apoorva javadekar -how does mutual fund reputation affect subsequent fund flows?

How Does Mutual Fund Reputation Affect SubsequentFund Flows?

Apoorva Javadekar

Boston University

February 8, 2016

Apoorva Javadekar (Boston University) Reputation and Fund Flows February 8, 2016 1 / 34

Introduction

Motivation I: Why Study Mutual Funds?

1 Mutual Funds: Important Vehicle of InvestmentManage 15 Tr $Mutual funds owns 30% US equities Vs 20% direct holdings46% of US household own mutual funds

2 Understand Behavioral Patterns:Investors learn about managerial ability through returns⇒ fund flows shed light on learning, information processing capacitiesetc.

3 Fund Flows Affect Managerial Risk Taking90% funds managers paid as a % of assets⇒ flow patterns can affect risk taking⇒ impacts on asset prices


Introduction

Motivation II: The Paper

1 Existing Literature:Studies link between fund performance and fund flows (flow-schedule)Finds and rationalizes evidence of return chasing and convexity in fundflowsBut not much is known about the importance of performance history(reputation)

2 This Paper:Explore the role of reputation for fund flowsHow history up to t − 1 affect link between time t performance andtime t + 1 flowsCan we explain the evidence?


Introduction

Role of Reputation

Better understanding of managerial incentives:High reputation ⇒ Low P(Getting Fired) (Khorana; 1996, Kostovetsky;2011)My sample: 30% of the fired managers belong to bottom 20%reputation rankBut compensation too determine the incentives and flows affectcompensation⇒ Important to know how reputation affect flows

But can reputation affect flows?Investor Heterogeneity ⇒ investor composition is history-specific⇒ subsequent reactions to fund performance become history-specific


Introduction

Agenda

Empirical EvidenceModelTesting model predictions in dataTests to check validity of model mechanism


Introduction

Literature Review

Return Chasing and flow convexity:Ippolito (1992), Sirri & Tuffano (1998), Chevallier & Ellison (1997)Lack Performance Persistence:Carhart (1997), Bollen & Busse (2004) test short and medium termpersistenceRisk shifting due to convex flows:Brown, Harlow, Starks (1996), Basak (2012)Theoretical Models:

Berk & Green (2004): rationalizes lack of persistence and returnchasing simultaneously using decreasing returns and competitive capitalsupplyLynch & Musto (2003): explains convexity using manager replacement

Berk & Tonks (2007): repeat losers have insensitive flows to theleft of flow-schedule


Empirical Evidence

Variables

Fund Flows:

FLOWit =qit − [qit−1 × (1 + rit)]

qit−1 × (1 + rit)

where rit denotes net of expense fund returns during time t and qitdenotes fund assets at the end of time t.

Fund Performance:Ranks within same ’investment objective’ based on raw net returns(Sirri & Tuffano; 1998)Ranks based upon ’CAPM-Alpha’ (Berk & Binsbergen; 2014)

Ranks are normalized to lie between [0, 1] interval.

Current Performance (Perfit): Based upon current year t

Reputation (reputeit): Based upon 5 year window ending withcurrent year t.


Empirical Evidence

Summary Statistics

Reputation Excess αLT Exp Front Turn σLT Size AgeRetLT Ratio Load over Mn$ Years

LowMean -0.042 -0.038 0.013 0.038 0.886 0.186 670.933 17.268Median -0.041 -0.037 0.013 0.041 0.700 0.176 122.750 12.000

MedMean -0.003 -0.001 0.012 0.038 0.715 0.172 1329.879 17.335Median -0.007 -0.004 0.012 0.043 0.550 0.167 208.500 12.000

TopMean 0.042 0.041 0.012 0.035 0.702 0.175 2019.931 16.014Median 0.031 0.032 0.012 0.038 0.520 0.170 351.650 11.000

Full SampleMean 0.000 0.002 0.012 0.037 0.743 0.175 1368.062 17.027Median -0.005 -0.002 0.012 0.042 0.570 0.169 211.475 12.000


Empirical Evidence

Basic Regression Framework

Objective: Asses impact of reputation starting at time t onflow-schedule for the period t + 1Regression:

FLOWit+1 = a +5∑

j=2φjQjit +

5∑j=2

ψj (Qjit × reputeit−1)

+(γ × reputeit−1) + controlsit + εit+1

Qjit denotes dummy for j th quantile of Perfit

Regression of t + 1 flows on time t recent performance givenreputation starting at time tRegression for each quantile of Perfit to account for non-linearity(Chevallier & Ellison; 1997)


Empirical Evidence

Regression Output: Fund Flows

Panel A: Raw Returns

Q2t − Q1t 0.034*** 0.037*** 0.013(0.006) (0.006) (0.011)

Q3t − Q1t 0.084*** 0.090*** 0.032***(0.007) (0.007) (0.012)

Q4t − Q1t 0.124*** 0.130*** 0.050***(0.007) (0.007) (0.014)

Q5t − Q1t 0.241*** 0.246*** 0.107***(0.010) (0.010) (0.018)

reputet−1 0.202*** 0.083***(0.013) (0.015)

reputet−1 × (Q2t − Q1t ) 0.043**(0.019)

reputet−1 × (Q3t − Q1t ) 0.108***(0.021)

reputet−1 × (Q4t − Q1t ) 0.149***(0.026)

reputet−1 × (Q5t − Q1t ) 0.261***(0.033)

Intercept 0.056 -0.067* -0.008(0.037) (0.035) (0.035)

Adj R2 0.176 0.208 0.215

Results


Empirical Evidence

Main Results Regression Table

Result 1: Significant return chasing effect ignoring reputationinteractions and even after controlling for reputation

Result 2: Return chasing effect is reduced by more than half afterincluding reputation interactions

Result 3: All the interaction terms are large and significantSignificant =⇒ (Qj − Q1|repute = high) > (Qj − Q1|repute = low).Large =⇒ Interaction effect more important than return chasingeffect

Result 4: Coefficients on Interaction term rise monotonically withperformance

⇒ Flow-Schedule more sensitive for higher reputed fundsFlow-schedule sensitive even at the lower end for high reputation fund.


Empirical Evidence

Flow-Schedule Graph


Empirical Evidence

Example

Best Fund: Q5t = 1 and reputet−1 = 0.90

Worst Fund: Q1t = 1 and reputet−1 = 0.10

∆FLOW ≡ FLOW (Best)− FLOW (Worst) = 40.8%

Break-Up:Source Contribution

∆FLOW Due to Return Chasing Effect 10.7%

∆FLOW Due to Reputation Effect 6.6%

∆FLOW Due to Interaction Effect 23.50%


Empirical Evidence

Robustness Checks

Change in Market Share as dependent variable (Spiegel & Zhang;2012) Result

Results valid across age and size categories Result

Results valid even if recent performance is computed over a longerhorizon Result


Empirical Evidence

Model


Model

Set-Up

Manager with unknown skill α and generates gross return as

Rt = α + εt

withεt ∼ N

(0, σ2

ε

)Convex cost of active management: C(x) = ηx2

Net Return Process:

rt = ht−1Rt − f − η[

(ht−1 × qt−1)2

qt−1

]

where ht−1 denotes actively managed share of assets during time t


Model

Investors and Beliefs

Investors:Unit mass of risk neutral investorsµ fraction of Always Attentive (AA)1− µ fraction Occasionally Attentive (OA)Each period, P(attention|OA) = δ < 1Have infinitely deep pockets

Beliefs About Managerial Skill: At the end of time t

α ∼ N(φt , σ2t )

⇒Et(α) = Et(Rt+1) = φt


Model

Mechanism I

Equilibrium Condition When δ = 1: (Berk & Green; 2004)Et(rt+1|ht , φt)=0

Deep pockets ensure that fund receive required inflowsFull attention ensures that no investor invests in negative NPVmanager.

Equilibrium Condition When δ < 1:Et(rt+1|ht , φt)≤0

Deep pockets ensure that no positive expected NPV project existsInattentive investors ⇒ capital outflows could be less than required toattain zero NPV condition

Inattention =⇒ Over-Sized funds relative to competitivebenchmark.


Model

Mechanism II

In equilibrium: Low reputation funds predominantly owned byOA-types

Because AA-types are fast to move out of poor performing funds

Implications For Flows:Dampened outflows after yet another bad performance by lowreputation fundsOver-Sized ⇒ Low required inflows after a good performance

Implications for Persistence:Over-Sized ⇒ Low reputation funds must under-perform


Model

Solution With δ < 1

Initial Investor Composition: A investor’s ownership at t = 0 is

λ0 =µ

µ+ (1− µ)δ︸︷︷︸Attentive Fraction In Economy

Competitive Size and Flows: q∗t satisfy

Et [rt+1|ht , q∗t ] = 0

and required flowse∗

t = q∗t − qt−1(1 + rt)

Attentive Capital:

zt =

λt−1 + (1− λt−1) δ︸︷︷︸Attentive Fraction Within Fund

qt−1(1 + rt)


Model

Investor Composition

Outflows ⇒ λt < λt−1If fund has enough attentive capital:

AA’s Contribution To Outflows =λt−1

λt−1 + (1− λt−1)δ>λt−1

If zt < |e∗t | ⇒ λt = 0 as every attentive investor liquidates

Inflows⇒ λt > λt−1AA-type contribute λ0 of new capital and outflows reduce λ ⇒ λ0 isupper limit of λt−1

λt is a weighted average of λ0 and λt−1

⇒ λt ∈ (λt−1, λ0)

Persistent outflows ⇒ High fraction of Inattentive Investors


Model

Learning and Fund Flows

Belief Updates:

φt = φt−1 +

( rt − Et−1(rt)

ht−1

)(σ2

t−1σ2

t−1 + σ2ε

)︸︷︷︸

=ωt−1

⇒ ∆φt bigger for over-sized funds as Et−1(rt) < 0

Fund Flows: Let qt−1 = q∗t−1 × (1 + ψt−1)

If capital adjustment is complete

FFt =

[1 + ωt−1

(rt2f + ψt−1

2

)]2

(1 + ψt−1)(1 + rt)− 1

In case zt is not enough to support outflows

FFt = − ztqt(1 + rt+1)


Model

Fund Flows Continued

Limited Outflows: Low reputed funds ⇒ low λt−1 =⇒ Flatflow-schedule on the left tail

Dampened Inflows:Over-Size Effect: Low reputed fund ⇒ ψt−1 > 0 =⇒ requiredinflows e∗

t = q∗t − qt−1(1 + rt) are smaller compared to competitively

sized fundLearning Effect: Et−1(rt) < 0 ⇒ q∗

t itself is pushed up for a given rt⇒ e∗

t is higher for a given rt

For reasonable parameter values, Over-Size effect dominates Learningeffect


Model

Flows With Various Parameter Values


Model

Performance Persistence

Reputation Market SMB HML Momentum 4-factor N AdjDecile Beta Beta Beta Beta Alpha R2

D1 (Low) 1.00426*** 0.16568*** -0.02126 0.00836 -0.00137*** 420 0.968(0.01232) (0.01845) (0.02147) (0.01435) (0.00045)

D2 1.00323*** 0.17559*** -0.00004 0.02108 -0.00138*** 420 0.976(0.00988) (0.01873) (0.01886) (0.01535) (0.00039)

D3 1.01012*** 0.14140*** 0.02330 0.01872 -0.00118*** 420 0.976(0.01136) (0.01883) (0.02081) (0.01400) (0.00040)

D4 0.98307*** 0.13459*** 0.03731** 0.00185 -0.00060* 420 0.978(0.01017) (0.01757) (0.01775) (0.01180) (0.00035)

D5 0.97228*** 0.13435*** 0.02788 0.00757 -0.00059 420 0.975(0.01108) (0.02109) (0.01739) (0.01116) (0.00037)

D6 0.96283*** 0.08781*** 0.00442 -0.00417 -0.00039 420 0.972(0.01688) (0.02009) (0.01763) (0.01291) (0.00045)

D7 0.96463*** 0.13536*** 0.01433 0.00991 -0.00022 420 0.974(0.01140) (0.01836) (0.02146) (0.01302) (0.00040)

D8 0.97028*** 0.16909*** -0.01974 0.01421 -0.00048 420 0.977(0.01387) (0.01493) (0.01666) (0.01190) (0.00041)

D9 0.94807*** 0.17254*** -0.02423 -0.00728 0.00023 420 0.972(0.01533) (0.01826) (0.02095) (0.01340) (0.00044)

D10 (Top) 0.98846*** 0.20101*** -0.00393 -0.01694 -0.00018 420 0.969(0.01092) (0.02160) (0.01902) (0.01344) (0.00044)


Model

Calibration Exercise

Parameter Value Source

f 1.76% Data (including loads)

ψlow 0.93 See below

ωt =

(σ2

t−1σ2

t−1+σ2ε

)0.0955 Berk, Green (2004)

δlow (1 − λlow ) + λlow 0.18 Moment Fitting

δhigh(1 − λhigh) + λhigh 0.49 Moment Fitting

Size Distortion ψt :

Et (rt+1)︸︷︷︸−1.64%

= −ηh2t q∗

t ψt = − f︸︷︷︸1.76%

×ψt


Model

Experiments To Validate Model Mechanism

Heterogeneity in Investors ⇒ Heterogeneity in Flows

What events damp this heterogeneity?Managerial Replacement:⇒ media news, and other soft information⇒ higher investor attention even from otherwise inattentive investors⇒ dampened investor heterogeneityLarge Front Loads Large front loads ⇒ potentially more attention byinvestors

In both these cases, interaction between reputation and recentperformance must lose its importance.

Replacement front loads


Model

Concluding Remarks

Return chasing gets stronger with reputationPersistence in poor performance for low-reputation fundsSimple model with inattentive investors explains the heterogeneity inflow-scheduleInteresting to study risk shifting conditional on reputation


Model

Thank You !


Model

Regression With Change in Market Share

Panel A: Raw Returns Panel B: CAPM-Alpha

Q2t − Q1t 0.042 -0.125*** 0.061** -0.085*(0.026) (0.046) (0.026) (0.044)

Q3t − Q1t 0.107*** -0.186** 0.131*** -0.130*(0.032) (0.079) (0.036) (0.071)

Q4t − Q1t 0.258*** -0.158*** 0.276*** -0.110**(0.033) (0.051) (0.035) (0.053)

Q5t − Q1t 0.510*** -0.167** 0.490*** -0.149**(0.046) (0.070) (0.047) (0.069)

reputet−1 -0.048 -0.023(0.060) (0.066)

reputet−1 × (Q2t − Q1t ) 0.326*** 0.297***(0.098) (0.088)

reputet−1 × (Q3t − Q1t ) 0.577*** 0.517***(0.169) (0.166)

reputet−1 × (Q4t − Q1t ) 0.811*** 0.753***(0.124) (0.121)

reputet−1 × (Q5t − Q1t ) 1.309*** 1.195***(0.186) (0.186)

Intercept -0.189 -0.217 -0.220 -0.305(0.240) (0.223) (0.231) (0.221)

Adj R2 0.062 0.088 0.055 0.077

Back to Robustness


Model

Age And Size Robustness With Raw Returns

Panel A: Age Bins Panel B: Size Bins

Young=1 Young=0 Small=1 Small=0Q2t − Q1t 0.004 0.011 -0.001 0.034**

(0.024) (0.012) (0.015) (0.016)Q3t − Q1t 0.029 0.035*** 0.016 0.045***

(0.026) (0.013) (0.017) (0.017)Q4t − Q1t 0.057* 0.046*** 0.041* 0.048***

(0.030) (0.014) (0.023) (0.017)Q5t − Q1t 0.157*** 0.087*** 0.116*** 0.086***

(0.039) (0.019) (0.028) (0.021)reputet−1 0.075*** 0.095*** 0.056** 0.093***

(0.028) (0.018) (0.024) (0.020)reputet−1 × (Q2t − Q1t ) 0.042 0.048** 0.058* 0.014

(0.038) (0.021) (0.031) (0.026)reputet−1 × (Q3t − Q1t ) 0.126*** 0.089*** 0.164*** 0.070**

(0.042) (0.024) (0.036) (0.028)reputet−1 × (Q4t − Q1t ) 0.177*** 0.125*** 0.214*** 0.122***

(0.052) (0.026) (0.053) (0.028)reputet−1 × (Q5t − Q1t ) 0.268*** 0.237*** 0.323*** 0.246***

(0.066) (0.036) (0.057) (0.037)

Intercept 0.096 -0.076** 0.024 -0.044(0.122) (0.039) (0.057) (0.041)

Adj R2 0.209 0.234 0.181 0.268

Back to Robustness


Model

Longer Horizon For Recent Performance


Q2t − Q1t 0.019** 0.008 0.005 0.039*** 0.029*** 0.001(0.008) (0.008) (0.015) (0.008) (0.008) (0.014)

Q3t − Q1t 0.060*** 0.042*** 0.021 0.058*** 0.041*** 0.017(0.009) (0.009) (0.016) (0.008) (0.008) (0.017)

Q4t − Q1t 0.101*** 0.074*** 0.024 0.123*** 0.097*** 0.035*(0.009) (0.009) (0.018) (0.010) (0.010) (0.018)

Q5t − Q1t 0.217*** 0.177*** 0.048* 0.212*** 0.173*** 0.034(0.013) (0.013) (0.028) (0.013) (0.013) (0.028)

reputet−2 0.158*** 0.066*** 0.156*** 0.040*(0.014) (0.022) (0.014) (0.022)

reputet−2 × (Q2t − Q1t ) 0.022 0.079***(0.029) (0.027)

reputet−2 × (Q3t − Q1t ) 0.063** 0.076**(0.030) (0.030)

reputet−2 × (Q4t − Q1t ) 0.117*** 0.144***(0.031) (0.031)

reputet−2 × (Q5t − Q1t ) 0.230*** 0.257***(0.043) (0.044)

Intercept 0.035 -0.035 -0.005 0.002 -0.074** -0.037(0.036) (0.036) (0.036) (0.036) (0.036) (0.037)

Adj. R2 0.329 0.343 0.347 0.326 0.339 0.344

Back to Robustness


Model

Regression With Managerial Replacement

Panel A: Raw Returns Panel B: CAPM-α

Replacement Yes No Yes No

Perft 0.135** 0.123*** 0.169*** 0.146***-0.052 -0.029 -0.061 -0.033

reputet−1 0.007 0.034 -0.013 -0.036-0.046 -0.024 -0.047 -0.023

Perft× reputet−1 0.196* 0.313*** 0.104 0.280***-0.102 -0.05 -0.099 -0.052

Intercept -0.123 0.008 -0.084 -0.009-0.087 -0.043 -0.088 -0.045

N 1136 7014 1136 7014Adj R2 0.158 0.21 0.152 0.208

Back to Experiments


Model

Regressions Across Fee Structures


Front Load Low High Low High

Perft 0.171*** 0.153*** 0.167*** 0.166***(0.042) (0.039) (0.049) (0.039)

reputet−1 0.054 0.096*** 0.058 0.098***(0.036) (0.035) (0.037) (0.032)

Perft× reputet−1 0.268*** 0.140** 0.222*** 0.102(0.071) (0.066) (0.081) (0.067)

Intercept 0.106 -0.057 0.108 -0.092(0.085) (0.066) (0.085) (0.066)

N 2581 2785 2581 2785Adj R2 0.239 0.169 0.223 0.164

Back to Experiments


apoorva javadekar -how does mutual fund reputation affect subsequent fund flows?

Documents