assessing and combining financial conditions indexes · assessing and combining financial...
TRANSCRIPT
Assessing and Combining Financial ConditionsIndexes∗
Sirio Aramonte,a Samuel Rosen,b and John W. Schindlera
aFederal Reserve BoardbUniversity of North Carolina at Chapel Hill, Kenan-Flagler
Business School
We evaluate the short-horizon predictive ability of finan-cial conditions indexes for stock returns and macroeconomicvariables. We find reliable predictability only when the sampleincludes the 2008 financial crisis, and we argue that this resultis driven by tailoring the indexes to the crisis and by non-synchronous trading. In addition, we suggest a simple proce-dure for aggregating the various indexes into a single proxy forfinancial conditions, which can help to reduce the uncertaintyfaced by policymakers when monitoring financial conditions.
JEL Codes: E32, G01, G17.
1. Introduction
The severity and the economic impact of the 2008 financial crisishave led to a proliferation of indexes that proxy for financial con-ditions or financial stress, which we collectively refer to as finan-cial conditions indexes (FCIs). An open question in the literature is
∗We would like to thank Kathryn Smith for research assistance, and partici-pants of several Federal Reserve Quantitative Surveillance Group meetings. Wewould also like to thank Malcolm Spittler and Troy Matheson for providing dataand guidance for, respectively, the Citi Financial Conditions Index and the Inter-national Monetary Fund U.S. Financial Conditions Index, and Mark Carlson,Kurt Lewis, and William Nelson, and Roberto Cardarelli, Selim Elekdag, andSubir Lall for providing data and guidance for, respectively, their Financial Mar-ket Stress Index and the International Monetary Fund U.S Financial Stress Index.This article represents the views of the authors and should not be interpreted asreflecting the views of the Board of Governors of the Federal Reserve System orother members of its staff. Corresponding author: [email protected]; Tel.:+1-202-912-4301.
1
2 International Journal of Central Banking February 2017
whether these indexes should be used as indicators of current finan-cial conditions, or whether they also have predictive power for futurefinancial and economic activity and could be used as early-warningindicators. In this paper we evaluate whether FCIs can predict stockreturns or innovations to macroeconomic variables over a one-monthor a one-quarter horizon. In addition, we study the Granger causalityrelation between the FCIs and several measures of credit availabilitythat we use as proxies for the “financial cycle” (Borio 2014). Finally,we suggest a simple procedure for aggregating the various FCIs intoa single proxy for financial conditions.
We find that most FCIs can predict monthly and quarterlyreturns on the S&P 500 and on a portfolio of financial companies andalso innovations to a number of macroeconomic variables. However,this predictability is not robust to excluding the period around the2008 financial crisis (2007–12). We posit three possible explanationsfor this pattern. The first is threshold effects, in the sense that finan-cial conditions may matter only after deteriorating beyond a certainthreshold. The second is data mining, because most of the FCIs wereintroduced after the financial crisis, and they are built using variablesthat displayed significant movements during the crisis. And the thirdis non-synchronous data, given that most FCIs include one variablederived from the prices of options that trade after equity marketsclose, hence FCIs include, by construction, information that willbe incorporated in stock prices the next day. We provide evidencethat all three effects are present, but we argue that predictability ismainly driven by data mining and by non-synchronous trading.
The FCIs we study measure financial conditions in one of twoways. The first approach is to evaluate whether broad financial con-ditions are loose or tight by historical standards (e.g., the Bloombergand Chicago Fed indexes). Alternatively, other indexes purport tomeasure whether the financial system is experiencing historicallyunusual stress (e.g., the Carlson, Lewis, and Nelson 2014 index).Despite these and other methodological differences that we detail insection 2, the FCIs exhibit a large amount of common variation, ascan be seen in the top graph in figure 1. Such a result is expectedbecause, while the concept of “financial conditions” is often onlyloosely defined by the articles that build the various FCIs, changesin the state of the financial system directly affect many of the vari-ables used to build the FCIs. However, the top graph in figure 1
Vol. 13 No. 1 Assessing and Combining Financial Conditions 3
Figure 1. Time-Series Plots of the Twelve FinancialConditions Indexes
Notes: The graphs show the time series of the twelve FCIs we study. For scalereasons, the International Monetary Fund (IMF) U.S. Financial Stress Index andthe Carlson, Lewis, and Nelson (2014) Financial Stress Index are not shown. Thebottom graph focuses on the years surrounding the 2008 financial crisis, and itshows the averages of standardized versions of (i) Citi FCI and IMF FCI, (ii)MS FCI and ANFCI, and (iii) the remaining FCIs. See table 1 for a list of FCIacronyms.
4 International Journal of Central Banking February 2017
shows that, at certain points in time, different indexes provide aconflicting reading of financial conditions. This was the case, forinstance, in 1997–98 for the Citi FCI and 2004–05 for one of theBloomberg FCIs. In section 3, we propose a method to extract asingle and more precise measure of financial conditions from the setof indexes we consider.
Having a reliable measure of financial conditions is importantfor policymakers because, as the literature has shown, financialconditions matter in a variety of ways. The importance of a well-functioning financial system to the broad economy is highlightedby the results in Bernanke and Blinder (1992), Kashyap, Stein,and Wilcox (1993), Kashyap, Lamont, and Stein (1994), Peek andRosengren (1997), and Paravisini (2008), who show that tight mon-etary policy, binding capital ratios, and bank financing constraintscan reduce the supply of credit, especially in the case of small bankswith less liquid assets (Kashyap and Stein 2000). A reduced creditsupply has particularly negative effects for small firms (Gertler andGilchrist 1994, Khwaja and Mian 2008) and bank-dependent bor-rowers (Chava and Purnanandam 2011). A tight credit supply ulti-mately affects investment (Campello, Graham, and Harvey 2010),inventories (Kashyap, Lamont, and Stein 1994), and the broadereconomy (Bernanke 1983; Peek and Rosengren 1997, 2000; Calomirisand Mason 2003).
We assess predictability using simple predictive regressions of aset of stock returns or innovations to macroeconomic variables onlagged FCI values. Our analysis differs from Hatzius et al. (2010),who also evaluate the predictive power of several FCIs, in a fewimportant ways. First, we only consider FCIs that are updated atleast monthly (we should emphasize that, as shown in figure 1, theFCIs still have meaningful monthly variation), and we study theirpredictive power for stock returns and innovations to macroeconomicvariables one month and one quarter ahead. By using such horizons,our predictability results are unlikely to reflect business-cycle effects,especially in the case of one-month-ahead regressions. Second, westudy a larger number of FCIs and a larger number of financial andmacroeconomic variables. Third, we explicitly discuss whether thepredictability that arises in coincidence with the 2008 financial crisisis hard-wired in the FCIs, in the sense that the variables includedin the FCIs may have been chosen on the basis of whether they
Vol. 13 No. 1 Assessing and Combining Financial Conditions 5
experienced large fluctuations in 2007 and 2008.1 Finally, asdescribed in more detail in section 2, we assess the statistical signif-icance of predictability with a methodology that is robust to biasesgenerated by the high persistence that typically characterizes FCIs,which we found had noticeable effects on the results discussed insection 2.
Our conclusions differ from those of English, Tsatsaronis, andZoli (2005), who focus on four- and eight-quarter horizons, and con-sider a longer sample than we do. They aggregate financial variableswith principal component analysis and find that the resulting proxiesfor financial conditions have some predictive power for macroeco-nomic variables. Our conclusions, however, are more in line withthose of Hatzius et al. (2010), because we find limited value in usingFCIs as reliable early-warning indicators, especially when (i) exclud-ing the period surrounding the 2008 financial crisis from the sample,and (ii) acknowledging that the choice of the variables that entermany FCIs may have been influenced by knowing, in hindsight, whatprecipitated the financial crisis.
In the second part of our paper, we discuss a simple two-stepmethodology for combining the various FCIs into a single proxy forfinancial conditions. The large number of FCIs is itself indicative ofthe uncertainty that surrounds the measurement of financial con-ditions. In the top graph in figure 1, one can see that the FCIsgenerally move together in the long run, but they can provide con-flicting assessments at a given point in time—even shortly beforea major financial crisis. For instance, the bottom graph in figure1 shows the averages of three sets of (standardized) FCIs between2006 and 2009. The shaded area highlights the period between Marchand August 2008, which corresponds to the months between the pur-chase of The Bear Stearns Companies, Inc. by JPMorgan Chase &Co. (March 16, 2008) and the bankruptcy filing of Lehman BrothersHoldings Inc. on September 15, 2008. The chart clearly shows that
1For example, Hatzius et al. (2010) develop their own FCI in addition to eval-uating others. They note on page 21 that “the better performance during themost recent five years . . . may reflect selection bias in our choice of variables toinclude in the index: naturally, our selection was governed in part by an under-standing of the types of financial variables that were used for monitoring andmeasuring the recent financial crisis. In this sense, we did not seek to mitigateobserver bias.”
6 International Journal of Central Banking February 2017
different FCIs give conflicting readings of the state of the financialsystem in the second and third quarters of 2008, ranging from anoticeable deterioration to a large improvement.
Aggregating the individual indexes can help average out modeluncertainty and provide policymakers with a single variable thatreflects the information available in all of the FCIs. We first identifythe indexes that best summarize the information provided by theremaining FCIs. In the second step we form all combinations of theidentified indexes and select the “best” combination on the basis ofhow well it summarizes the information in the remaining FCIs. Thisprocedure is discussed in detail in section 3.
2. Assessing the Predictive Ability of FCIs
We consider twelve FCIs that (i) focus on the United States, (ii)are updated at least every month, and (iii) have a sufficiently longhistory.2 Table 1 contains a detailed list of data sources and includesthe abbreviated FCI names we use in the paper.
The construction of the FCIs varies considerably, although allof them are largely based on financial market variables, includingimplied volatilities, Treasury yields, yield spreads, commercial paperrates, stock returns, and exchange rates (see Kliesen, Owyang, andVermann 2012 for a detailed list of variables that underlie a range ofthe FCIs we study here). Some FCIs only include a small set of vari-ables, as in the case of the Bloomberg Financial Conditions Index,which is based on ten underlying variables. One index, the ChicagoFed Adjusted National Financial Conditions Index, has more than100 underlying variables.
In general, the constituent variables are aggregated using prin-cipal component analysis or simple weighted sums. Principal com-ponent analysis is used extensively in the literature to extract infor-mation from a large set of macroeconomic or financial variables forforecasting purposes. For example, Ludvigson and Ng (2007) rely
2In order to increase the power of our test, we require that the FCIs coverthe stressful episodes that characterized the late 1990s (the Long-Term Capi-tal Management collapse, and the Asian and Russian financial crises). For thisreason, we do not include the HSBC Financial Clog Index (Bloomberg tickerHSCLOG Index) or the Westpac U.S. Financial Stress Index (Bloomberg tickerWRAISTRS Index), whose series start in 2007 and 1998, respectively.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 7
Tab
le1.
Sum
mar
yD
escr
iption
sof
the
Tw
elve
FC
Is
Ava
ilab
ility
Shor
tan
dSta
rtY
ear
Full
Nam
eN
ame
Agg
rega
tion
Yea
rEst
.R
efer
ence
Dat
aSou
rce
Blo
ombe
rgU
.S.Fin
anci
alC
ondi
tion
sIn
dex
BFC
ID
-WA
1994
2009
Ros
enbe
rg(2
009)
Blo
ombe
rg
Blo
ombe
rgU
.S.Fin
anci
alC
ondi
tion
sIn
dex
Plu
sB
FC
I+D
-WA
1994
2009
Ros
enbe
rg(2
009)
Blo
ombe
rg
Cle
vela
ndFin
anci
alSt
ress
Inde
xC
FSI
D-W
A19
9120
09O
etet
al.(2
011)
Cle
vela
ndFe
d
Mor
gan
Stan
ley
Fin
anci
alC
ondi
tion
sIn
dex
U.S
.M
SFC
ID
-WA
1995
N/A
Blo
ombe
rg
Fin
anci
alM
arke
tSt
ress
Inde
xC
LN
FSI
W-F
PC
1994
2003
Car
lson
,Lew
is,an
dN
elso
n(2
014)
Aut
hors
Nat
iona
lFin
anci
alC
ondi
tion
sIn
dex
NFC
IW
-FP
C19
7320
06B
rave
and
But
ters
(201
0)C
hica
goFe
d
Adj
uste
dN
atio
nalFin
anci
alC
ondi
tion
sIn
dex
AN
FC
IW
-FP
C19
7320
06B
rave
and
But
ters
(201
0)C
hica
goFe
d
St.Lou
isFe
dFin
anci
alSt
ress
Inde
xST
LFSI
W-F
PC
1993
2009
Klie
sen
and
Smit
h(2
010)
St.Lou
isFe
d
Kan
sas
City
Fin
anci
alSt
ress
Inde
xK
CFSI
M-F
PC
1990
2009
Hak
kio
and
Kee
ton
(200
9)K
ansa
sC
ity
Fed
Cit
iFin
anci
alC
ondi
tion
sIn
dex
Cit
iFC
IM
-WA
1983
2000
D’A
nton
io(2
008)
Cit
iR
esea
rch
IMF
U.S
.Fin
anci
alC
ondi
tion
sIn
dex
IMF
FC
IM
-DFM
1994
2009
Mat
heso
n(2
012)
Aut
hor
IMF
U.S
.Fin
anci
alSt
ress
Inde
xIM
FFSI
M-W
A19
8020
11C
arda
relli
,E
lekd
ag,
and
Lal
l(2
011)
Aut
hors
Note
s:T
heta
ble
prov
ides
sum
mar
yin
form
atio
non
the
twel
veFC
ISw
est
udy.
The
abbr
evia
tion
sin
the
“Ava
ilabi
lity
and
Agg
rega
-ti
on”
colu
mn
indi
cate
whe
ther
the
inde
xis
upda
ted
daily
(D),
wee
kly
(W),
orm
onth
ly(M
),an
dw
heth
erth
eva
riab
les
unde
rlyi
ngth
ein
dex
are
aggr
egat
edw
ith
aw
eigh
ted
aver
age
(WA
),fir
stpr
inci
palco
mpon
ent
(FP
C),
ordy
nam
icfa
ctor
mod
el(D
FM
).T
heFC
Iin
Car
lson
,Lew
is,an
dN
elso
n(2
014)
isba
sed
onan
inde
xde
velo
ped
in20
03.A
llin
dexe
sar
eex
pres
sed
asz-
scor
es,w
ith
the
exce
ptio
nof
the
CLN
FSI
,w
hich
isa
prob
abili
ty.In
the
empi
rica
lan
alys
is,w
ere
plac
eth
eC
LN
FSI
wit
hit
sna
tura
llo
gari
thm
(the
inde
xne
ver
isex
actl
yze
roin
our
sam
ple)
.
8 International Journal of Central Banking February 2017
on dynamic factor analysis to summarize a broad cross-section ofvariables, and find that several of the resulting factors can predictone-quarter-ahead excess stock returns. Stock and Watson (2002)use principal component analysis to build factors used to predictseveral macroeconomic variables in the medium run (six, twelve,and twenty-four months ahead).
For indexes constructed as weighted sums, weights are typicallyassigned subjectively by the authors, although a few of the indexesuse more sophisticated methods. The Cleveland Financial StressIndex (CFSI), for instance, calculates weights dynamically basedon the relative dollar flow observed in the Federal Reserve Board’sFlow of Funds statistical release (Z.1) (Oet et al. 2011). All theindexes are expressed in terms of z-scores, with the exception of theFinancial Stress Index of Carlson, Lewis, and Nelson (2014), whichis expressed in terms of probabilities.
2.1 Predictive Regressions
We evaluate the predictive power of the twelve FCIs with a series ofmonthly and quarterly predictive regressions of the form
yt = α + β × FCIt−1 + εt.
The FCIs enter the regression in levels, rather than changes,because the developers of most of the FCIs emphasize that theirindexes are ordinal measures of financial conditions/stress.
We focus on one-month-ahead and one-quarter-ahead regressionsbecause our predictability results might in part reflect business-cycleeffects if the dependent variable were measured over longer hori-zons. In particular, many FCIs include the implied volatility indexVIX as a constituent variable. Bollerslev, Tauchen, and Zhou (2009)find that the variance risk premium, which is the difference betweenthe squared value of VIX and a measure of realized variance, canpredict stock returns about three to six months out, with the asso-ciated adjusted R2 peaking at a five-month horizon and declininggradually at longer horizons. The variance risk premium is drivenby the dynamics of both volatility risk and risk aversion; the latteris embedded in VIX, and its evolution over time has a business-cyclecomponent. Note that we do not control for the predictive power ofother variables, like the variance risk premium, because the results
Vol. 13 No. 1 Assessing and Combining Financial Conditions 9
from regressions in which FCIs are the only predictor already suggesta lack of reliable predictive power at the horizons we focus on.
The FCI coefficient is estimated with OLS, and we assess itsstatistical significance with either heteroskedasticity-consistent stan-dard errors or the local-to-unity asymptotics procedure of Campbelland Yogo (2006). Local-to-unity asymptotics are useful in evaluatingthe statistical significance of persistent predictors, because, in suchcases, the standard t-test can give a rejection rate that is inconsistentwith its nominal size.
As is clear from the time-series plots in figure 1, FCIs tend tobe quite persistent. The high autocorrelation that characterizes theFCIs is also evident in the confidence intervals for the autoregressiveroots shown in table 2.3 We report confidence intervals, rather thanpoint estimates, to highlight that, after accounting for statisticaluncertainty, the FCIs plausibly have autoregressive roots very closeto one. In five cases, we are actually unable to reject the hypothesisthat the FCIs have a unit root. Given the nature of the problem westudy, asymptotic standard errors would normally be biased againstfinding predictability, and the local-to-unit asymptotics procedure ofCampbell and Yogo (2006) corrects for this bias. Our results showthat the Campbell and Yogo (2006) procedure finds predictability25 percent more often than OLS for innovations to macroeconomicvariables, and six times as often for stock returns.4
3The procedure for calculating the confidence intervals in table 2 is describedin the online appendix to Campbell and Yogo (2006).
4The direction and severity of the bias depend on the persistence of the pre-dictor, on the correlation between the innovations to the predictor and to thepredicted variable (δ), and on the sign of the predictive relation (see the discus-sion on the skewness of the t-statistic distribution in section 3.2 of Campbell andYogo 2006). Given the persistence of the FCIs, and assuming that poor finan-cial conditions lead to negative returns/innovations to macroeconomic variables,OLS should find predictability less often than the Campbell and Yogo (2006)procedure for FCIs that increase in value when financial conditions deteriorate.In order to run the Campbell and Yogo (2006) procedure, we need to make surethat the correlation δ is negative, and we do so by multiplying an FCI by –1 ifneeded. The figures mentioned in the text refer to predictive regressions basedon the full sample and where the FCIs, after being multiplied by –1 if needed,increase in value when financial conditions deteriorate. This is the case for all thefull-sample regressions that predict returns, and for 78 percent of the full-sampleregressions that predict innovations to macroeconomic variables.
10 International Journal of Central Banking February 2017
Tab
le2.
Sum
mar
ySta
tist
ics
ofth
eT
wel
veFC
Is
Val
ue
on90
%C
Ifo
rA
RM
ean
Med
ian
10th
90th
Sep
t.20
08#
Obs.
Root
Dai
ly
BFC
I0.
426
0.06
0−
0.88
71.
910
7.65
144
350.
922
0.99
6B
FC
I+0.
264
−0.
057
−1.
237
1.78
25.
980
4435
0.90
10.
986
CFSI
0.13
80.
024
−1.
065
1.61
32.
592
4435
0.93
91.
004
MS
FC
I0.
075
−0.
139
−1.
144
1.71
80.
764
4429
0.83
40.
947
Wee
kly
CLN
FSI
−4.
311
−4.
625
−7.
264
−0.
660
−0.
001
887
0.92
30.
997
NFC
I−
0.36
1−
0.53
4−
0.78
30.
229
1.84
788
70.
936
1.00
2A
NFC
I−
0.03
0−
0.22
4−
0.81
50.
923
3.31
088
70.
757
0.89
8ST
LFSI
0.05
0−
0.12
7−
0.90
50.
909
2.90
488
70.
936
1.00
2
Mon
thly
KC
FSI
0.13
2−
0.11
5−
0.80
01.
010
2.73
020
40.
945
1.00
6C
itiFC
I0.
137
0.01
3−
1.17
21.
694
3.22
420
40.
916
0.99
3IM
FFC
I0.
078
−0.
189
−0.
781
1.10
02.
930
204
0.97
41.
015
IMF
FSI
−0.
216
−1.
065
−3.
425
3.39
38.
930
204
0.80
70.
930
Note
s:T
heta
ble
show
sth
em
ean,
med
ian,
10th
and
90th
per
cent
iles,
pea
kcr
isis
valu
e(S
epte
mber
2008
),to
talob
serv
atio
ns,an
d90
per
cent
confi
denc
ein
terv
alfo
rth
eA
R(1
)au
tore
gres
sive
root
(see
Cam
pbel
lan
dY
ogo
2006
).T
ime
per
iod:
July
1995
toJu
ne20
12.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 11
In practice, we assume that each of the FCIs follows an AR(1)process, and use local-to-unity asymptotics unless the autoregressiveroot of the FCI is sufficiently distant from one, in a sense definedin footnote 5, or unless there is no correlation between the innova-tions to the FCI’s autoregressive process and the innovations in aregression of the predicted variable on the FCI.5 Note that both apersistent predictor and a non-zero correlation are necessary for thestandard OLS asymptotics to be inappropriate.
The dependent variables in our predictive regressions are (i)returns on a broad market index and on various industry portfo-lios, and (ii) autoregressive residuals of log-changes in several eco-nomic variables; we study residuals because, unlike returns, changesin macroeconomic variables can be autocorrelated. We considermonthly and quarterly returns on the S&P 500 and on sevenequally weighted industry portfolios: finance, construction, manu-facturing, transportation, wholesale trade, retail trade, and services.The macroeconomic variables we consider measure the availability ofcredit (total consumer credit, and commercial and industrial loans),the state of the housing market (housing starts), and manufactur-ing activity (durable goods orders, industrial production, and totalmanufacturing inventory).6
We present the results for the predictability of stock returns intables 4 through 7. For each portfolio/FCI combination we report thecoefficient on the FCI (βFCI) and the regression root mean squarederror (RMSE). We show an asterisk next to a coefficient when thecoefficient is statistically significant at the 90 percent confidencelevel.
5Using the notation in Campbell and Yogo (2006), we rely onheteroskedasticity-consistent standard errors if the DF-GLS statistic is less than–5 (a more negative DF-GLS statistic shifts the confidence interval for the autore-gressive root of the predictor away from one), or if the parameter δ̂ (which meas-ures the correlation between the innovations) is equal to zero. When needed,we linearly interpolate the values obtained from the lookup tables in the onlineappendix to Campbell and Yogo (2006), which only provide a discrete set ofvalues for δ̂ and of the DF-GLS statistic.
6Stock returns are from the Center for Research in Security Prices (CRSP)through Wharton Research Data Services. The macroeconomic data are from theFederal Reserve Economic Data (FRED) database of the Federal Reserve Bankof St. Louis. See table 3 for summary statistics and additional details on theconstruction of the industry portfolios.
12 International Journal of Central Banking February 2017Tab
le3.
Sum
mar
ySta
tist
ics
ofIn
dust
rySto
ckR
eturn
san
dLog
-Chan
ges
ofM
acro
Var
iable
s
Mea
nStd
.D
ev.
Med
ian
10th
90th
#O
bs.
Stoc
kRet
urns
S&P
500
0.00
60.
046
0.01
0−
0.06
00.
059
204
Fin
ance
0.00
90.
046
0.01
2−
0.03
30.
059
204
Con
stru
ctio
n0.
012
0.08
00.
016
−0.
085
0.10
020
4M
anuf
actu
ring
0.01
20.
073
0.01
5−
0.08
10.
099
204
Tra
nspo
rtat
ion
0.00
90.
061
0.01
7−
0.07
10.
072
204
Who
lesa
leTra
de0.
012
0.06
70.
020
−0.
070
0.08
520
4R
etai
lTra
de0.
011
0.07
20.
009
−0.
066
0.07
820
4Se
rvic
es0.
012
0.08
40.
020
−0.
086
0.10
120
4
Mac
roVar
iabl
es
C&
I0.
004
0.01
00.
005
−0.
011
0.01
520
4C
onsu
mer
Cre
dit
0.00
50.
006
0.00
4−
0.00
30.
009
204
Dur
able
Goo
ds0.
002
0.04
20.
003
−0.
047
0.04
920
4H
ousi
ngSt
arts
−0.
003
0.06
9−
0.00
7−
0.08
90.
080
204
Ind.
Pro
duct
ion
0.00
20.
007
0.00
2−
0.00
60.
009
204
Tot
alIn
vent
ory
0.00
20.
007
0.00
2−
0.00
80.
010
204
Note
s:T
heta
ble
repor
tssu
mm
ary
stat
isti
csfo
rth
ere
turn
son
the
S&P
500
inde
xan
don
indu
stry
por
tfol
ios,
and
for
log-
chan
ges
inse
lect
edm
acro
econ
omic
vair
able
s.T
hein
dust
rypor
tfol
ios
are
equa
llyw
eigh
ted
and
are
built
usin
gC
RSP
data
onth
eba
sis
ofSt
anda
rdIn
dust
rial
Cla
ssifi
cati
on(S
IC)
code
s(fi
nanc
e:60
00to
6231
and
6712
to67
26;co
nstr
ucti
on:15
21to
1799
;m
anuf
actu
ring
:20
11to
3999
;tr
ansp
orta
tion
:40
11to
4971
;w
hole
sale
trad
e:50
12to
5199
;re
tail
trad
e:52
11to
5999
;se
rvic
es:
7011
to89
99).
The
mac
roec
onom
icva
riab
les
are
(wit
hth
eFe
dera
lR
eser
veB
ank
ofSt
.Lou
is’s
FR
ED
mne
mon
icin
pare
nthe
ses)
:co
mm
erci
alan
din
dus-
tria
llo
ans
atal
lco
mm
erci
alba
nks
(BU
SLO
AN
S),
tota
lco
nsum
ercr
edit
owne
dan
dse
curi
tize
d(T
OTA
LSL
),m
anuf
actu
rers
’ne
wor
ders
for
dura
ble
good
s(D
GO
RD
ER
),ho
usin
gst
arts
(HO
UST
),in
dust
rial
prod
ucti
on(I
ND
PR
O),
and
valu
eof
man
ufac
ture
rs’t
otal
inve
ntor
ies
(AM
TM
TI)
.T
ime
per
iod:
July
1995
toJu
ne20
12.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 13
Table 4 shows results for the 1995–2012 sample, and it indicatesthat eleven of the twelve FCIs can predict returns on the financeindustry portfolio, that four can predict the S&P 500, and that theMorgan Stanley index can predict returns on the finance and threeadditional industry portfolios. There is noticeable dispersion in theRMSEs across industry portfolios for a given FCI, with the construc-tion and services portfolios generally showing the largest RMSE andthe finance portfolios the lowest, but the RMSEs are remarkablysimilar across FCIs for a given portfolio. The statistically significantcoefficients have the expected negative sign, indicating that higherlevels of financial stress (or tighter financial conditions) tend to befollowed by lower returns.
The severity of the 2008 financial crisis naturally raises the ques-tion of whether the predictive power of the FCIs mainly arises fromthe events that started in early 2007, or whether it is also present inthe broader sample. In table 5 we report the coefficients and RMSEsestimated over the 1995–2006 sample, and the results highlight that,essentially, there is no meaningful predictability left. Later in thissection we discuss whether the lack of predictive power outside ofthe 2008 crisis is due to predictability only being there during peri-ods of financial stress, or whether it is the result of the FCIs beingtailored to the recent financial crisis.
The conclusions that we can draw from monthly returns alsocarry over to quarterly returns (tables 6 and 7). Five FCIs have pre-dictive power for returns on the S&P 500 or on the financial industryportfolio in the 1995–2012 sample. Similar to the monthly analysis,returns on the remaining industry portfolios are not predictable.
Restricting the sample to the pre-crisis period (1995–2006, table7) all but eliminates the predictability, with the exception of theMorgan Stanley index, which can now predict returns on five port-folios. Note that the coefficients for the Morgan Stanley index arepositive, while they were negative at the monthly horizon in thefull sample (table 4). This difference is consistent with the possibil-ity that measuring returns over longer horizons captures the earlysigns of a turning business cycle. The Morgan Stanley index maysignal poor financial conditions when the business cycle is close toits trough, and, while one-month-ahead returns may still be negativeas the economy bottoms out, one-quarter-ahead returns may alreadycapture the initial pickup in economic activity.
14 International Journal of Central Banking February 2017
Tab
le4.
Pre
dic
tive
Reg
ress
ions:
Mon
thly
Sto
ckR
eturn
s,19
95–2
012
BFC
IB
FC
I+C
FSI
MS
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−0.
432∗
4.60
4−
0.28
84.
629
−0.
180
4.64
90.
022
4.65
2Fin
ance
−0.
624∗
4.51
7−
0.42
0∗4.
569
−0.
778∗
4.55
0−
0.24
1∗4.
611
Con
stru
ctio
n−
0.16
48.
079
−0.
054
8.08
20.
291
8.07
7−
0.22
78.
079
Man
ufac
turi
ng−
0.26
37.
265
−0.
055
7.27
60.
186
7.27
4−
0.41
5∗7.
262
Tra
nspo
rtat
ion
−0.
315
6.12
2−
0.11
96.
138
0.00
56.
141
−0.
369∗
6.12
7W
hole
sale
Tra
de−
0.30
36.
676
−0.
100
6.69
00.
115
6.69
1−
0.30
1∗6.
684
Ret
ailTra
de−
0.19
47.
227
0.02
97.
233
0.11
47.
232
0.01
67.
233
Serv
ices
−0.
108
8.39
30.
111
8.39
30.
481
8.38
0−
0.18
28.
392
AN
FC
IN
FC
IC
LN
FSI
ST
LFSI
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−0.
069
4.65
2−
1.23
9∗4.
596
−0.
117
4.64
4−
0.51
14.
622
Fin
ance
−1.
068∗
4.54
5−
1.84
3∗4.
491
−0.
271∗
4.57
3−
0.66
0∗4.
567
Con
stru
ctio
n−
0.20
58.
081
−0.
500
8.07
7−
0.02
78.
082
0.23
08.
079
Man
ufac
turi
ng0.
020
7.27
6−
0.57
57.
269
0.03
57.
276
0.07
67.
276
Tra
nspo
rtat
ion
0.04
26.
141
−0.
840
6.12
1−
0.05
76.
139
−0.
116
6.13
9W
hole
sale
Tra
de−
0.36
86.
686
−0.
697
6.68
00.
074
6.69
00.
080
6.69
2R
etai
lTra
de−
0.41
77.
226
−0.
223
7.23
20.
100
7.22
90.
444
7.21
8Se
rvic
es0.
279
8.39
2−
0.39
38.
391
0.17
98.
384
0.27
58.
390
(con
tinu
ed)
Vol. 13 No. 1 Assessing and Combining Financial Conditions 15
Tab
le4.
(Con
tinued
)
Cit
iFC
IIM
FFSI
KC
FSI
IMF
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−0.
566
4.60
8−
0.04
14.
650
−0.
583∗
4.60
8−
0.92
0∗4.
565
Fin
ance
−0.
406
4.59
5−
0.20
0∗4.
558
−0.
766∗
4.54
2−
0.77
0∗4.
557
Con
stru
ctio
n0.
030
8.08
30.
078
8.07
70.
063
8.08
2−
0.17
08.
081
Man
ufac
turi
ng−
0.00
77.
276
0.03
97.
275
−0.
013
7.27
6−
0.32
87.
269
Tra
nspo
rtat
ion
−0.
281
6.13
20.
000
6.14
1−
0.20
16.
137
−0.
587
6.11
4W
hole
sale
Tra
de−
0.06
06.
692
−0.
000
6.69
2−
0.04
86.
692
−0.
186
6.69
0R
etai
lTra
de−
0.07
17.
232
0.10
77.
222
0.28
27.
226
0.16
77.
231
Serv
ices
−0.
122
8.39
30.
078
8.38
90.
130
8.39
3−
0.20
98.
392
Note
s:T
heta
ble
repor
tsth
eco
effici
ents
and
root
mea
nsq
uare
der
rors
(bot
hin
%)
ofpr
edic
tive
regr
essi
ons
ofon
e-m
onth
-ahe
adst
ock
retu
rns
onFC
Ile
vels
.Se
eta
bles
1an
d3
for
mor
ede
tails
onth
eFC
Isan
dfo
rth
ede
finit
ion
ofin
dust
rypor
tfol
ios.
Ast
eris
ksin
dica
tesi
gnifi
canc
eat
the
90per
cent
leve
l.St
atis
tica
lsi
gnifi
canc
eis
eval
uate
don
the
basi
sof
hete
rosk
edas
tici
ty-c
onsi
sten
tst
anda
rder
rors
or,w
here
appr
opri
ate,
wit
hth
elo
cal-to
-uni
tyas
ympt
otic
sof
Cam
pbel
lan
dY
ogo
(200
6)(s
eese
ctio
n2
for
deta
ils).
Tim
eper
iod:
July
1995
toJu
ne20
12.
16 International Journal of Central Banking February 2017
Tab
le5.
Pre
dic
tive
Reg
ress
ions:
Mon
thly
Sto
ckR
eturn
s,19
95–2
006
BFC
IB
FC
I+C
FSI
MS
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
0.14
74.
277
0.03
84.
279
0.71
84.
245
0.43
64.
253
Fin
ance
0.18
03.
392
0.27
93.
384
−0.
040
3.39
50.
512
3.35
0C
onst
ruct
ion
0.52
46.
737
0.05
76.
751
1.08
06.
703
−0.
027
6.75
1M
anuf
actu
ring
0.97
96.
980
0.35
17.
019
1.73
96.
907
0.04
57.
028
Tra
nspo
rtat
ion
0.62
55.
983
0.20
96.
002
1.29
15.
928
0.11
96.
004
Who
lesa
leTra
de0.
715
6.17
00.
210
6.19
61.
192
6.13
50.
310
6.19
0R
etai
lTra
de0.
444
6.14
00.
154
6.14
90.
911
6.11
30.
707
6.10
4Se
rvic
es1.
474
8.85
30.
861
8.89
52.
400
8.75
70.
327
8.93
2
AN
FC
IN
FC
IC
LN
FSI
ST
LFSI
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
1.11
7∗4.
228
−0.
070
4.27
90.
045
4.27
8−
0.33
14.
275
Fin
ance
−0.
547
3.38
00.
214
3.39
50.
156
3.38
20.
889
3.36
6C
onst
ruct
ion
0.09
66.
751
1.12
06.
746
0.24
26.
735
1.14
96.
726
Man
ufac
turi
ng0.
856
7.01
03.
335
6.98
10.
434
6.97
81.
724
6.97
4Tra
nspo
rtat
ion
1.02
75.
975
1.16
95.
999
0.24
25.
988
0.77
35.
993
Who
lesa
leTra
de0.
053
6.19
91.
695
6.18
50.
428
6.14
41.
529
6.15
1R
etai
lTra
de−
0.10
76.
151
1.01
56.
146
0.38
96.
105
1.34
76.
113
Serv
ices
1.54
28.
892
3.74
88.
891
0.73
28.
825
2.44
08.
853
(con
tinu
ed)
Vol. 13 No. 1 Assessing and Combining Financial Conditions 17
Tab
le5.
(Con
tinued
)
Cit
iFC
IIM
FFSI
KC
FSI
IMF
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−0.
451
4.25
30.
177
4.26
0−
0.49
34.
268
−0.
929
4.23
9Fin
ance
0.21
83.
388
−0.
069
3.39
2−
0.02
93.
395
0.59
03.
375
Con
stru
ctio
n0.
089
6.75
10.
029
6.75
10.
421
6.74
60.
265
6.74
9M
anuf
actu
ring
0.40
17.
016
0.26
77.
003
1.00
57.
001
0.39
67.
024
Tra
nspo
rtat
ion
−0.
063
6.00
50.
248
5.98
00.
373
6.00
1−
0.37
16.
001
Who
lesa
leTra
de0.
222
6.19
50.
084
6.19
60.
575
6.18
90.
606
6.18
8R
etai
lTra
de−
0.02
06.
151
0.02
96.
151
0.47
96.
144
0.83
86.
129
Serv
ices
0.14
58.
937
0.36
98.
900
1.25
28.
905
0.56
08.
932
Note
s:T
heta
ble
repor
tsth
eco
effici
ents
and
root
mea
nsq
uare
der
rors
(bot
hin
%)
ofpr
edic
tive
regr
essi
ons
ofon
e-m
onth
-ahe
adst
ock
retu
rns
onFC
Ile
vels
.Se
eta
bles
1an
d3
for
mor
ede
tails
onth
eFC
Isan
dfo
rth
ede
finit
ion
ofin
dust
rypor
tfol
ios.
Ast
eris
ksin
dica
tesi
gnifi
canc
eat
the
90per
cent
leve
l.St
atis
tica
lsi
gnifi
canc
eis
eval
uate
don
the
basi
sof
hete
rosk
edas
tici
ty-c
onsi
sten
tst
anda
rder
rors
,or
,w
here
appr
opri
ate,
wit
hth
elo
cal-to
-uni
tyas
ympt
otic
sof
Cam
pbel
lan
dY
ogo
(200
6)(s
eese
ctio
n2
for
deta
ils).
Tim
eper
iod:
July
1995
toD
ecem
ber
2006
.
18 International Journal of Central Banking February 2017
Tab
le6.
Pre
dic
tive
Reg
ress
ions:
Quar
terl
ySto
ckR
eturn
s,19
95–2
012
BFC
IB
FC
I+C
FSI
MS
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−1.
099∗
8.88
2−
0.66
88.
989
−0.
650∗
9.02
70.
800
9.00
5Fin
ance
−1.
401∗
9.04
6−
0.96
09.
188
−2.
191∗
9.04
30.
041
9.31
4C
onst
ruct
ion
−0.
006
15.6
160.
203
15.6
121.
144
15.5
720.
481
15.6
06M
anuf
actu
ring
−0.
025
15.0
420.
565
15.0
161.
181
14.9
940.
436
15.0
34Tra
nspo
rtat
ion
−0.
498
12.5
290.
165
12.5
510.
141
12.5
530.
196
12.5
52W
hole
sale
Tra
de0.
179
14.1
990.
714
14.1
560.
919
14.1
711.
210
14.1
35R
etai
lTra
de0.
620
14.8
851.
278
14.7
781.
101
14.8
751.
897
14.7
60Se
rvic
es0.
462
17.0
741.
090
17.0
021.
927
16.9
772.
035
16.9
32
AN
FC
IN
FC
IC
LN
FSI
ST
LFSI
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−0.
946
9.01
9−
3.64
7∗8.
787
−0.
175
9.04
2−
1.17
18.
972
Fin
ance
−3.
325∗
8.91
7−
4.61
5∗8.
900
−0.
611
9.19
7−
1.38
79.
206
Con
stru
ctio
n−
1.65
715
.558
−1.
267
15.5
970.
432
15.5
811.
217
15.5
66M
anuf
actu
ring
−1.
195
15.0
11−
0.72
615
.036
0.56
714
.980
1.47
914
.966
Tra
nspo
rtat
ion
−1.
208
12.5
15−
2.06
612
.493
0.18
712
.545
0.55
912
.540
Who
lesa
leTra
de−
1.62
614
.140
−0.
311
14.2
000.
713
14.0
971.
882
14.0
71R
etai
lTra
de−
1.26
514
.882
1.25
614
.898
0.87
114
.769
3.08
514
.580
Serv
ices
−1.
424
17.0
51−
0.70
117
.085
0.98
116
.926
1.94
716
.974
(con
tinu
ed)
Vol. 13 No. 1 Assessing and Combining Financial Conditions 19
Tab
le6.
(Con
tinued
)
Cit
iFC
IIM
FFSI
KC
FSI
IMF
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−1.
408
8.90
3−
0.07
89.
047
−1.
467∗
8.91
8−
2.32
4∗8.
756
Fin
ance
−0.
630
9.28
6−
0.49
49.
148
−1.
665∗
9.14
7−
0.74
39.
154
Con
stru
ctio
n1.
202
15.5
530.
232
15.5
940.
602
15.6
03−
0.17
315
.615
Man
ufac
turi
ng1.
201
14.9
780.
384
14.9
811.
092
14.9
980.
034
15.0
42Tra
nspo
rtat
ion
0.05
012
.553
0.20
612
.532
0.05
412
.553
−0.
921
12.5
21W
hole
sale
Tra
de1.
122
14.1
420.
448
14.1
131.
456
14.1
180.
653
14.1
87R
etai
lTra
de1.
309
14.8
390.
650
14.7
392.
559
14.6
711.
777
14.8
13Se
rvic
es0.
861
17.0
610.
440
17.0
191.
339
17.0
320.
235
17.0
88
Note
s:T
heta
ble
repor
tsth
eco
effici
ents
and
root
mea
nsq
uare
der
rors
(bot
hin
%)
ofpr
edic
tive
regr
essi
ons
ofon
e-qu
arte
r-ah
ead
stoc
kre
turn
son
FC
Ile
vels
.Se
eta
bles
1an
d3
for
mor
ede
tails
onth
eFC
Isan
dfo
rth
ede
finit
ion
ofin
dust
rypor
tfol
ios.
Ast
eris
ksin
dica
tesi
gnifi
canc
eat
the
90per
cent
leve
l.St
atis
tica
lsi
gnifi
canc
eis
eval
uate
don
the
basi
sof
hete
rosk
edas
tici
ty-c
onsi
sten
tst
anda
rder
rors
or,w
here
appr
opri
ate,
wit
hth
elo
cal-to
-uni
tyas
ympt
otic
sof
Cam
pbel
land
Yog
o(2
006)
(see
sect
ion
2fo
rde
tails
).T
ime
per
iod:
July
1995
toJu
ne20
12.
20 International Journal of Central Banking February 2017
Tab
le7.
Pre
dic
tive
Reg
ress
ions:
Quar
terl
ySto
ckR
eturn
s,19
95–2
006
BFC
IB
FC
I+C
FSI
MS
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
0.37
58.
351
0.16
88.
355
1.12
28.
316
2.05
8∗7.
984
Fin
ance
1.33
86.
788
1.17
36.
776
−0.
921∗
6.85
31.
698∗
6.57
9C
onst
ruct
ion
3.09
413
.512
1.31
413
.707
2.63
913
.638
1.06
213
.717
Man
ufac
turi
ng4.
004
13.9
131.
732
14.2
255.
299
13.7
991.
906
14.1
57Tra
nspo
rtat
ion
2.07
711
.938
1.02
012
.025
2.76
911
.899
1.73
411
.893
Who
lesa
leTra
de3.
342
12.1
321.
416
12.3
852.
961
12.2
812.
681∗
12.0
52R
etai
lTra
de3.
199
12.5
071.
841
12.6
652.
178
12.7
103.
560∗
12.0
80Se
rvic
es5.
677
17.2
993.
557
17.5
957.
327
17.1
573.
757∗
17.4
12
AN
FC
IN
FC
IC
LN
FSI
ST
LFSI
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−1.
628
8.30
1−
3.30
38.
308
0.25
78.
341
−0.
956
8.34
2Fin
ance
−2.
247
6.75
61.
664
6.87
10.
398
6.84
13.
035
6.70
5C
onst
ruct
ion
−2.
492
13.6
962.
746
13.7
550.
958
13.6
444.
777
13.5
52M
anuf
actu
ring
−1.
420
14.3
155.
456
14.2
621.
710
13.9
336.
194
13.9
77Tra
nspo
rtat
ion
−0.
722
12.0
65−
1.53
212
.065
0.99
011
.912
2.81
311
.985
Who
lesa
leTra
de−
2.84
612
.358
2.63
312
.452
1.50
912
.109
5.26
312
.172
Ret
ailTra
de−
2.43
112
.730
2.91
912
.786
1.54
012
.442
4.96
412
.551
Serv
ices
−1.
607
17.9
604.
453
17.9
442.
613∗
17.2
228.
253
17.4
71
(con
tinu
ed)
Vol. 13 No. 1 Assessing and Combining Financial Conditions 21
Tab
le7.
(Con
tinued
)
Cit
iFC
IIM
FFSI
KC
FSI
IMF
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−1.
548
8.18
60.
073
8.35
5−
1.73
08.
284
−2.
812
8.16
1Fin
ance
0.93
96.
810
−0.
399
6.81
40.
438
6.88
12.
134
6.75
0C
onst
ruct
ion
1.38
813
.693
−0.
299
13.7
551.
901
13.7
231.
514
13.7
42M
anuf
actu
ring
1.87
714
.195
0.19
714
.331
2.81
314
.229
1.48
514
.309
Tra
nspo
rtat
ion
0.07
612
.073
0.09
512
.070
0.50
012
.069
−0.
890
12.0
59W
hole
sale
Tra
de1.
257
12.3
98−
0.12
512
.469
1.67
312
.428
1.71
012
.425
Ret
ailTra
de0.
851
12.7
78−
0.18
512
.803
1.74
512
.763
2.55
312
.707
Serv
ices
1.38
517
.922
0.42
817
.953
3.29
117
.864
1.84
717
.947
Note
s:T
heta
ble
repor
tsth
eco
effici
ents
and
root
mea
nsq
uare
der
rors
(bot
hin
%)
ofpr
edic
tive
regr
essi
ons
ofon
e-qu
arte
r-ah
ead
stoc
kre
turn
son
FC
Ile
vels
.Se
eta
bles
1an
d3
for
mor
ede
tails
onth
eFC
Isan
dfo
rth
ede
finit
ion
ofin
dust
rypor
tfol
ios.
Ast
eris
ksin
dica
tesi
gnifi
canc
eat
the
90per
cent
leve
l.St
atis
tica
lsi
gnifi
canc
eis
eval
uate
don
the
basi
sof
hete
rosk
edas
tici
ty-c
onsi
sten
tst
anda
rder
rors
or,w
here
appr
opri
ate,
wit
hth
elo
cal-to
-uni
tyas
ympt
otic
sof
Cam
pbel
land
Yog
o(2
006)
(see
sect
ion
2fo
rde
tails
).T
ime
per
iod:
July
1995
toD
ecem
ber
2006
.
22 International Journal of Central Banking February 2017
We now discuss whether the FCIs are informative about futureinnovations to the macroeconomic variables we mentioned earlier inthis section. In order to implement the Campbell and Yogo (2006)procedure, the only covariate we include in the predictive regressionsis one of the FCIs, and we account for autocorrelation in macro-economic variable changes by using the residuals from log-changeautoregressions as the dependent variable, where the number of lagsis chosen on the basis of the Schwarz Bayesian information criterion.7
The first set of results (table 8) focuses on one-month-aheadpredictability, with the sample running from 1995 to 2012.8 TheChicago Fed, St. Louis Fed, Kansas City Fed, and InternationalMonetary Fund (IMF) FCI indexes predict all the variables; mostother indexes predict at least three of the six variables. When thesample excludes the 2008 financial crisis (table 9), however, the evi-dence in favor of predictability is weak, with most indexes beingable to predict only one macroeconomic variable, typically indus-trial production. Similar conclusions can be drawn when focusing onquarterly horizons, as shown in tables 10 and 11 (only the MorganStanley index can predict more variables in the short sample thanin the full sample).9
7For the monthly series, we use two lags for total consumer credit, commercialand industrial loans, industrial production, and total manufacturing inventory,and one lag for durable goods orders and housing starts. For the quarterly series,we use two lags for total consumer credit and total manufacturing inventory, onelag for commercial and industrial loans and industrial production, and zero lagsfor durable goods orders and housing starts.
8While we expect a negative relation between a worsening of current financialconditions and future economic activity, some coefficients in tables 8 and 9 arepositive. The reason is that the Campbell and Yogo (2006) procedure requires anegative correlation between current innovations to the dependent variable andthe predictor, hence we sometimes need to multiply the FCI by –1. We system-atically check that the sign of the estimated coefficient is as expected: positiveif the FCI (multiplied by –1 as applicable) signals stress when low, and negativeotherwise.
9In an unreported analysis, we obtain similar results when, instead of exclud-ing the period around the financial crisis, we orthogonalize the innovations rela-tive to the appropriately lagged Chicago Fed National Activity Index, which isa proxy for the current state of the economy. The similarity of the results when(i) excluding the crisis and (ii) controlling for the state of the economy is notsurprising, especially in a relatively short sample like ours, because the financialcrisis also coincided with a severe recession, which exerts a strong leverage effect.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 23
Tab
le8.
Pre
dic
tive
Reg
ress
ions:
Innov
atio
ns
toM
onth
lyLog
-Chan
ges
inM
acro
econ
omic
Var
iable
s,19
95–2
012
BFC
IB
FC
I+C
FSI
MS
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I0.
078∗
0.66
20.
105∗
0.65
20.
072
0.66
9−
0.05
10.
671
Con
sum
erC
redi
t−
0.03
40.
519
−0.
032
0.51
9−
0.04
70.
519
0.01
80.
521
Dur
able
Goo
ds−
0.88
3∗3.
733
−0.
765∗
3.78
1−
0.92
0∗3.
860
−0.
667∗
3.90
5H
ousi
ngSt
arts
−1.
184∗
6.24
1−
0.89
5∗6.
344
−0.
759∗
6.45
6−
0.76
4∗6.
448
Ind.
Pro
duct
ion
−0.
120∗
0.64
5−
0.09
4∗0.
654
−0.
167∗
0.64
9−
0.12
6∗0.
657
Tot
alIn
vent
ory
−0.
059∗
0.43
7−
0.05
4∗0.
438
−0.
043
0.44
4−
0.07
1∗0.
439
AN
FC
IN
FC
IC
LN
FSI
ST
LFSI
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I0.
059
0.67
20.
240∗
0.65
8−
0.03
80.
667
0.18
0∗0.
647
Con
sum
erC
redi
t−
0.02
50.
521
−0.
129∗
0.51
6−
0.01
70.
520
−0.
073∗
0.51
6D
urab
leG
oods
−1.
147∗
3.87
5−
2.36
6∗3.
725
−0.
317∗
3.90
0−
1.15
3∗3.
786
Hou
sing
Star
ts−
1.87
5∗6.
342
−2.
984∗
6.26
5−
0.32
1∗6.
457
−1.
220∗
6.37
6In
d.P
rodu
ctio
n−
0.05
10.
670
−0.
321∗
0.64
50.
072∗
0.64
9−
0.16
2∗0.
650
Tot
alIn
vent
ory
−0.
050
0.44
4−
0.14
3∗0.
438
−0.
021
0.44
3−
0.07
3∗0.
439
(con
tinu
ed)
24 International Journal of Central Banking February 2017
Tab
le8.
(Con
tinued
)
Cit
iFC
IIM
FFSI
KC
FSI
IMF
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I0.
119∗
0.66
00.
034∗
0.66
10.
168∗
0.64
70.
183∗
0.64
9C
onsu
mer
Cre
dit
−0.
034
0.52
0−
0.01
30.
519
−0.
061∗
0.51
7−
0.06
8∗0.
517
Dur
able
Goo
ds−
0.83
4∗3.
860
−0.
335∗
3.77
2−
1.24
7∗3.
731
−1.
356∗
3.74
6H
ousi
ngSt
arts
1.16
2∗6.
369
−0.
414∗
6.31
7−
1.28
2∗6.
349
−1.
390∗
6.35
9In
d.P
rodu
ctio
n0.
180∗
0.64
0−
0.04
4∗0.
651
0.17
7∗0.
643
−0.
213∗
0.63
8Tot
alIn
vent
ory
−0.
075∗
0.43
8−
0.01
7∗0.
442
−0.
071∗
0.43
9−
0.10
6∗0.
434
Note
s:T
heta
ble
repor
tsth
eco
effici
ents
and
root
mea
nsq
uare
der
rors
(bot
hin
%)
ofpr
edic
tive
regr
essi
ons
ofon
e-m
onth
-ahe
adin
nova
tion
sto
log-
chan
ges
inm
acro
econ
omic
vari
able
son
FC
Ile
vels
.Se
eta
bles
1an
d3
for
mor
ede
tails
onth
eFC
Isan
don
the
mac
roec
onom
icva
riab
les.
Ast
eris
ksin
dica
tesi
gnifi
canc
eat
the
90per
cent
leve
l.St
atis
tica
lsi
gnifi
canc
eis
eval
uate
don
the
basi
sof
hete
rosk
edas
tici
ty-c
onsi
sten
tst
anda
rder
rors
or,w
here
appr
opri
ate,
wit
hth
elo
cal-to
-uni
tyas
ympt
otic
sof
Cam
pbel
lan
dY
ogo
(200
6)(s
eese
ctio
n2
for
deta
ils).
Tim
eper
iod:
July
1995
toJu
ne20
12.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 25
Tab
le9.
Pre
dic
tive
Reg
ress
ions:
Innov
atio
ns
toM
onth
lyLog
-Chan
ges
inM
acro
econ
omic
Var
iable
s,19
95–2
006
BFC
IB
FC
I+C
FSI
MS
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I−
0.08
50.
583
−0.
087∗
0.58
00.
042
0.58
6−
0.04
10.
586
Con
sum
erC
redi
t0.
030
0.33
20.
040
0.33
10.
005
0.33
3−
0.02
90.
332
Dur
able
Goo
ds0.
339
3.68
30.
309
3.68
10.
420
3.68
1−
0.17
03.
690
Hou
sing
Star
ts0.
225
5.24
90.
382
5.23
80.
414
5.24
40.
455
5.23
0In
d.P
rodu
ctio
n0.
138∗
0.54
70.
059
0.55
6−
0.09
20.
555
−0.
082∗
0.55
2Tot
alIn
vent
ory
−0.
040
0.38
3−
0.05
20.
380
−0.
022
0.38
4−
0.07
5∗0.
375
AN
FC
IN
FC
IC
LN
FSI
ST
LFSI
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I−
0.05
80.
586
0.21
70.
585
−0.
057∗
0.57
7−
0.22
1∗0.
576
Con
sum
erC
redi
t−
0.05
80.
332
−0.
118
0.33
20.
015
0.33
2−
0.01
60.
333
Dur
able
Goo
ds0.
179
3.69
32.
128
3.65
7−
0.18
53.
677
0.66
73.
679
Hou
sing
Star
ts0.
120
5.25
20.
465
5.25
10.
145
5.24
5−
0.82
75.
236
Ind.
Pro
duct
ion
−0.
054
0.55
90.
363∗
0.55
20.
058∗
0.54
80.
153
0.55
4Tot
alIn
vent
ory
0.04
80.
383
−0.
179
0.38
2−
0.03
3∗0.
379
−0.
098
0.38
1
(con
tinu
ed)
26 International Journal of Central Banking February 2017
Tab
le9.
(Con
tinued
)
Cit
iFC
IIM
FFSI
KC
FSI
IMF
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I0.
076
0.58
20.
009
0.58
70.
128
0.58
20.
227∗
0.57
0C
onsu
mer
Cre
dit
0.02
10.
333
0.02
0∗0.
330
0.05
80.
332
0.05
50.
332
Dur
able
Goo
ds−
0.27
43.
683
0.18
73.
670
0.87
63.
655
−0.
807
3.65
9H
ous.
Star
ts−
0.25
65.
246
−0.
091
5.24
90.
144
5.25
20.
490
5.24
4In
d.P
rodu
ctio
n0.
132∗
0.54
20.
041∗
0.55
20.
169∗
0.55
0−
0.14
7∗0.
552
Tot
alIn
vent
ory
0.05
20.
380
−0.
016
0.38
3−
0.07
90.
381
−0.
149∗
0.37
3
Note
s:T
heta
ble
repor
tsth
eco
effici
ents
and
root
mea
nsq
uare
der
rors
(bot
hin
%)
ofpr
edic
tive
regr
essi
ons
ofon
e-m
onth
-ahe
adin
nova
tion
sto
log-
chan
ges
inm
acro
econ
omic
vari
able
son
FC
Ile
vels
.Se
eta
bles
1an
d3
for
mor
ede
tails
onth
eFC
Isan
don
the
mac
roec
onom
icva
riab
les.
Ast
eris
ksin
dica
tesi
gnifi
canc
eat
the
90per
cent
leve
l.St
atis
tica
lsi
gnifi
canc
eis
eval
uate
don
the
basi
sof
hete
rosk
edas
tici
ty-c
onsi
sten
tst
anda
rder
rors
or,w
here
appr
opri
ate,
wit
hth
elo
cal-to
-uni
tyas
ympt
otic
sof
Cam
pbel
lan
dY
ogo
(200
6)(s
eese
ctio
n2
for
deta
ils).
Tim
eper
iod:
July
1995
toD
ecem
ber
2006
.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 27
Tab
le10
.P
redic
tive
Reg
ress
ions:
Innov
atio
ns
toQ
uar
terl
yLog
-Chan
ges
inM
acro
econ
omic
Var
iable
s,19
95–2
012
BFC
IB
FC
I+C
FSI
MS
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I0.
406∗
1.22
90.
477∗
1.16
10.
571∗
1.26
0−
0.41
4∗1.
305
Con
sum
erC
redi
t−
0.09
91.
003
−0.
122
0.99
6−
0.12
11.
007
−0.
029
1.01
4D
urab
leG
oods
−1.
858∗
5.30
8−
1.63
7∗5.
483
−1.
413∗
5.89
6−
1.45
0∗5.
840
Hou
sing
Star
ts−
1.37
3∗8.
645
−0.
803
8.82
2−
1.07
7∗8.
846
0.78
18.
870
Ind.
Pro
duct
ion
−0.
248∗
1.29
1−
0.16
71.
324
−0.
371∗
1.29
6−
0.34
8∗1.
290
Tot
alIn
vent
ory
−0.
285∗
1.11
2−
0.25
7∗1.
128
−0.
238∗
1.17
5−
0.19
71.
179
AN
FC
IN
FC
IC
LN
FSI
ST
LFSI
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I0.
548∗
1.31
41.
238∗
1.17
60.
204∗
1.29
80.
752∗
1.15
5C
onsu
mer
Cre
dit
−0.
064
1.01
4−
0.32
90.
996
−0.
014
1.01
4−
0.17
40.
999
Dur
able
Goo
ds−
2.64
3∗5.
679
−4.
935∗
5.31
1−
0.70
8∗5.
824
−2.
278∗
5.60
4H
ousi
ngSt
arts
−2.
657
8.65
1−
4.17
1∗8.
562
−0.
386
8.86
5−
0.78
48.
878
Ind.
Pro
duct
ion
−0.
061
1.34
9−
0.65
7∗1.
292
−0.
163∗
1.29
1−
0.28
9∗1.
317
Tot
alIn
vent
ory
0.42
0∗1.
151
−0.
578∗
1.15
0−
0.10
1∗1.
175
−0.
258∗
1.17
1
(con
tinu
ed)
28 International Journal of Central Banking February 2017
Tab
le10
.(C
ontinued
)
Cit
iFC
IIM
FFSI
KC
FSI
IMF
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I0.
526∗
1.24
60.
201∗
1.19
10.
780∗
1.11
70.
812∗
1.13
2C
onsu
mer
Cre
dit
−0.
089
1.01
0−
0.03
81.
006
−0.
145
1.00
3−
0.20
8∗0.
994
Dur
able
Goo
ds−
1.63
5∗5.
765
−0.
635∗
5.63
7−
2.41
5∗5.
508
−2.
866∗
5.36
9H
ousi
ngSt
arts
−1.
025
8.83
4−
0.37
58.
815
−1.
163
8.82
9−
2.01
1∗8.
690
Ind.
Pro
duct
ion
−0.
384∗
1.27
4−
0.11
7∗1.
285
−0.
393∗
1.28
5−
0.43
0∗1.
281
Tot
alIn
vent
ory
0.33
4∗1.
136
−0.
094∗
1.15
3−
0.28
2∗1.
163
−0.
375∗
1.14
1
Note
s:T
heta
ble
repor
tsth
eco
effici
ents
and
root
mea
nsq
uare
der
rors
(bot
hin
%)
ofpr
edic
tive
regr
essi
ons
ofon
e-qu
arte
r-ah
ead
inno
vati
ons
tolo
g-ch
ange
sin
mac
roec
onom
icva
riab
les
onFC
Ile
vels
.Se
eta
bles
1an
d3
for
mor
ede
tails
onth
eFC
Isan
don
the
mac
roec
onom
icva
riab
les.
Ast
eris
ksin
dica
tesi
gnifi
canc
eat
the
90per
cent
leve
l.St
atis
tica
lsi
gnifi
canc
eis
eval
uate
don
the
basi
sof
hete
rosk
edas
tici
ty-c
onsi
sten
tst
anda
rder
rors
or,w
here
appr
opri
ate,
wit
hth
elo
cal-to
-uni
tyas
ympt
otic
sof
Cam
pbel
lan
dY
ogo
(200
6)(s
eese
ctio
n2
for
deta
ils).
Tim
eper
iod:
July
1995
toJu
ne20
12.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 29
Tab
le11
.P
redic
tive
Reg
ress
ions:
Innov
atio
ns
toQ
uar
terl
yLog
-Chan
ges
inM
acro
econ
omic
Var
iable
s,19
95–2
006
BFC
IB
FC
I+C
FSI
MS
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I0.
431∗
1.11
10.
399∗
1.09
50.
436
1.12
7−
0.29
21.
118
Con
sum
erC
redi
t0.
058
0.60
70.
028
0.60
90.
094
0.60
6−
0.06
50.
605
Dur
able
Goo
ds0.
956
4.42
6−
0.96
04.
389
0.74
84.
469
−0.
768∗
4.40
8H
ousi
ngSt
arts
1.51
35.
349
1.75
8∗5.
190
1.21
05.
434
1.50
6∗5.
203
Ind.
Pro
duct
ion
−0.
261
1.09
7−
0.09
61.
115
−0.
087
1.11
8−
0.21
6∗1.
089
Tot
alIn
vent
ory
−0.
364∗
0.87
6−
0.37
4∗0.
845
−0.
232
0.91
5−
0.31
6∗0.
851
AN
FC
IN
FC
IC
LN
FSI
ST
LFSI
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I0.
061
1.17
21.
375∗
1.11
0−
0.19
1∗1.
109
0.76
1∗1.
104
Con
sum
erC
redi
t−
0.17
60.
601
−0.
352
0.60
20.
030
0.60
70.
027
0.60
9D
urab
leG
oods
0.48
04.
494
3.93
04.
372
−0.
492∗
4.39
6−
1.33
74.
450
Hou
sing
Star
ts0.
501
5.49
92.
044
5.47
8−
0.33
65.
467
2.54
65.
347
Ind.
Pro
duct
ion
−0.
160
1.11
6−
0.69
31.
103
−0.
118
1.09
5−
0.31
01.
108
Tot
alIn
vent
ory
0.00
90.
931
−1.
189∗
0.87
2−
0.16
5∗0.
872
−0.
625∗
0.87
3
(con
tinu
ed)
30 International Journal of Central Banking February 2017
Tab
le11
.(C
ontinued
)
Cit
iFC
IIM
FFSI
KC
FSI
IMF
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
C&
I0.
344∗
1.11
10.
137∗
1.12
10.
699∗
1.08
50.
810∗
1.05
2C
onsu
mer
Cre
dit
0.05
00.
607
−0.
047
0.59
80.
191
0.59
7−
0.06
60.
608
Dur
able
Goo
ds−
0.41
34.
481
−0.
320
4.43
21.
607
4.38
6−
2.05
4∗4.
308
Hou
sing
Star
ts0.
255
5.50
0−
0.08
55.
503
0.51
55.
497
1.04
35.
467
Ind.
Pro
duct
ion
0.36
2∗1.
049
−0.
142∗
1.06
2−
0.42
6∗1.
087
−0.
303
1.10
3Tot
alIn
vent
ory
0.31
7∗0.
865
0.14
2∗0.
861
0.52
3∗0.
870
−0.
605∗
0.84
7
Note
s:T
heta
ble
repor
tsth
eco
effici
ents
and
root
mea
nsq
uare
der
rors
(bot
hin
%)
ofpr
edic
tive
regr
essi
ons
ofon
e-qu
arte
r-ah
ead
inno
vati
ons
tolo
g-ch
ange
sin
mac
roec
onom
icva
riab
les
onFC
Ile
vels
.Se
eta
bles
1an
d3
for
mor
ede
tails
onth
eFC
Isan
don
the
mac
roec
onom
icva
riab
les.
Ast
eris
ksin
dica
tesi
gnifi
canc
eat
the
90per
cent
leve
l.St
atis
tica
lsi
gnifi
canc
eis
eval
uate
don
the
basi
sof
hete
rosk
edas
tici
ty-c
onsi
sten
tst
anda
rder
rors
or,w
here
appr
opri
ate,
wit
hth
elo
cal-to
-uni
tyas
ympt
otic
sof
Cam
pbel
lan
dY
ogo
(200
6)(s
eese
ctio
n2
for
deta
ils).
Tim
eper
iod:
July
1995
toD
ecem
ber
2006
.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 31
2.2 Granger Causality Tests
In the final part of our predictability analysis, we study the relationbetween each of the FCIs and selected measures of credit availability.Borio (2014) highlights that the financial cycle can amplify macro-economic fluctuations and contribute to significant economic dislo-cations, and that one key component of the financial cycle is creditavailability. Borio (2014) also suggests that peaks in the financialcycle are closely associated with financial crises, which implies thatthe results of our tests in this section can provide additional evidenceon whether FCIs are a useful signal for an upcoming financial crisis.The measures of credit availability that we consider are consumercredit, mortgage credit, and the issuance of commercial mortgage-backed securities (CMBS), all expressed as a fraction of nominalgross domestic product. The last two variables are directly relatedto the real estate boom that set the stage for the 2008 financial crisis,with one of them also measuring the importance of credit provisionthrough the securitization channel.10
We depart from the methodology of Campbell and Yogo (2006)when studying the relation between the FCIs and the measures ofcredit availability, and we conduct a series of Granger causality testsinstead. The reason is that it is the overall level of credit availabil-ity, rather than its changes or innovations, that contributes to thebuildup of financial vulnerabilities during financial booms and thatsets the stage for future financial crises (see the discussion in Borio2014). By using Granger causality tests based on quarterly vectorautoregressions with an appropriate number of lags, we can focuson levels and still account for the persistence of both the FCIs andthe credit availability variables. We use the procedure described inToda and Yamamoto (1995) to account for the fact that the FCIsand the credit availability variables may not only be persistent butalso potentially integrated in small samples.
10The consumer credit and mortgage credit data is from the Federal Reserve’sZ.1 release. Consumer credit is series LA153166000.Q (household and non-profitorganizations; consumer credit), while mortgage credit is series LA153165105.Q(household and non-profit organizations; home mortgages). CMBS issuance istotal issuance less Fannie Mae/Freddie Mac, resecuritized, and foreign propertyissuance, and it is provided by Commercial Mortgage Alert. Nominal gross domes-tic product is the series GDP from the FRED database of the Federal ReserveBank of St. Louis.
32 International Journal of Central Banking February 2017
For each FCI/credit variable pair, we select the optimal num-ber of vector autoregression lags (m) with the Schwarz Bayesianinformation criterion. We then use augmented Dickey-Fuller teststo determine each series’ order of integration, and define n as themaximum order of integration for the two series. Next, we estimatea vector autoregression with m + n lags. Using heteroskedasticity-consistent standard errors, we construct Wald test statistics to deter-mine the joint statistical significance of the coefficients on the firstm lags of the FCI in the equation for the credit availability variable,and the joint statistical significance of the coefficients on the first mlags of the credit availability variable in the equation for the FCI.These are the Granger causality tests, and the null hypothesis isthat the coefficients are jointly zero. Table 12 reports the resultingp-values.
The top panel of table 12 shows the p-values of tests of whetherthe FCIs Granger-cause the indicated credit availability measure,and the bottom panel shows the results of tests that the credit avail-ability measures Granger-cause the FCIs. In the left panels we usethe full sample, which runs from 1995 to 2012. In the right panelswe evaluate to what extent the results are affected by observationscorresponding to the financial crisis by setting the 2008 observa-tions of the dependent variables to missing.11 We assess the robust-ness of the results to the financial crisis by excluding only one year(2008), because the credit availability variables have a strong trendbefore the crisis and only after 2007 do they exhibit significant vari-ation (Borio 2014 highlights that the financial cycle evolves moreslowly than the economic cycle). Truncating the sample after 2007would have eliminated the variation that is likely most importantfor identifying the effect we are interested in.
In the full sample, consumer credit, mortgage credit, and CMBSissuance are Granger-caused by five, four, and three FCIs, respec-tively. After setting the 2008 observations of the credit availabilitymeasures to missing, no FCIs Granger-cause CMBS issuance, andthree out of twelve still Granger-cause consumer credit and mort-gage credit. Focusing on whether the FCIs are Granger-caused bycredit availability, only six tests of the seventy-two that we report
11We set the observations to missing rather than dropping them to avoid using2007:Q4 as the first lag for 2009:Q1.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 33
Tab
le12
.FC
Isan
dM
easu
res
ofC
redit
Ava
ilab
ility:G
range
rC
ausa
lity
Tes
ts
Full
Sam
ple
Excl
udin
g20
08
CC
R/
MT
G/
CM
BS/
CC
R/
MT
G/
CM
BS/
GD
PG
DP
GD
PG
DP
GD
PG
DP
FC
IsG
rang
er-C
ausi
ngC
redi
tA
vaila
bilit
y
BFC
I0.
120.
170.
340.
360.
200.
66B
FC
I+0.
200.
160.
690.
720.
620.
75C
FSI
0.62
0.51
0.06
∗0.
680.
570.
16M
SFC
I0.
650.
740.
980.
870.
800.
59C
LN
FSI
0.21
0.90
0.56
0.21
0.93
0.71
NFC
I0.
03∗∗
0.06
∗0.
100.
07∗
0.20
0.23
AN
FC
I0.
04∗∗
0.40
0.07
∗0.
270.
06∗
0.30
STLFSI
0.13
0.02
∗∗0.
210.
200.
160.
16K
CFSI
0.07
∗0.
06∗
0.16
0.13
0.15
0.33
Cit
iFC
I0.
01∗∗
∗0.
430.
230.
02∗∗
0.26
0.73
IMF
FC
I0.
01∗∗
∗0.
140.
09∗
0.01
∗∗∗
0.08
∗0.
11IM
FFSI
0.34
0.06
∗0.
510.
460.
05∗∗
0.61
(con
tinu
ed)
34 International Journal of Central Banking February 2017Tab
le12
.(C
ontinued
)
Full
Sam
ple
Excl
udin
g20
08
CC
R/
MT
G/
CM
BS/
CC
R/
MT
G/
CM
BS/
GD
PG
DP
GD
PG
DP
GD
PG
DP
Cre
ditA
vaila
bilit
yG
rang
er-C
ausi
ngFC
Is
BFC
I0.
420.
190.
930.
170.
270.
63B
FC
I+0.
230.
09∗
0.66
0.07
∗0.
320.
45C
FSI
0.27
0.82
0.33
0.39
0.73
0.56
MS
FC
I0.
110.
800.
780.
07∗
0.56
0.34
CLN
FSI
0.17
0.72
0.25
0.19
0.54
0.12
NFC
I0.
610.
300.
810.
410.
780.
84A
NFC
I0.
950.
440.
180.
770.
920.
50ST
LFSI
0.29
0.14
0.61
0.15
0.36
0.52
KC
FSI
0.98
0.06
∗0.
500.
270.
710.
70C
itiFC
I0.
100.
720.
340.
10∗
0.94
0.65
IMF
FC
I0.
560.
05∗∗
0.33
0.41
0.16
0.26
IMF
FSI
0.75
0.40
0.73
0.80
0.53
0.89
Note
s:T
heta
ble
show
sp-
valu
esfr
omte
sts
that
the
give
nFC
IG
rang
er-c
ause
sth
ein
dica
ted
mea
sure
ofcr
edit
avai
labi
lity
(top
pane
l)an
dth
atth
egi
ven
cred
itav
aila
bilit
ym
easu
reG
rang
er-c
ause
sth
ein
dica
ted
FC
I(b
otto
mpa
nel)
.T
hem
easu
res
are
cons
umer
cred
itov
erno
min
alG
DP
(CC
R/G
DP
),m
ortg
age
cred
itov
erno
min
alG
DP
(MT
G/G
DP
),an
dis
suan
ceof
com
mer
cial
mor
tgag
e-ba
cked
secu
riti
esov
erno
min
alG
DP
(CM
BS/
GD
P).
We
choo
seth
enu
mber
ofla
gsfo
rth
eve
ctor
auto
regr
essi
ons
unde
rlyi
ngth
ete
sts
bysu
mm
ing
the
opti
mal
lag
chos
enw
ith
the
Schw
arz
Bay
esia
nin
form
atio
ncr
iter
ion
and
the
max
imum
orde
rof
inte
grat
ion
ofth
etw
ose
ries
inth
eve
ctor
auto
regr
essi
onch
osen
wit
hau
gmen
ted
Dic
key-
Fulle
rte
sts
(Tod
aan
dY
amam
oto
1995
).W
eco
nclu
deth
atth
ere
isG
rang
erca
usal
ity
ifa
Wal
dte
stre
ject
sth
enu
llhy
pot
hesi
sth
atth
ela
gged
valu
esof
the
inde
pen
dent
vari
able
are
equa
lto
zero
.N
ote
that
,fo
llow
ing
Tod
aan
dY
amam
oto
(199
5),
we
only
test
that
the
first
mla
gsar
eeq
ual
toze
ro,
whe
rem
isth
enu
mber
ofla
gsid
enti
fied
asop
tim
alw
ith
the
Schw
arz
crit
erio
n.T
hesy
mbol
s**
*,**
,an
d*
indi
cate
stat
isti
calsi
gnifi
canc
eat
the
1per
cent
,5
per
cent
,an
d10
per
cent
leve
ls,re
spec
tive
ly.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 35
in the bottom panel reject the null of no Granger causality. Theseresults suggest that, if we exclude 2008, neither the FCIs nor thecredit availability measures Granger-cause each other. The reasonis that only six tests out of thirty-six (16.7 percent) reject the nullthat the FCIs do not Granger-cause the credit variables, and thenumber drops to three out of thirty-six (8.3 percent) when we testwhether the credit variables Granger cause the FCIs. In both cases,the incidence of rejections is roughly in line with the expected num-ber of false positives for tests that have a significance level of 10percent. In the full sample, the evidence is less clear on whether theFCIs Granger-cause the credit variables, because twelve tests out ofthirty-six (33.3 percent) reject the null of no Granger causality. Thenumber of tests that reject the null that the credit variables do notGranger-cause the FCIs remains three out of thirty-six.
2.3 Interpreting Our Results
Overall, our opinion is that the empirical evidence in favor of theFCIs having reliable predictive power is weak. Given that the FCIsare built by combining public data for typically highly liquid finan-cial instruments, they can hardly be characterized as containingprivileged information. We discuss three possible explanations forthe predictability we find in some of our results: threshold effects,data mining, and non-synchronous trading. We ultimately concludethat those results are mainly driven by data mining and by non-synchronous trading.
The predictability could be the result of threshold effects, in thatfinancial conditions matter only after they deteriorate sufficiently.12
In figure 2 we show a scatter plot of innovations to log-changesof total inventories against the St. Louis FCI, where observationsfor which the index is above its 80th percentile are highlighted bydifferent shapes. Triangles indicate an observation from between
12In Kashyap, Lamont, and Stein (1994), for instance, financing constraints atthe firm level only become binding in tight-money environments. Hubrich andTetlow (2012) find that macroeconomic dynamics crucially depend on financialstress, with stress being “of negligible importance in ‘normal’ times, but of criticalimportance when the economy is in a high-stress . . . state” (page 30).
36 International Journal of Central Banking February 2017
Figure 2. Threshold Effects
Notes: The plot shows a scatter of innovations to log-changes in total invento-ries against the Federal Reserve Bank of St. Louis’s Financial Stress Index. Thesquares and triangles correspond to observations for which the index is above its80th percentile, with the triangles indicating an observation from between 2007and 2012 and the squares indicating an observation from outside that range. Timeperiod: July 1995 to June 2012.
2007 and 2012, and squares indicate an observation from outsidethat range.13
The evidence in figure 2 does not clearly point to a thresh-old effect. There is indeed a negative relationship between the FCIand log-changes in inventories; however, this negative relationship isstrong only in the observations from the recent crisis. One possibleexplanation is that the variables underlying the many FCIs con-structed after the 2008 crisis were chosen based on their movementsduring the crisis. For instance, the FCIs typically include variableslike the spread between three-month LIBOR and the three-monthTreasury-bill rate (the TED spread), which had unusually largemovements during the 2008 crisis. Figure 3 shows the time series of
13Figure 2 is representative of other index/variable combinations. The unre-ported figures are available upon request from the authors.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 37
Figure 3. Time Series of the TED Spread
Notes: The graph shows the time series of the spread, in basis points, betweenthe three-month USD LIBOR and the three-month U.S. Treasury yield, which iscommonly referred to as the “TED spread.” Data are from the British Bankers’Association and the Federal Reserve H.15 statistical release. The vertical lineshighlight five events that range from the Asian crisis in 1997 to the Europeandebt crisis in 2010.
the TED spread from 1986 onwards, and it highlights the unusuallyhigh level of the TED spread that was a feature of the 2008 finan-cial crisis. Unfortunately, in practice it is difficult to distinguish athreshold effect from possible data mining, not least because onlyone major financial crisis is included in the sample.
Another potential source of predictability is non-synchronoustrading across markets. Many FCIs, for instance, include the impliedvolatility index VIX, which is derived from the prices of S&P 500index options. These options trade for fifteen minutes after tradingin the underlying index ends (4:15 p.m. versus 4:00 p.m. easternstandard time)14 with the consequence that VIX on day t containsinformation that will be reflected in stock prices only on day t + 1—afact that can generate spurious predictability (see Atchison, Butler,and Simonds 1987 for a discussion of the effects of non-synchronous
14Details on the trading hours of S&P 500 index options can be found athttp://www.cboe.com/products/indexopts/spx spec.aspx
38 International Journal of Central Banking February 2017
trading on the autocorrelation of equity index returns). We explorethis possibility by running predictive regressions on returns thatexclude the first day of each holding period.
Table 13 shows that now only five of the twelve FCIs have sta-tistically significant predictive power for the finance portfolio, downfrom eleven in table 4, suggesting that non-synchronous trading doesplay a role but is not the sole driver of predictability. In unreportedresults, we find that excluding the first day of each quarter does notchange the significance patterns of tables 6 and 7; most likely, theeffect of non-synchronous trading is diluted by the larger variationof returns measured over longer horizons.
The notion that FCIs only have weak predictive power is alsosupported by a series of Granger causality tests. These tests focuson the dynamic relation between the FCIs and a set of measuresfor the availability of consumer credit, mortgage credit, and CMBSissuance. Two of these variables are closely related to the housingboom that characterized the 2000s, and CMBS issuance also meas-ures the importance of the securitization channel for the provisionof commercial real estate credit. Even though both the variables ofinterest and the statistical approach are different than in the predic-tive regression analysis, this additional set of results supports ourmain conclusion that the FCIs do not have predictive power if oneexcludes the 2008 financial crisis from the sample.
Finally, we should emphasize that we do not conduct an out-of-sample analysis, which would be a more stringent test of pre-dictability, because our in-sample analysis already finds a lack ofreliable predictive power.
3. Combining FCIs
We now turn to the question of how to consolidate the increasinglylarge number of indexes into a single proxy for financial conditions.The FCIs themselves are already an aggregation of underlying vari-ables, and the procedure we describe below can be seen as a higher-level consolidation that aggregates across different variable sets andmethodologies, with the objective of reducing model uncertainty forpolicymakers. As discussed in the introduction, the various FCIsfollow similar long-run trends, but they can give significantly differ-ent readings on financial conditions at a given point in time (e.g.,
Vol. 13 No. 1 Assessing and Combining Financial Conditions 39Tab
le13
.P
redic
tive
Reg
ress
ions:
Mon
thly
Sto
ckR
eturn
sExcl
udin
gth
eFir
stD
ayof
the
Mon
th,19
95–2
012
BFC
IB
FC
I+C
FSI
MS
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−0.
230
4.62
0−
0.11
04.
630
−0.
027
4.63
30.
308
4.62
1Fin
ance
−0.
453∗
4.58
1−
0.26
34.
615
−0.
634∗
4.58
90.
007
4.63
4C
onst
ruct
ion
0.07
18.
295
0.19
88.
289
0.43
88.
283
0.19
28.
293
Man
ufac
turi
ng−
0.06
07.
146
0.12
47.
144
0.31
47.
139
−0.
039
7.14
6Tra
nspo
rtat
ion
−0.
156
6.06
70.
013
6.07
20.
117
6.07
0−
0.05
76.
071
Who
lesa
leTra
de−
0.09
16.
529
0.09
16.
529
0.23
86.
526
0.08
66.
530
Ret
ailTra
de0.
062
7.19
00.
247
7.18
00.
278
7.18
50.
448
7.17
4Se
rvic
es0.
075
8.14
90.
273
8.13
90.
579
8.12
90.
216
8.14
7
AN
FC
IN
FC
IC
LN
FSI
ST
LFSI
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
0.11
34.
633
−0.
670
4.61
7−
0.04
14.
632
−0.
193
4.62
9Fin
ance
−0.
922∗
4.58
0−
1.34
9∗4.
567
−0.
197
4.61
0−
0.38
64.
616
Con
stru
ctio
n0.
039
8.29
50.
188
8.29
50.
078
8.29
30.
602
8.27
1M
anuf
actu
ring
0.13
17.
146
−0.
081
7.14
60.
121
7.14
00.
355
7.13
7Tra
nspo
rtat
ion
0.12
96.
071
−0.
422
6.06
70.
008
6.07
20.
103
6.07
1W
hole
sale
Tra
de−
0.19
96.
529
−0.
186
6.53
00.
147
6.52
10.
361
6.52
0R
etai
lTra
de−
0.17
57.
189
0.36
47.
187
0.18
77.
177
0.76
27.
147
Serv
ices
0.36
98.
145
0.04
28.
150
0.25
88.
127
0.52
18.
132
(con
tinu
ed)
40 International Journal of Central Banking February 2017
Tab
le13
.(C
ontinued
)
Cit
iFC
IIM
FFSI
KC
FSI
IMF
FC
I
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
βFCI
RM
SE
S&P
500
−0.
318
4.61
90.
050
4.63
0−
0.32
94.
619
−0.
633∗
4.59
2Fin
ance
−0.
260
4.62
5−
0.12
14.
612
−0.
540∗
4.59
6−
0.54
74.
603
Con
stru
ctio
n0.
189
8.29
30.
183
8.26
70.
360
8.28
60.
160
8.29
4M
anuf
actu
ring
0.17
37.
144
0.12
57.
131
0.19
27.
143
−0.
085
7.14
6Tra
nspo
rtat
ion
−0.
139
6.07
00.
067
6.06
6−
0.04
16.
071
−0.
399
6.05
9W
hole
sale
Tra
de0.
089
6.53
00.
078
6.52
40.
159
6.52
80.
054
6.53
1R
etai
lTra
de0.
155
7.18
80.
204
7.15
10.
531
7.16
70.
424
7.17
9Se
rvic
es0.
048
8.15
00.
147
8.13
20.
296
8.14
40.
003
8.15
0
Note
s:T
heta
ble
repor
tsth
eco
effici
ents
and
root
mea
nsq
uare
der
rors
(bot
hin
%)
ofpr
edic
tive
regr
essi
ons
ofon
e-m
onth
-ahe
adst
ock
retu
rns
onFC
Ile
vels
.R
etur
nsar
eca
lcul
ated
afte
rex
clud
ing
the
first
day
ofea
chm
onth
.Se
eta
bles
1an
d3
for
mor
ede
tails
onth
eFC
Isan
dfo
rth
ede
finit
ion
ofin
dust
rypor
tfol
ios.
Ast
eris
ksin
dica
tesi
gnifi
canc
eat
the
90per
cent
leve
l.St
atis
tica
lsi
gnifi
canc
eis
eval
uate
don
the
basi
sof
hete
rosk
edas
tici
ty-c
onsi
sten
tst
anda
rder
rors
or,w
here
appr
opri
ate,
wit
hth
elo
cal-to
-uni
tyas
ympt
otic
sof
Cam
pbel
lan
dY
ogo
(200
6)(s
eese
ctio
n2
for
deta
ils).
Tim
eper
iod:
July
1995
toJu
ne20
12.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 41
see figure 1). When aggregating the FCIs, we first select a smallersubsample of indexes, and then search for the “best” combination.
In the first step, we sort the individual indexes on the basis of howwell they capture the information contained in the remaining FCIs.We measure this ability to capture information using the adjustedR2 from regressions15 of changes in the first principal component ofall indexes except for index i on changes in index i. Letting i denotethe FCI of interest, with i = 1, . . . , 12 and “fpc” the first principalcomponent,
PC−i = fpc({FCIj}j �=i)
ΔPC−i = γ + δ × ΔFCIi + εt.
Table 14 reports the adjusted R2s of the regressions described above,which we run on two different samples, 1995–2006 and 1995–2012.The rankings in the two samples are quite similar, with the St. LouisFCI, in particular, having a noticeable margin over the other indexes.We use the ranking to select the five FCIs with the highest adjustedR2 for further aggregation. The two Bloomberg FCIs are rankedamong the top five when the sample includes the period surround-ing the 2008 financial crisis; however, given that the two indexes arebuilt in a similar way, and that BFCI+ ranks noticeably worse thanBFCI in the shorter sample, we exclude BFCI+ from the set of best-performing indexes. We replace BFCI+ with the Kansas City index,which ranks sixth in the 1995–2012 sample and fifth when excludingthe years around the 2008 financial crisis.
Three aspects of our setup warrant further discussion. First, theranking criterion does not compare the FCIs with a benchmark,because financial conditions are an unobservable factor. We also donot use forecasting errors to select or optimally combine the variouspredictors (e.g., Timmermann 2006), because in the first part of thepaper we find that FCIs are not particularly reliable at forecastingeither returns or macroeconomic variables. Second, it is in princi-ple possible that a regression yields a low R2 because index i is aradically better proxy for financial conditions and does not span theremaining FCIs. However, the overlap and the encompassing nature
15We use robust regressions (Hamilton 1991), which reduce the influence ofoutliers.
42 International Journal of Central Banking February 2017
Tab
le14
.Expla
nat
ory
Pow
erof
FC
Is
July
1995
–Dec
.20
06Ju
ly19
95–J
une
2012
ΔP
CR
es.Δ
Ave
rage
Ran
kΔ
PC
Res
.ΔA
vera
geR
ank
BFC
I48
.10
44.3
346
.22
256
.99
52.9
854
.98
4B
FC
I+34
.99
30.7
632
.88
658
.35
54.2
056
.27
3C
itiFC
I34
.43
37.1
835
.81
444
.62
45.3
945
.00
5C
FSI
32.4
929
.44
30.9
724
.70
21.8
923
.30
NFC
I44
.36
41.3
442
.85
365
.07
60.7
962
.93
2A
NFC
I15
.85
11.6
213
.74
35.9
229
.30
32.6
1IM
FFSI
22.3
327
.99
25.1
633
.35
35.6
034
.48
KC
FSI
35.4
431
.61
33.5
25
42.2
336
.41
39.3
26
IMF
FC
I14
.04
19.1
816
.61
29.3
638
.77
34.0
6M
SFC
I17
.79
17.7
817
.79
21.3
424
.40
22.8
7C
LN
FSI
29.9
627
.70
28.8
333
.61
30.9
632
.28
STLFSI
63.7
260
.01
61.8
61
75.6
373
.82
74.7
21
Note
s:T
heta
ble
repor
tsth
ead
just
edR
2s
(in
%)of
robu
stre
gres
sion
s(H
amilt
on19
91)of
(i)ch
ange
sin
the
first
prin
cipa
lcom
pon
ent
ofal
lth
eFC
Is(a
side
from
the
one
indi
cate
din
each
row
)on
chan
ges
inth
ein
dica
ted
FC
I(c
olum
nsla
bel
ed“Δ
PC
”),
and
of(i
i)on
e-la
gau
tore
gres
sive
resi
dual
sof
chan
ges
inth
efir
stpr
inci
palco
mpon
ent
ofal
lth
eFC
Is(a
side
from
the
one
indi
cate
din
each
row
)on
one-
lag
auto
regr
essi
vere
sidu
als
ofth
ech
ange
sin
the
indi
cate
dFC
I(c
olum
nsla
bel
ed“R
es.Δ
”).
Vol. 13 No. 1 Assessing and Combining Financial Conditions 43
of the variables that underlie the different indexes make such a pos-sibility unlikely. Third, we choose to select five FCIs, and the choiceof this number is admittedly arbitrary, but the key point is that weare able to reduce the number of combinations we consider, thuslowering the likelihood that our results are driven by data mining.
In the second step we form all combinations of the five indexesselected above,16 calculate each combination’s first principal compo-nent, and regress17 changes in the first principal component of theFCIs that are not in the combination under consideration (out ofthe twelve we study) on changes in the first principal component ofthe combination. Letting C denote the combination of interest,
PC/∈C = fpc({FCIj}j /∈C)
PC∈C = fpc({FCIj}j∈C)
ΔPC/∈C = γ + δ × ΔPC∈C + εt.
In order to minimize the risk of overfitting, the regressions are runon several subsamples, and we use the resulting set of adjusted R2 toselect the “best” combination of FCIs. Specifically, we calculate, foreach combination and in each subsample, the squared deviation ofthe combination’s adjusted R2 relative to the highest adjusted R2 ineach subsample. We then average, for each combination, the squareddeviations across time periods, and use the averages to identify the“best” composite FCI. Table 15 reports five averages: the first (col-umn A) shows arithmetic averages; in the second (B) the averageis weighted by the ratio of daily S&P 500 return volatility in eachsubsample over the volatility in the full sample; in the third (C) itis weighted by the ratio of the average VIX level in each subsam-ple over the average VIX level in the full sample; in the fourth (D)weights are based on the volatility of daily VIX changes; in the fifth(E) the arithmetic average is calculated on the four non-overlappingsamples (7/95–12/98 through 1/06–6/12).
The criterion we use to rank the FCIs is, of course, one of poten-tially many. For example, we could have selected the FCIs with the
16We form thirty-one different combinations: five individual indexes, ten setsof two indexes, ten sets of three indexes, five sets of four indexes, and one set offive indexes.
17We again use robust regressions (see Hamilton 1991).
44 International Journal of Central Banking February 2017Tab
le15
.Expla
nat
ory
Pow
erof
the
Fir
stP
rinci
pal
Com
pon
ent
ofFC
IC
ombin
atio
ns
7/95
–7/
95–
7/95
–1/
99–
1/02
–1/
06–
Ave
rage
Squar
edD
evia
tion
Com
bin
atio
n6/
1212
/06
12/9
812
/05
12/0
56/
12A
BC
DE
STLFSI
–B
FC
I70
.36
58.6
856
.74
63.4
449
.81
76.6
91.
481.
411.
461.
431.
80ST
LFSI
–N
FC
I76
.12
62.1
760
.36
63.2
453
.72
81.6
80.
840.
800.
840.
791.
14ST
LFSI
–K
CFSI
70.2
757
.41
51.5
552
.77
63.8
071
.02
2.41
2.41
2.49
2.39
3.11
STLFSI
–C
itiFC
I73
.31
54.2
363
.40
53.7
750
.30
83.1
61.
721.
691.
771.
571.
97B
FC
I–
NFC
I62
.43
53.9
548
.65
58.8
945
.55
67.2
03.
453.
393.
433.
463.
99B
FC
I–
KC
FSI
66.1
061
.43
59.8
164
.21
56.6
571
.91
1.30
1.35
1.33
1.36
1.44
BFC
I–
Cit
iFC
I70
.96
57.9
365
.49
51.2
551
.55
77.3
41.
801.
851.
911.
692.
26N
FC
I–
KC
FSI
61.0
348
.32
48.9
148
.74
47.1
463
.46
4.76
4.80
4.85
4.77
5.34
NFC
I–
Cit
iFC
I64
.41
48.9
752
.43
42.6
440
.08
77.5
84.
644.
534.
724.
315.
48C
itiFC
I–
KC
FSI
58.7
845
.10
40.3
439
.41
42.3
665
.57
7.11
7.00
7.22
6.90
8.29
STLFSI
–B
FC
I-
NFC
I71
.32
61.9
660
.88
70.8
350
.99
76.0
40.
910.
880.
880.
901.
12ST
LFSI
–B
FC
I–
KC
FSI
73.7
366
.84
68.0
071
.86
61.5
173
.02
0.41
0.48
0.43
0.49
0.56
STLFSI
–B
FC
I–
Cit
iFC
I77
.09
64.2
970
.55
62.9
455
.97
81.6
60.
480.
490.
500.
440.
65ST
LFSI
–N
FC
I–
KC
FSI
69.6
659
.32
54.6
656
.15
63.3
372
.18
1.85
1.86
1.91
1.85
2.35
STLFSI
–N
FC
I–
Cit
iFC
I78
.31
62.2
272
.37
57.6
255
.16
86.0
70.
770.
790.
830.
691.
04ST
LFSI
–C
itiF
CI
–K
CFSI
71.3
658
.55
52.9
052
.85
54.6
078
.65
2.09
2.03
2.15
1.98
2.74
BFC
I–
NFC
I–
KC
FSI
67.9
664
.93
63.8
570
.14
56.9
768
.97
0.98
1.08
1.01
1.11
1.16
BFC
I–
NFC
I–
Cit
iFC
I71
.05
62.8
667
.08
60.4
455
.34
75.3
20.
940.
990.
990.
931.
18B
FC
I–
Cit
iFC
I–
KC
FSI
74.8
963
.92
72.1
658
.57
57.0
479
.65
0.72
0.76
0.78
0.68
0.99
NFC
I–
Cit
iFC
I–
KC
FSI
69.5
254
.51
55.6
848
.79
47.9
575
.87
2.83
2.80
2.92
2.67
3.53
(con
tinu
ed)
Vol. 13 No. 1 Assessing and Combining Financial Conditions 45
Tab
le15
.(C
ontinued
)
7/95
–7/
95–
7/95
–1/
99–
1/02
–1/
06–
Ave
rage
Squar
edD
evia
tion
Com
bin
atio
n6/
1212
/06
12/9
812
/05
12/0
56/
12A
BC
DE
STLFSI
–B
FC
I–
NFC
I–
KC
FSI
72.8
068
.44
69.1
176
.10
59.1
871
.79
0.46
0.53
0.48
0.55
0.61
STLFSI
–B
FC
I–
NFC
I–
Cit
iFC
I75
.70
67.1
171
.43
68.0
357
.75
81.4
30.
230.
240.
240.
220.
32
STLFSI
–B
FC
I–
Cit
iFC
I–
KC
FSI
74.7
567
.53
73.3
463
.24
60.3
179
.45
0.39
0.43
0.43
0.39
0.55
STLFSI
–N
FC
I–
Cit
iFC
I–
KC
FSI
74.2
862
.26
61.2
256
.73
57.1
780
.84
1.09
1.09
1.15
1.02
1.48
BFC
I–
NFC
I–
Cit
iFC
I–
KC
FSI
76.2
667
.30
73.2
862
.83
57.9
879
.01
0.44
0.48
0.48
0.43
0.65
STLFSI
–B
FC
I–
NFC
I–
Cit
iFC
I-
KC
FSI
75.4
568
.96
72.4
568
.46
60.1
277
.55
0.26
0.29
0.28
0.28
0.36
Note
s:T
hefir
stsi
xco
lum
nsre
por
tth
ead
just
edR
2s
(in
%)
ofro
bust
regr
essi
ons
(Ham
ilton
1991
)of
(i)
chan
ges
inth
efir
stpr
inci
pal
com
pon
ent
ofal
lth
eFC
Is(a
side
from
thos
ein
dica
ted
inea
chro
w)
onch
ange
son
the
first
prin
cipa
lco
mpon
ent
ofth
ein
dica
ted
FC
Ico
mbi
nati
on(t
here
sult
sfo
rin
divi
dual
FC
Isar
eun
tabu
late
d).
The
last
five
colu
mns
show
the
aver
ages
,ac
ross
sub-
per
iods
,of
the
squa
red
devi
atio
nsof
each
com
bina
tion
’sad
just
edR
2fr
omea
chti
me
per
iod’
sla
rges
tad
just
edR
2.C
olum
nA
show
ssi
mpl
eav
erag
esac
ross
the
six
sub-
per
iods
.C
olum
nsB
–Dre
por
tw
eigh
ted
aver
ages
,w
here
the
wei
ghts
assi
gned
toea
chsu
b-per
iod
are
base
don
the
vola
tilit
yof
daily
S&P
500
retu
rns
(B),
onth
eav
erag
eim
plie
dvo
lati
lity
inde
xV
IX(C
),an
don
the
vola
tilit
yof
daily
chan
ges
inV
IX(D
).T
hew
eigh
tsar
eca
lcul
ated
byno
rmal
izin
gth
evo
lati
litie
s/av
erag
esin
each
sub-
per
iod
wit
hth
evo
lati
lity/
aver
age
over
the
full
sam
ple.
Col
umn
Esh
ows
sim
ple
aver
ages
acro
ssth
efo
urno
n-ov
erla
ppin
gsu
b-per
iods
(7/9
5–12
/98
thro
ugh
1/06
–6/1
2).
46 International Journal of Central Banking February 2017
lowest volatility. Such choice would have implied an assumption onthe way financial conditions change over time, namely that theyevolve smoothly. Choosing the FCIs with the highest volatility wouldhave implied that we assume financial conditions can change rapidly,and that we are looking for a more reactive proxy. Precisely to avoidimposing strong assumptions on the nature of the process for finan-cial conditions, we have adopted a criterion that only assumes that(i) all the indexes we study provide some information about financialconditions, and (ii) none of the indexes is likely to contain uniquelyaccurate information about financial conditions.
The results in table 15 show (in bold) that the first principalcomponent of the St. Louis Fed, Bloomberg, Chicago Fed, and Citiindexes has the lowest average squared deviations in all columns(A)–(E): hence we consider such combination our composite FCI(CFCI). Figure 4 highlights that the CFCI follows the generalpattern of the individual FCIs, but its volatility exhibits differentregimes depending on whether financial conditions are loose or tight.The three graphs in figure 4 show the CFCI against three alter-native proxies for financial conditions: STLFSI, which is the best-performing individual FCI; the first principal component of STLFSIand of the index with the lowest correlation with STLFSI (MS FCI);and the implied volatility index VIX, which is one of the variablesunderlying many FCIs. The CFCI tracks STLFSI closely, althoughthe latter is less volatile in the years following the bull market of thelate 1990s. The first principal component of STLFSI and of MS FCIis more volatile than the CFCI, especially in the earlier part of thesample, and it points to much more improved conditions than theCFCI in early 2008, just before the crisis gained full traction.
A comparison of VIX and the CFCI shows that the two trackeach other quite well, with the exception of the period between mid-2007 and late 2008, when VIX remains stable, and the CFCI showsa largely steady deterioration in financial conditions. In addition,the CFCI points to loose financial conditions in the second half ofthe 1990s until late 1998, while, over the same period, VIX pointsto slowly deteriorating conditions starting in 1995. In figure 5, weplot the twenty-four-month exponentially weighted rolling correla-tion between VIX changes and changes in the CFCI, where obser-vations are weighted so that the weight decays by 50 percent everytwelve months (see figure 5 for details). The correlation is initially
Vol. 13 No. 1 Assessing and Combining Financial Conditions 47
Figure 4. The Composite FCI and Other Proxies ofFinancial Conditions
Notes: Each of the three graphs shows the composite FCI (CFCI, solid line)against one of three alternative proxies for financial conditions: the STLFSI (topgraph), which is the individual FCI that best summarizes the information in theremaining FCIs (see table 14); the first principal component of the STLFSI and ofthe index that is least correlated with the STLFSI (the MS FCI—middle graph);and the implied volatility index VIX (bottom graph). VIX data are from theChicago Board Options Exchange.
48 International Journal of Central Banking February 2017
Figure 5. Exponentially Weighted Correlation betweenVIX Changes and Changes in the Composite FCI
Notes: The graph shows exponentially weighted correlation (dotted line) betweenmonthly changes in the volatility index VIX and monthly changes in the com-posite FCI, together with the VIX index, which is standardized for scale reasons(dashed line), and the composite FCI (CFCI, solid line). The correlation is cal-culated on the basis of a twenty-four-month rolling window, where the weightsdecay by 50 percent every twelve months. Specifically, the weights assigned toobservations {t − i}23
i=0 are given by 10.75 · α · (1 − α)i, where α = 1 − e
−ln(4)24 .
low, but it jumps to about 80 percent with the Russian default inAugust 1998. With the exception of two relatively short periods inlate 2000 and 2005/06, it stays mostly above 50 percent, and it hasbeen around 80–90 percent since the events of the late summer of2008.
4. Conclusion
We provide an assessment of the one-month-ahead and one-quarter-ahead predictive power that a selection of financial conditionsindexes (FCIs) have for returns on a broad equity index and a setof equity industry portfolios and for innovations to log-changes inmacroeconomic variables. Our analysis is based on local-to-unityasymptotics, which allows for accurate statistical inference in the
Vol. 13 No. 1 Assessing and Combining Financial Conditions 49
presence of persistent predictors. We find that the evidence for pre-dictive power at the horizons we consider is generally weak, unlessthe financial crisis is included. Further, part of the predictive poweris driven by the effects of non-synchronous trading and, potentially,data mining. We also study the relation between the FCIs andconsumer credit, mortgage credit, and the issuance of commercialmortgage-backed securities, and find only limited evidence that theFCIs Granger-cause these variables. Based on these results, we con-clude that FCIs are better used as aggregate indicators of currentfinancial conditions.
We also observe that, even in the months leading to the 2008financial crisis, different FCIs can provide conflicting assessments offinancial conditions. We suggest a procedure for combining the vari-ous FCIs into a single proxy, which can help reduce the uncertaintyfacing policymakers when monitoring financial conditions. Our pro-cedure builds on the modest assumptions that all the FCIs we studyprovide some information about financial conditions, yet no indexcontains uniquely accurate information.
References
Atchison, M., K. Butler, and R. Simonds. 1987. “NonsynchronousSecurity Trading and Market Index Autocorrelation.” Journal ofFinance 42 (1): 111–18.
Bernanke, B. 1983. “Nonmonetary Effects of the Financial Crisis inthe Propagation of the Great Depression.” American EconomicReview 73 (3): 257–76.
Bernanke, B., and A. Blinder. 1992. “The Federal Funds Rate andthe Channels of Monetary Transmission.” American EconomicReview 82 (4): 901–21.
Bollerslev, T., G. Tauchen, and H. Zhou. 2009. “Expected StockReturns and Variance Risk Premia.” Review of Financial Studies22 (11): 4463–92.
Borio, C. 2014. “The Financial Cycle and Macroeconomics: WhatHave We Learnt?” Journal of Banking and Finance 45: 182–98.
Brave, S., and R. Butters. 2010. “Gathering Insights on the For-est from the Trees: A New Metric for Financial Conditions.”Working Paper No. 2010-07, Federal Reserve Bank of Chicago.
50 International Journal of Central Banking February 2017
Calomiris, C., and J. Mason. 2003. “Consequences of Bank DistressDuring the Great Depression.” American Economic Review 93(3): 937–47.
Campbell, J., and M. Yogo. 2006. “Efficient Tests of Stock ReturnPredictability.” Journal of Financial Economics 81 (1): 27–60.
Campello, M., J. Graham, and C. Harvey. 2010. “The Real Effectsof Financial Constraints: Evidences from a Financial Crisis.”Journal of Financial Economics 97 (3): 470–87.
Cardarelli, R., S. Elekdag, and S. Lall. 2011. “Financial Stress andEconomic Contractions.” Journal of Financial Stability 7 (2):78–97.
Carlson, M., K. Lewis, and W. Nelson. 2014. “Using Policy Inter-vention to Identify Financial Stress.” International Journal ofFinance and Economics 19 (1): 59–72.
Chava, S., and A. Purnanandam. 2011. “The Effect of Banking Crisison Bank-Dependent Borrowers.” Journal of Financial Economics99 (1): 116–35.
D’Antonio, P. 2008. “Appendix.” In EMA Special Report, 26–28.Citigroup Global Markets Inc.
English, W., K. Tsatsaronis, and E. Zoli. 2005. “Assessing the Pre-dictive Power of Measures of Financial Conditions for Macroeco-nomic Variables.” BIS Papers 22 (April): 228–52.
Gertler, M., and S. Gilchrist. 1994. “Monetary Policy, BusinessCycles, and the Behavior of Small Manufacturing Firms.” Quar-terly Journal of Economics 109 (2): 309–40.
Hakkio, C., and W. Keeton. 2009. “Financial Stress: What Is It,How Can It Be Measured, and Why Does It Matter?” EconomicReview (Federal Reserve Bank of Kansas City) (Q2): 5–50.
Hamilton, L. 1991. “srd1: How Robust Is Robust Regression?” StataTechnical Bulletin 2: 21–26.
Hatzius, J., P. Hooper, F. Mishkin, K. Schoenholtz, and M. Wat-son. 2010. “Financial Conditions Indexes: A Fresh Look Afterthe Financial Crisis.” NBER Working Paper No. 16150.
Hubrich, K., and R. Tetlow. 2012. “Financial Stress and EconomicDynamics: The Transmission of Crises.” FEDS Paper No. 2012-82, Board of Governors of the Federal Reserve System.
Vol. 13 No. 1 Assessing and Combining Financial Conditions 51
Kashyap, A., O. Lamont, and J. Stein. 1994. “Credit Conditionsand the Cyclical Behavior of Inventories.” Quarterly Journal ofEconomics 109 (3): 565–92.
Kashyap, A., and J. Stein. 2000. “What Do a Million Observa-tions on Banks Say about the Transmission of Monetary Policy?”American Economic Review 90 (3): 407–28.
Kashyap, A., J. Stein, and D. Wilcox. 1993. “Monetary Policy andCredit Conditions: Evidence from the Composition of ExternalFinance.” American Economic Review 83 (1): 78–98.
Khwaja, A., and A. Mian. 2008. “Tracing the Impact of Bank Liquid-ity Shocks: Evidence from an Emerging Market.” American Eco-nomic Review 98 (4): 1413–42.
Kliesen, K., M. Owyang, and E. Vermann. 2012. “DisentanglingDiverse Measures: A Survey of Financial Stress Indexes.” Review(Federal Reserve Bank of St. Louis) 94 (5): 369–97.
Kliesen, K., and D. Smith. 2010. “Measuring Financial MarketStress.” Economic Synopses (Federal Reserve Bank of St. Louis)(2).
Ludvigson, S., and S. Ng. 2007. “The Empirical Risk-Return Rela-tion: A Factor Analysis Approach.” Journal of Financial Eco-nomics 83 (1): 171–222.
Matheson, T. 2012. “Financial Conditions Indexes for the UnitedStates and Euro Area.” Economics Letters 115 (3): 441–46.
Oet, M., R. Eiben, T. Bianco, D. Gramlich, and S. Ong. 2011. “TheFinancial Stress Index: Identification of Systemic Risk Condi-tions.” Working Paper No. 11-30, Federal Reserve Bank of Cleve-land.
Paravisini, D. 2008. “Local Bank Financial Constraints and FirmAccess to External Finance.” Journal of Finance 63 (5): 2161–93.
Peek, J., and E. Rosengren. 1997. “The International Transmissionof Financial Shocks: The Case of Japan.” American EconomicReview 87 (4): 495–505.
———. 2000. “Collateral Damage: Effects of the Japanese Bank Cri-sis on Real Activity in the United States.” American EconomicReview 90 (1): 30–45.
Rosenberg, M. R. 2009. “Global Financial Market Trends and Pol-icy.” Bloomberg Financial Conditions Watch 2 (6).
52 International Journal of Central Banking February 2017
Stock, J., and M. Watson. 2002. “Macroeconomic Forecasting UsingDiffusion Indexes.” Journal of Business and Economic Statistics20 (2): 147–62.
Timmermann, A. 2006. “Forecast Combinations.” In Handbook ofEconomic Forecasting, Vol. 1, ed. G. Elliott, C. W. J. Granger,and A. Timmermann, 135–96 (chapter 4). North-Holland.
Toda, H., and T. Yamamoto. 1995. “Statistical Inference in VectorAutoregressions.”