.21
12 210,421 184.98 15.82 767.51 210,421 6.10 0.48 76.43
13 202,166 198.30 16.99 845.43 202,166 6.54 0.51 82.32
14 195,972 213.38 18.50 952.35 195,972 7.12 0.55 109.62
15 189,943 228.41 19.91 1134.33 189,943 7.58 0.57 105.64
16 185,782 246.63 21.74 1447.15 185,782 8.22 0.61 130.33
17 178,734 260.94 23.08 1899.71 178,734 8.76 0.63 169.36
18 173,472 276.99 24.64 2289.29 173,472 9.33 0.66 185.39
19 168,305 293.50 26.19 2689.59 168,305 9.72 0.68 181.77
20 164,741 314.19 28.37 3137.10 164,741 10.28 0.72 203.26
(continued on the next page)
61
(table 4 continued)
Panel B: Operating liquidity reserve - forecast accuracy for distressed sample (41,664 financially distressed firm quarters from 1980Q1-2010Q4)
Absolute OLR forecast inaccuracy (in Millions of Dollars) (OLRpredict β OLRactual)
Relative OLR forecast inacurracy (OLRpredict - OLRactual)/ OLRactual
Qtr ahead N Mean Median Std.dev. N Mean Median Std.dev. 1 38,549 1.91 -0.16 59.79 38,549 0.09 -0.03 2.69
2 35,241 6.10 0.25 82.13 35,241 0.39 0.04 3.39
3 33,242 9.80 0.60 100.91 33,242 0.61 0.09 4.49
4 31,717 13.42 1.01 118.35 31,717 0.82 0.14 5.29
5 29,971 16.86 1.47 138.11 29,971 1.03 0.19 6.03
6 28,597 20.29 2.06 148.66 28,597 1.22 0.25 6.27
7 27,275 24.27 2.68 174.49 27,275 1.46 0.31 8.61
8 26,154 28.21 3.36 201.08 26,154 1.65 0.36 9.88
9 24,894 32.02 3.94 222.51 24,894 1.87 0.42 11.62
10 23,836 35.35 4.67 205.14 23,836 2.05 0.47 11.11
11 22,831 39.36 5.41 221.25 22,831 2.22 0.52 11.17
12 21,998 43.72 6.17 249.41 21,998 2.42 0.57 12.44
13 21,005 48.66 6.81 315.31 21,005 2.66 0.63 14.04
14 20,157 53.94 7.48 430.19 20,157 2.89 0.68 15.60
15 19,338 55.83 8.18 265.53 19,338 3.14 0.73 16.78
16 18,664 60.23 8.88 274.92 18,664 3.36 0.78 17.72
17 17,860 64.18 9.58 282.78 17,860 3.55 0.81 18.11
18 17,187 68.26 10.30 293.91 17,187 3.72 0.86 18.78
19 16,515 73.45 11.06 313.53 16,515 3.92 0.91 19.68
20 15,943 85.40 12.01 756.83 15,943 4.43 0.97 38.98
This table reports the forecasted model inaccuracy of the s-prob model. Panel A reports the forecast inaccuracy for the complete sample (i.e., solvent and financially distressed firm quarters). Panel B restrict the comparison to financially distressed firm quarters. For each of the twenty firm quarters predicted ahead we take the median over all 10,000 simulated (adjusted) OLR paths and compare the forecasted median OLR to the actual OLR (where OLR=actq+lineun-lctq). Next, we calculate the absolute and relative forecast inaccuracy. For example, we compare the actual OLR with the median predicted OLR simulated three quarters ago to yield the three quarters ahead absolute and relative forecast accuracy.
62
Figure 3: Selected Operating Liquidity Reserve-Predictions by Industry
This figure shows the median of OLR-prediction (1, 2, and 3 years ahead) compared to the actual OLR by four randomly selected industries (FFI48-No. 3, 17, 21, and 36) using the Fama-French (2015) 48-industry classification.
63
Table 5: Summary Statistics for OLR by Firm Quarter to Delist
Panel A: OLR by quarter to delisting (41,664 financially distressed firm quarters) OLR by quarter to delist
Qtr to delist Mean Median Std.dev. 1 -4.51 0.07 133.23 2 6.58 0.68 139.16 3 14.79 1.32 106.96 4 20.52 1.98 125.13 5 21.48 2.48 180.40 6 25.03 2.90 135.57 7 29.72 3.28 195.35 8 33.08 3.89 225.85 9 35.21 4.33 203.80
10 34.23 4.78 207.49 11 36.31 5.13 205.32 12 33.97 5.30 267.50 13 34.10 5.63 278.88 14 37.76 6.02 260.61 15 39.77 6.35 253.29 16 43.16 6.64 233.13 17 45.76 6.86 234.26 18 47.32 7.00 236.16 19 44.11 7.23 173.95 20 45.41 7.28 180.37
This table shows the distribution of the operating liquidity reserve (OLR=actq+lineun-lctq) for the twenty quarters ahead of firmβs delisting.
64
Table 6: Descriptive Statistics
Panel A: Univariate summary statistics (solvent vs. financially distressed firm quarters from 1980Q1-2010Q4)
Variable Status Mean Median Std.dev. 1% 99% N
Rating Solvent 10.34 11.00 3.63 2.00 17.00 56,991
Financially Distressed 14.37 14.00 2.29 8.00 21.00 3,827
Z-Prob Solvent 0.16 0.10 0.19 0.00 0.99 288,885
Financially Distressed 0.38 0.31 0.31 0.00 1.00 41,664
O-Prob Solvent 0.05 0.01 0.12 0.00 0.66 288,885
Financially Distressed 0.21 0.10 0.26 0.00 1.00 41,664
S-Prob Solvent 0.18 0.03 0.29 0.00 1.00 288,885
Financially Distressed 0.58 0.70 0.37 0.00 1.00 41,664
Z-Probu Solvent 0.03 0.02 0.05 0.00 0.18 288,885
Financially Distressed 0.08 0.04 0.13 0.00 0.76 41,664
O-Probu Solvent 0.02 0.00 0.05 0.00 0.23 288,885
Financially Distressed 0.06 0.02 0.12 0.00 0.66 41,664
Panel B: Correlations (Pearson above, Spearman rank below the diagonal)
Variable Financial Distressed Rating Z-Prob O-Prob S-Prob Z-Probu O-Probu
Financially Distressed 0.27 0.30 0.24 0.31 0.20 0.22 Rating 0.28 0.58 0.36 0.42 0.28 0.31 Z-Prob 0.26 0.60 0.61 0.58 0.57 0.47 O-Prob 0.26 0.57 0.73 0.53 0.46 0.66 S-Prob 0.24 0.45 0.48 0.53 0.44 0.38 Z-Probu 0.18 0.31 0.67 0.48 0.37 0.50
O-Probu 0.18 0.51 0.42 0.54 0.25 0.52
This table reports summary statistics for solvent vs. financially distressed observations by model. Panel A presents summary statistics of the evaluated model outcomes. RATING (=Standard & Poor's Credit Ratings, coded from 1 (βAAAβ) to 21 (βDβ or "SD") based on the Compustat item splticrm, Z-PROB (=original Altman Z-Score model form 1968). The z-score is transformed into probabilities by the logit function (1/(1+exp(z-score) ), O-PROB (=original Ohlson O-Score model 1 probabilities 1980), S-PROB are the probability outcomes from the financial distressed prediction model using stochastic processes, Z-PROBu (=updated z-score model by a logistic regression). The updated z-scores are transformed into probabilities by the standard logit transformation (1/(1+exp(-z-score) to allow the probabilities to increase with the absolute values of z-scores), O-PROBu (=updated o-score model probabilities by a logistic regression). The total number N of observations are 330,549 firm quarters / 10,687 firms over the sample period 1980Q1 to 2010Q4. FINANCIALLY DISTRESSED represents an indicator variable, which is one if delisting (indicated by CRSP delisting codes 400, 500-585) occurs within the next 20 quarters (i.e., a delisting firm quarter is determined if the firm experience a delisting within the next 20 quarters as defined in table 1). The differences in probabilities of the solvent and financial distressed samples are significant at the 1%-level using a t-test and the Wilcoxon rank sum test (two-tailed). Note that for the Rating variable the sample size is reduced to 56,991 firm quarters / 2,243 firms covered by Standard & Poorβs credit ratings and CRSP delistings. Panel B shows the Pearson correlation above and the Spearman rank correlation below the diagonal for the evaluated models.
65
Figure 4: Median Estimated Default Probability per S&P Rating Class
This figure shows the median estimated distress/default probability per Standard & Poorβs rating class. RATING CLASS (=Standard & Poor's Credit Ratings, based on the Compustat item splticrm). The non-investment grade starts with (βBBβ). For this analysis the sample size is reduced to 60,562 firm quarter observations / 2,118 firms that are covered by the major rating agency Standard & Poorβs. The historic default rate as benchmark is from the 2013 Annual U.S. Corporate Default Study by Standard & Poor's Global Fixed Income Research.
66
Figure 5: Default Probabilities Over Time
This figure shows the median of the estimated distress probabilities for the different models through the period 1980Q1 to 2010Q4 together with the actual delisting rate (where DELISTING RATE = number of delisted firms (CRSP delisting codes 400, 500-585) / number of active firms per quarter). Note that the peak in the distress probability in the o-probu
model during 1982q1 β 1983q1 is attributable to a high proportion of young and small firms entering the sample in 1980-1982 (i.e., approximately an increase by around 700 new unique firms). This leads to a decrease (i) in the median sample firm size (i.e. total assets), and (ii) the firmβs average age. These two effects influence the size-coefficient of the o-probu model significantly.
67
Figure 6: Evolvement of Default Probabilities Before Delisting
This figure shows the evolvement of the median estimated default probabilities for the different models up to 40 quarters before delisting (from 1980Q1 to 2010Q4). The zero represents the quarterly end of an actual listing at one of the following stock exchanges (NYSE, AMEX, NASDAQ or NYSE ARCA) using CRSP delistings codes (dlstcd=400, 500-585).
68
Table 7: Comparative Receiver Operating Characteristic (ROC) and Area Under the Curve (AUROC)
Panel A: Receiver Operating Characteristic (ROC): Area under the Curve (AUROC) by years (from 1980-2010) # firm quarters S-Prob Z-Prob O-Prob Z-Probu O-Probu
Year solvent financially distressed AUROC SE AUROC SE AUROC SE AUROC SE AUROC SE
1980 5,650 124 0.8257 0.0212 0.8293 0.0185 0.8780 0.0158 0.7965 0.0218 0.8422 0.0174
1981 5,378 141 0.7574 0.0235 0.8374 0.0168 0.8720 0.0147 0.7600 0.0210 0.8009 0.0192
1982 5,905 251 0.8195 0.0140 0.7583 0.0188 0.8570 0.0132 0.7556 0.0153 0.8271 0.0141
1983 8,156 752 0.8363 0.0080 0.6742 0.0124 0.8170 0.0082 0.7414 0.0102 0.8131 0.0080
1984 8,991 1,068 0.8470 0.0065 0.7419 0.0094 0.8317 0.0065 0.7774 0.0082 0.8370 0.0064
1985 9,131 1,172 0.8279 0.0064 0.7640 0.0082 0.8266 0.0065 0.7614 0.0081 0.8079 0.0070
1986 9,200 1,362 0.8096 0.0062 0.7389 0.0079 0.8173 0.0060 0.7415 0.0075 0.7941 0.0065
1987 8,866 1,544 0.8037 0.0059 0.7294 0.0076 0.8037 0.0060 0.7365 0.0073 0.7945 0.0060
1988 8,895 1,722 0.8129 0.0054 0.7427 0.0072 0.7975 0.0060 0.7395 0.0070 0.7859 0.0061
1989 9,260 1,823 0.8268 0.0049 0.7562 0.0069 0.8024 0.0057 0.7524 0.0066 0.7963 0.0057
1990 9,277 1,690 0.8301 0.0051 0.7616 0.0070 0.8038 0.0058 0.7461 0.0071 0.8017 0.0055
1991 9,462 1,573 0.8210 0.0056 0.7415 0.0077 0.7925 0.0063 0.7379 0.0075 0.8080 0.0057
1992 9,676 1,279 0.8118 0.0063 0.7160 0.0087 0.7752 0.0072 0.7202 0.0085 0.7999 0.0065
1993 10,001 1,362 0.8100 0.0061 0.6955 0.0086 0.7627 0.0071 0.7302 0.0081 0.7924 0.0064
1994 10,393 1,613 0.8115 0.0055 0.7042 0.0077 0.7576 0.0065 0.7143 0.0074 0.7776 0.0062
1995 10,937 1,802 0.7852 0.0056 0.7018 0.0073 0.7460 0.0064 0.7144 0.0069 0.7563 0.0062
1996 11,171 2,187 0.7940 0.0051 0.6938 0.0067 0.7399 0.0060 0.7135 0.0064 0.7539 0.0057
1997 11,229 2,542 0.7943 0.0049 0.7426 0.0057 0.7589 0.0054 0.7305 0.0057 0.7628 0.0053
1998 11,125 2,718 0.8000 0.0046 0.7411 0.0055 0.7543 0.0052 0.7245 0.0055 0.7452 0.0053
1999 10,847 2,459 0.7980 0.0046 0.7157 0.0062 0.7574 0.0053 0.7272 0.0058 0.7522 0.0054
2000 10,360 2,315 0.8089 0.0048 0.7399 0.0061 0.7574 0.0055 0.7397 0.0060 0.7567 0.0055
2001 10,136 1,769 0.8130 0.0051 0.7794 0.0065 0.7867 0.0059 0.7590 0.0066 0.7664 0.0062
2002 10,495 1,335 0.8096 0.0058 0.7652 0.0074 0.7769 0.0069 0.7505 0.0077 0.7687 0.0070
2003 10,381 974 0.7776 0.0074 0.7308 0.0087 0.7480 0.0081 0.7196 0.0090 0.7524 0.0080
2004 10,011 1,040 0.8019 0.0065 0.7343 0.0086 0.7662 0.0075 0.7256 0.0088 0.7729 0.0076
2005 9,577 964 0.8231 0.0066 0.7502 0.0086 0.7711 0.0079 0.7414 0.0090 0.7764 0.0079
2006 9,329 914 0.8126 0.0072 0.7346 0.0095 0.7745 0.0083 0.7309 0.0096 0.7845 0.0080
2007 8,852 1,030 0.8308 0.0071 0.7542 0.0090 0.8016 0.0077 0.7745 0.0087 0.8092 0.0075
2008 8,695 935 0.8435 0.0072 0.7661 0.0088 0.7965 0.0079 0.7916 0.0085 0.8027 0.0077
2009 8,796 680 0.8283 0.0091 0.7550 0.0105 0.7666 0.0101 0.7656 0.0109 0.7895 0.0093
2010 8,703 524 0.8180 0.0106 0.7588 0.0116 0.7724 0.0109 0.7683 0.0123 0.7940 0.0106 (continued on the next page)
69
(Table 7 continued)
Panel B: Receiver Operating Characteristic (ROC): complete sample
Model: # solvent # financially distressed AUROC SE
S-Prob 288,885 41,664 0.8122 0.0011
Z-Prob 288,885 41,664 0.7327 0.0014 O-Prob 288,885 41,664 0.7818 0.0012 Z-Probu 288,885 41,664 0.7272 0.0014
O-Probu 288,885 41,664 0.7207 0.0013
Panel C: Receiver Operating Characteristic (ROC): bootstrapped samples
Model: MEAN MEDIAN MIN MAX SE S-Prob 0.8123 0.8123 0.8087 0.8160 0.0011 Z-Prob 0.7328 0.7328 0.7280 0.7366 0.0014 O-Prob 0.7819 0.7819 0.7771 0.7855 0.0012 Z-Probu 0.7272 0.7271 0.7225 0.7321 0.0014 O-Probu 0.7207 0.7207 0.7165 0.7253 0.0013
Panel D: Proportion of financially distressed firm quarters outperformed by best model (N=41,664 delisted firm quarters and N=330,549 total firm quarters)
Model: S-Prob Z-Prob O-Prob Z-ProbU O-ProbU
N 39,786 - 1,878 - -
% 95.49% 0.00% 4.51% 0.00% 0.00% This table shows the Receiver Operating Characteristic curve for the evaluated models. Panel A reports the area under Receiver Operating Characteristic curve (AUROC) by years. The area under the ROC curve, AUROC and SE is calculated following the nonparametric approach by DeLong et al. (1988), Hanley and McNeil (1982). By definition a firm quarter is considered as "financially distressed" if delisting occurs within the next 20 quarters (carefully note that CRSP delisting information are only available up to the end of year 2014). Panel B compares the AUROC for the whole sample (1980-2010). The equality of the model performances related to the AUROC is pairwise compared following the U-statistic algorithm by DeLong et al. (1988). The AUROCβs of the S-Prob model are significantly different from other models (pairwise comparison) at a p-value < 0.0001. In Panel C we confirm the statistical inference and sample independence of the ROC curve by employing 1,000 bootstrap replications of the original sample to obtain bootstrap standard errors (SE). Panel D compares the number of firm quarters outperformed by the best model per year.
70
Figure 7: ROC Analysis for Different Forecast Horizons
This figure shows the area under Receiver Operating Characteristic curve (AUROC) for different periods to delisting (in quarters) using the broad sample (1980Q1-2010Q4 with 330,549 firm quarter observations). Depending on the forecast horizon a binary variable is set to one if a delisting occurs within the next 1-24 quarter ahead and zero otherwise. The AUROC is calculated following the nonparametric approach by DeLong et al. (1988), Hanley and McNeil (1982).
71
Figure 8: ROC Analysis over Time
This figure shows the area under Receiver Operating Characteristic curve (AUROC) over time. The AUROC is calculated following the nonparametric approach by DeLong et al. (1988) and Hanley and McNeil (1982) for each single quarter (1980Q1-2010Q4) using a five-year forecast horizon.
.65
.7.7
5.8
.85
.9
1980
q119
81q1
1982
q119
83q1
1984
q119
85q1
1986
q119
87q1
1988
q119
89q1
1990
q119
91q1
1992
q119
93q1
1994
q119
95q1
1996
q119
97q1
1998
q119
99q1
2000
q120
01q1
2002
q120
03q1
2004
q120
05q1
2006
q120
07q1
2008
q120
09q1
2010
q1
subprime crisis S-ProbZ-Prob O-ProbZ-Prob_Upd O-Prob_Upd
72
Table 8: Receiver Operating Characteristic by Industry
Receiver Operating Characteristic (ROC): Area under the Curve (AUROC) by Fama and French 10-Industry classification
Model: S-Prob Z-Prob O-Prob Z-Probu O-Probu FFI10 Desc solvent distressed AUROC SE AUROC SE AUROC SE AUROC SE AUROC SE
1 NoDur 24,247 3,396 0.7878 0.0044 0.7321 0.0051 0.7547 0.0047 0.6959 0.0048 0.7010 0.0048 2 Durbl 10,547 1,733 0.8178 0.0055 0.7431 0.0071 0.7838 0.0060 0.7350 0.0068 0.7380 0.0064 3 Manuf 60,632 6,476 0.8376 0.0025 0.7619 0.0036 0.8164 0.0029 0.7541 0.0034 0.7420 0.0032 4 Enrgy 16,197 2,527 0.8353 0.0043 0.7162 0.0065 0.7771 0.0051 0.6759 0.0067 0.7798 0.0048 5 HiTec 67,159 9,581 0.8296 0.0021 0.7429 0.0030 0.8114 0.0023 0.7577 0.0028 0.7414 0.0028 6 Telcm 6,916 1,273 0.7721 0.0066 0.6728 0.0091 0.6746 0.0084 0.7189 0.0085 0.6841 0.0077 7 Shops 41,363 7,020 0.7924 0.0029 0.7515 0.0033 0.7623 0.0032 0.6943 0.0035 0.6787 0.0035 8 Hlth 26,646 2,910 0.8215 0.0041 0.7333 0.0055 0.7971 0.0042 0.7637 0.0048 0.7366 0.0049 9 Utils 822 62 0.9447 0.0118 0.8583 0.0365 0.8062 0.0377 0.8331 0.0326 0.8788 0.0193
10 Other 32,733 5,398 0.7724 0.0034 0.7050 0.0041 0.7445 0.0037 0.7014 0.0042 0.6865 0.0039 n/a n/a 1,623 1,288 0.7667 0.0089 0.6743 0.0104 0.6793 0.0099 0.7732 0.0086 0.6680 0.0099
This table reports the area under the ROC curve for Fama-French 10-industry classifications (2015). Note that if neither Compustat nor CRSP SIC codes were available we assign these firms to a mixed industry classification. There are 327,638 / 10,563 firm quarter observations / firms with a four digit SIC code classified in the Fama-French 10-industry definition (2015). Similar results are obtained, in untabulated analysis, by Fama-French 48-industry classifications (2015).
73
Table 9: Information content tests
Panel A: (1-year ahead prediction, firm quarters from 1980Q1-2010Q4, 327,451 solvent firm quarters vs. 3,098 financially distressed firm quarters) Variables: Model(1) Model(2) Model(3) Model(4) Model(5) Model(6) Model(7) Model(8) Model(9) Constant -4.616 -4.490 -3.839 -2.170 -2.615 -4.094 -2.542 -3.900 -1.730 (38.43)*** (37.98)*** (26.06)*** (7.85)*** (10.07)*** (25.15)*** (7.05)*** (25.31)*** (5.98)*** coeff S-Score 0.148 . . . . 0.112 0.093 . . (9.85)*** . . . . (6.11)*** (4.75)*** . . coeff Z-Score . 0.176 . . . 0.046 . 0.068 . . (7.27)*** . . . (1.34) . (1.79)** . coeff O-Score . . 0.302 . . 0.153 . 0.243 . . . (8.53)*** . . (3.02)*** . (4.96)*** . coeff Z-Scoreu . . . 0.752 . . 0.336 . 0.459 . . . (9.11)*** . . (3.10)*** . (4.49)*** coeff O-Scoreu . . . . 0.481 . 0.207 . 0.315 . . . . (7.89)*** . (2.64)*** . (4.19)*** Pseudo-RΒ² 0.0831 0.0450 0.0689 0.0745 0.0685 0.1069 0.1164 0.0725 0.0939 LRa 2,132*** 1,155*** 1,768*** 1,913*** 1,758*** 2,743*** 2,989*** 1,861*** 2,411***
(continued on next page)
74
(Table 9 continued)
Panel B: (3-year ahead prediction, firm quarters from 1980Q1-2010Q4, 327,451 solvent firm quarters vs. 3,098 financially distressed firm quarters) Variables: Model(1) Model(2) Model(3) Model(4) Model(5) Model(6) Model(7) Model(8) Model(9) Constant -4.662 -4.654 -4.084 -2.836 -3.329 -4.227 -3.238 -4.086 -2.454 (32.65)*** (28.78)*** (20.99)*** (6.50)*** (8.84)*** (19.77)*** (6.19)*** (20.83)*** (5.57)*** coeff S-Score 0.126 . . . . 0.100 0.091 . . (6.13)*** . . . . (4.04)*** (3.54)*** . . coeff Z-Score . 0.115 . . . 0.004 . 0.004 . . (2.90)*** . . . (0.09) . (0.08) . coeff O-Score . . 0.242 . . 0.137 . 0.239 . . . (5.43)*** . . (2.01)** . (3.93)*** . coeff Z-Scoreu . . . 0.563 . . 0.234 . 0.365 . . . (4.70)*** . . (1.67)* . (2.74)*** coeff O-Scoreu . . . . 0.322 . 0.122 . 0.220 . . . . (4.20)*** . (1.33) . (2.50)** Pseudo-RΒ² 0.0489 0.0139 0.0348 0.0318 0.0283 0.0577 0.0605 0.0348 0.0420 LRa 859*** 244*** 611*** 559*** 498*** 1,014*** 1,064*** 611*** 738***
(continued on next page)
75
(Table 9 continued)
Panel C: (5-year ahead prediction, firm quarters from 1980Q1-2010Q4, 327,451 solvent firm quarters vs. 3,098 financially distressed firm quarters)
Variables: Model(1) Model(2) Model(3) Model(4) Model(5) Model(6) Model(7) Model(8) Model(9) Constant -4.756 -4.854 -4.236 -3.317 -3.660 -4.337 -3.574 -4.227 -2.890 (27.77)*** (22.87)*** (17.37)*** (5.75)*** (7.77)*** (16.39)*** (5.42)*** (17.40)*** (5.03)*** coeff S-Score 0.111 . . . . 0.085 0.082 . . (4.28)*** . . . . (2.74)*** (2.58)*** . . coeff Z-Score . 0.071 . . . -0.023 . -0.028 . . (1.42) . . . (0.49) . (0.56) . coeff O-Score . . 0.223 . . 0.152 . 0.248 . . . (4.00)*** . . (1.81)* . (3.51)*** .
coeff Z-Scoreu . . . 0.450 . . 0.175 . 0.290
. . . (2.97)*** . . (1.08) . (1.85)*
coeff O-Scoreu . . . . 0.267 . 0.109 . 0.198 . . . . (2.94)*** . (1.03) . (1.97)** Pseudo-RΒ² 0.0348 0.0052 0.0255 0.0192 0.0193 0.0417 0.0421 0.0262 0.0279 LRa 437*** 65*** 321*** 242*** 242*** 524*** 529*** 330*** 351***
(continued on next page)
76
(Table 9 continued)
Panel D: Vuong-test statistics
S-Score Z-Score O-Score Z-Scoreu O-Scoreu
Model(1) vs. Model(2) Model(3) Model(4) Model(5)
1-year ahead 12.97*** 4.87*** 3.17*** 5.00*** 3-year ahead 12.17*** 5.74*** 6.79*** 7.14*** 5-year ahead 10.60*** 3.82*** 6.51*** 5.82*** */**/*** asterisks refer to significance at a 10%/ 5%/ 1% level for a one sided t-test. Figures in brackets are the z-statistics. This table shows the coefficients, z-statistics (in parentheses, which accounts for firm dependence between firm quarter observations), McFaddenβs (1974)-Pseudo-R2 and the likelihood ratio statistic (LR=2(L1-L0), where L1 is the maximized log likelihood for the model and L0 is the maximized log likelihood for the baseline model with a constant only using a dynamic logit models as in Chava and Jarrow 2004; Shumway 2001. Due to the panel structure of the data there are less independent observations than assumed by a standard logit regression model. The panels compare the contribution of the s-score estimation with the results from univariate and combined regressions results for 1, 3 and 5-year ahead predictions. The scores are calculated in the same way outlined in table 6 (i.e., original and updated z-score and o-score values (Altman 1968, Ohlson 1980)). The probabilities are converted into scores according to: π π πππ π ππππππ = ππππ[πππππ π ππππ/(1 β πππππ π ππππ)]. We bound all scores between Β±18.420680999. Panel D reports the Vuong LR test statistic results for strictly non-nested models (model(1) vs. model(2-5)). A positive LR test statistic indicates that the s-score model (i.e. model(1)) is preferable. If the LR test statistic is negative the compared model is favored. a in unreported likelihood ratio tests we show that the s-score model provides incremental information beyond (updated) combined z-scores/o-scores models. The difference between the models (comparing nested models Model(6) vs. Model(8) and Model(7) vs. Model(9)) is significant at the 1%- level for all forecast horizon (1, 3 and 5-year ahead predictions).
77
Table 10: Homogeneity
Panel A1: Number of delisting firm quarters by decile (20 qtr ahead) S-Prob Z-Prob O-Prob Z-Probu O-Probu
Decile Distressed firm quarters Distressed firm quarters Distressed firm quarters Distressed firm quarters Distressed firm quarters 10 14,874 14,874 13,647 13,647 14,510 14,510 14,386 14,386 11,835 11,835 9 9,618 24,492 7,840 21,487 8,323 22,833 5,990 20,376 7,149 18,984 8 5,828 30,320 4,848 26,335 5,417 28,250 4,067 24,443 5,387 24,371 7 3,671 33,991 3,329 29,664 3,709 31,959 3,530 27,973 4,017 28,388 6 2,482 36,473 2,504 32,168 2,736 34,695 3,230 31,203 3,239 31,627 5 1,850 38,323 2,186 34,354 2,206 36,901 2,835 34,038 2,814 34,441 4 1,361 39,684 1,838 36,192 1,658 38,559 2,579 36,617 2,590 37,031 3 840 40,524 1,645 37,837 1,305 39,864 2,247 38,864 2,094 39,125 2 673 41,197 1,637 39,474 1,065 40,929 1,863 40,727 1,678 40,803 1 467 41,664 2,190 41,664 735 41,664 937 41,664 861 41,664
Panel A2: Number of distressed firm quarters by decile (8-20 qtr ahead) of total 23,309 distressed firm quarters S-Prob Z-Prob O-Prob Z-Probu O-Probu
Decile Distressed firm quarters Distressed firm quarters Distressed firm quarters Distressed firm quarters Distressed firm quarters 10 7,397 7,397 5,894 5,894 6,701 6,701 6,379 6,379 5,353 5,353 9 4,933 12,330 4,025 9,919 4,580 11,281 2,879 9,258 3,651 9,004 8 3,501 15,831 2,803 12,722 3,102 14,383 2,377 11,635 2,971 11,975 7 2,271 18,102 2,085 14,807 2,329 16,712 2,231 13,866 2,348 14,323 6 1,641 19,743 1,775 16,582 1,804 18,516 2,077 15,943 2,010 16,333 5 1,278 21,021 1,517 18,099 1,448 19,964 1,944 17,887 1,840 18,173 4 923 21,944 1,281 19,380 1,148 21,112 1,781 19,668 1,772 19,945 3 576 22,520 1,176 20,556 918 22,030 1,579 21,247 1,472 21,417 2 463 22,983 1,171 21,727 746 22,776 1,363 22,610 1,230 22,647 1 326 23,309 1,582 23,309 533 23,309 699 23,309 662 23,309
(continued on the next page)
78
(Table 10 continued)
Panel B: Homogeneity (distressed firm quarters identified by s-prob = distressed firm quarters identified by other models) S-Prob Z-Prob O-Prob Z-Probu O-Probu
Decile No. of
distressed firm quarters
Equal No. of distressed
firm quarters Percentage
Equal No. of distressed
firm quarters Percentage
Equal No. of distressed
firm quarters Percentage
Equal No. of distressed
firm quarters Percentage
10 14,874 9,068 61.0% 8,950 60.2% 9,876 66.4% 7,830 52.6% 9 24,492 16,806 68.6% 17,560 71.7% 16,913 69.1% 15,382 62.8% 8 30,320 21,908 72.3% 23,670 78.1% 21,449 70.7% 21,064 69.5% 7 33,991 25,784 75.9% 28,198 83.0% 25,211 74.2% 25,472 74.9% 6 36,473 29,091 79.8% 31,745 87.0% 28,799 79.0% 29,202 80.1% 5 38,323 32,129 83.8% 34,752 90.7% 32,127 83.8% 32,547 84.9% 4 39,684 34,787 87.7% 37,212 93.8% 35,254 88.8% 35,678 89.9% 3 40,524 36,996 91.3% 39,055 96.4% 37,974 93.7% 38,226 94.3% 2 41,197 39,085 94.9% 40,543 98.4% 40,309 97.8% 40,359 98.0% 1 41,664 41,664 100.0% 41,664 100.0% 41,664 100.0% 41,664 100.0%
This table reports the modelβs power and their homogeneity compared to the s-prob model. Panel A1 reports the frequency of classified financially distressed firm quarters by estimated distress/default probability deciles (10=high probability of financial distress β 1=low probability of financial distress). In the highest decile the s-prob model identify 14,874 out of 41,664 financially distressed firm quarters. The corresponding frequencies are 13,647/14,510/14,386/11,835 for the z-prob/o-prob/z-probu/o-probu βmodels. A firm quarter is considered as distressed firm quarter if the company is delisted in the next 20 quarters ahead. For the long-term performance we split the financial distress sample into a second part (2-5 years before a financial distressed related delisting occurs with 23,309 financially distressed firm quarter observations). The results show that the s-prob model can identify 35.7% of the indicated financially distressed firm quarter in the highest probability decile. Panel B shows the absolute and relative proportion of equally identified financially distressed firm quarter observations for the corresponding probability deciles (10=high probability of financial distress β 1=low probability of financial distress).
79
Table 11: Robustness Check: Descriptive Statistics (Market Sample)
Panel A: Univariate summary statistics (Solvent vs. financially distressed firm quarters from 1980Q1-2010Q4)
Variable Status Mean Median Std.dev. 1% 99% N S-Prob Solvent 0.18 0.03 0.29 0.00 1.00 243,523 Financially
Distressed 0.58 0.71 0.37 0.00 1.00 36,711
EDF Solvent 0.05 0.00 0.16 0.00 0.84 243,523 Financially
Distressed 0.23 0.05 0.30 0.00 0.99 36,711
C-Prob Solvent 0.00 0.00 0.05 0.00 0.09 243,523 Financially
Distressed 0.05 0.00 0.16 0.00 0.93 36,711
C-Probu Solvent 0.00 0.00 0.01 0.00 0.02 243,523 Financially
Distressed 0.01 0.00 0.03 0.00 0.14 36,711
Panel B: Correlations (Pearson above, Spearman rank below the diagonal)
Variable Financially Distressed S-Prob EDF C-Prob C-Probu
Financially Distressed 0.41 0.30 0.20 0.28 S-Prob 0.36 0.39 0.24 0.35 EDF 0.31 0.52 0.38 0.42 C-Prob 0.36 0.60 0.71 0.76
C-Probu 0.40 0.65 0.67 0.72 This table reports summary statistics for solvent vs. financially distressed observations by model. Panel A presents summary statistics of the evaluated (market) model outcomes. EDF (Expected Default Frequency) is the default probability estimated monthly following the sequential-iterations algorithm specified by Crosbie and Bohn (2003) and Bharath and Shumway (2008). C-PROB is defined as the monthly-estimated default probability using the original coefficient from the model No. 2 by Campbell et al. (2008), where βπ π πππ π ππππ = β9.08 β 29.67 β (πππππππππΆπΆπΆπΆππππ) + 3.36 β (πππππππππΆπΆ) β 7.35 β(πΈπΈπππππΈπΈπππΆπΆππππ) + 1.48 β (πππππππππΆπΆ) + 0.082 β (πππππππππΈπΈ) β 2.40 β (πΆπΆπΆπΆπππΆπΆπππππΆπΆ) + 0.054 β (πππ΅π΅) β 0.937 β(πΆπΆπππππΆπΆπΈπΈ) ). C-PROBU
is the corresponding model No. 2 (re-estimated by a growing window logit regression model, starting with month 1976|01, see the Appendix 1 for a detailed description). We winsorize all modelβs parameter at the 1st and 99th percentile. The c-scores are transformed into probabilities using the standard logistic function (c-prob=1/(1+exp(-c-score)). The total number N of observations are 280,234 firm quarters / 10,143 firms over the sample period 1980Q1 to 2010Q4 with nonmissing values for the (market) default probabilities. FINANCIALLY DISTRESSED represents an indicator variable, which is one if delisting (indicated by CRSP delisting codes 400, 500-585) occurs within the next 20 quarters (i.e., a delisting firm quarter is determined if the firm experience a delisting within the next 20 quarters as defined in table 1). The differences in probabilities of the solvent and financially distressed means are significant at the 1%-level using a t-test or the Wilcoxon rank sum test (two-tailed). Panel B shows the Pearson correlation above and the Spearman rank correlation below the diagonal for the evaluated models.
80
Table 12: Robustness Check: Accounting vs. Market Models
Panel A: Receiver Operating Characteristic (ROC): complete sample
Model: # solvent # delisted AUROC SE S-Prob 243,523 36,711 0.8121 0.0012
EDF 243,523 36,711 0.7629 0.0013 C-Prob 243,523 36,711 0.8109 0.0012
C-Probu 243,523 36,711 0.8398 0.0010
Panel B: Receiver Operating Characteristic (ROC): bootstrapped samples
Model: MEAN MEDIAN MIN MAX
S-Prob 0.8121 0.8121 0.8084 0.8158
EDF 0.7630 0.7630 0.7590 0.7676
C-Prob 0.8109 0.8109 0.8072 0.8143
C-Probu 0.8398 0.8398 0.8365 0.8433
Panel A reports the area under Receiver Operating Characteristic curve (AUROC) for the s-prob model compared to (mixed)-market models. The EDF (Expected Default Frequency) equals the default probability estimated monthly following the sequential-iterations algorithm of Crosbie and Bohn (2003) and Bharath and Shumway (2008). We winsorize all modelβs parameter at the 1st and 99th percentile. C-PROB is defined as the monthly-updated default probability using the original coefficient from the mixed-model No. 2 by Campbell et al. (2008). C-PROBu
is the corresponding model No. 2 (re-estimated by a growing window logit regression model, starting with month 1976|01). We winsorize all measures at the 1st and 99th percentile before computing the logits. The area under the ROC Curve (AUROC) and its standard deviation (SE) is calculated following the nonparametric approach by DeLong et al. (1988), Hanley and McNeil (1982). By definition a firm is considered as "delisted" if delisting occurs within the next 20 quarters ahead. In Panel B we confirm the statistical inference and sample independence of the ROC curve by employing 1,000 bootstrap replications of the original sample to obtain bootstrap standard errors (SE).
81
Table 13: Robustness Check: Information content tests (Market Sample)
Panel A: (1-year ahead prediction, firm quarters from 1980Q1-2010Q4, 277,373 solvent firm quarters vs. 2,861 financially distressed firm quarters)
Variables: Model(1) Model(2) Model(3) Model(4) Model(5) Model(6) Model(7) Model(8) Constant -4.590 -3.576 -1.825 0.113 -2.453 -0.787 -2.032 -0.186 (37.55)*** (24.26)*** (7.68)*** (0.29) (8.35)*** (1.57) (7.88)*** (0.43) coeff S-Score 0.147 . . . 0.085 0.058 . . (9.60)*** . . . (4.27)*** (2.59)*** . . coeff EDF . 0.167 . . 0.069 0.044 0.079 0.042 . (7.53)*** . . (2.73)*** (1.77)* (3.12)*** (1.67)* coeff C-Score . . 0.359 . 0.212 . 0.265 . . . (11.35)*** . (4.56)*** . (6.09)*** . coeff C-Scoreu . . . 0.710 . 0.526 . 0.624 . . . (10.91)*** . (5.76)*** . (7.52)*** Pseudo RΒ² 0.0814 0.0962 0.1229 0.1620 0.1542 0.1728 0.1358 0.1656 LRa 1,777*** 2,100*** 2,683*** 3,537*** 3,367*** 3,772*** 2,965*** 3,615***
(continued on the next page)
82
(Table 13 continued)
Panel B: (3-year ahead prediction, firm quarters from 1980Q1-2010Q4, 277,373 solvent firm quarters vs. 2,861 financially distressed firm quarters)
Variables: Model(1) Model(2) Model(3) Model(4) Model(5) Model(6) Model(7) Model(8) Constant -4.633 -3.903 -2.365 -0.875 -3.120 -1.541 -2.669 -1.056 (31.39)*** (18.87)*** (6.13)*** (1.67)* (6.38)*** (2.33)** (6.11)*** (1.81)* coeff S-Score 0.126 . . . 0.079 0.049 . . (5.88)*** . . . (2.91)*** (1.67)* . . coeff EDF . 0.112 . . 0.054 0.021 0.063 0.021 . (4.96)*** . . (1.92)* (0.76) (2.24)** (0.74) coeff C-Score . . 0.284 . 0.126 . 0.187 . . . (6.31)*** . (1.82)* . (2.90)*** . coeff C-Scoreu . . . 0.542 . 0.413 . 0.492 . . . (7.00)*** . (3.66)*** . (4.81)*** Pseudo RΒ² 0.0478 0.0502 0.0543 0.0909 0.0770 0.0962 0.0638 0.0919 LRa 691*** 726*** 786*** 1,315*** 1,114*** 1,393*** 923*** 1,330***
(continued on the next page)
83
(Table 13 continued)
Panel C: (5-year ahead prediction, firm quarters from 1980Q1-2010Q4, 277,373 solvent firm quarters vs. 2,861 financially distressed firm quarters)
Variables: Model(1) Model(2) Model(3) Model(4) Model(5) Model(6) Model(7) Model(8) Constant -4.750 -4.204 -2.880 -1.453 -3.624 -1.843 -3.173 -1.483 (26.26)*** (15.73)*** (5.45)*** (2.22)** (5.44)*** (2.28)** (5.29)*** (2.07)** coeff S-Score 0.111 . . . 0.074 0.037 . . (3.99)*** . . . (2.12)** (0.97) . . coeff EDF . 0.086 . . 0.043 0.003 0.051 0.003 . (3.34)*** . . (1.30) (0.10) (1.53) (0.11) coeff C-Score . . 0.236 . 0.088 . 0.150 . . . (3.99)*** . (0.96) . (1.78)* . coeff C-Scoreu . . . 0.468 . 0.400 . 0.460 . . . (5.08)*** . (2.96)*** . (3.79)*** Pseudo RΒ² 0.0333 0.0307 0.0319 0.0674 0.0486 0.0697 0.0381 0.0674 LRa 333*** 307*** 320*** 675*** 487*** 699*** 382*** 675***
(continued on the next page)
84
(Table 13 continued)
Panel D: Vuong-test statistics
S-Score EDF C-Score C-Scoreu Model(1) vs. Model(2) Model(3) Model(4)
1-year ahead -2.16** -9.62** -18.19** 3-year ahead 0.21 -1.17 -10.78*** 5-year ahead 1.81* 1.26 -8.60*** */**/*** asterisks refer to significance at a 10%/ 5%/ 1% level for a one sided t-test. This table shows the coefficients, z-statistics (in parentheses, which accounts for firm dependence between firm quarter observations), McFaddenβs (1974)-Pseudo-R2 and the likelihood ratio statistic (LR=2(L1-L0), where L1 is the maximized log likelihood for the model and L0 is the maximized log likelihood for the baseline model with a constant only) using a dynamic logit models (Chava and Jarrow 2004; Shumway 2001). The sample is limited to nonmissing observations of the default probabilities using market models (in total there are 280,234 / 10,143 firm quarter / firm observations). Due to the panel structure of the data there are less independent observations than assumed by a standard logit regression model. The panels compare the contribution of the s-score estimation with the results from univariate and combined regressions results for 1, 3 and 5-year ahead predictions. The scores are calculated in the same way outlined in table 9. The probabilities are converted into scores according to: π π πππ π ππππππ = ππππ[πππππ π ππππ/(1 β πππππ π ππππ)]. We bound all scores between Β±18.420680999. Panel D reports the Vuong LR test statistic results for strictly non-nested models (model(1) vs. model(2-4)). A positive LR test statistic indicates that the s-score model (i.e. model(1)) is preferable. If the LR test statistic is negative the compared model is favored. a in unreported likelihood ratio tests we show that the s-score model provides incremental information beyond (updated) combined edf/c-scores models. The difference between the models (comparing nested models Model(5) vs. Model(7) and Model(6) vs. Model(8)) is significant at the 1%- level for all forecast horizon (1, 3 and 5-year ahead predictions).