Negativity Bias in Attention Allocation: Retail Investors Reaction to Stock Returns

Transcription

1 Negativity Bias in Attention Allocation: Retail Investors Reaction to Stock Returns Tomás Reyes 1 1 Pontificia Universidad Católica de Chile

2 Research Question Do retail investors display a negativity bias in attention allocation with respect to stock returns? Negativity Bias in Attention Allocation 1 / 28

3 Preview of the Results We use direct measures of attention using search volume from Google. We relate these measures with other proxies for attention. We find that retail investors display a negativity bias with respect to extreme stock returns at the U.S., state, and company level. By redistributing returns, we rule out the bias is due because negative and positive returns are not symmetric. Negativity Bias in Attention Allocation 2 / 28

4 Motivation Hirshleifer et al. (2004) show individuals are net buyers after negative and positive earning surprises. Net purchases are greater after extreme negative earnings surprises than after extreme favorable ones. Barber and Odean (2008) show individuals are net buyers of attention grabbing stocks. Buy-sell imbalances are greater after negative return days than after positive return days. Negativity Bias in Attention Allocation 3 / 28

5 Literature Review Negativity bias in psychology: Pratto and John (1991) find that bad traits attract more attention in an automatic and non-intentional fashion than good traits; Baumeister et al. (2001) argue negative events will produce larger and more intense consequences than positive events. Use of search volume from Google: Varian and Choi (2009) show that search volume can predict home sales, automotive sales, and tourism; Da et al. (2009) use searches on tickers to proxy for attention in individual stocks and show increases in search volume lead to contemporaneous price increases and future return reversals. Negativity Bias in Attention Allocation 4 / 28

6 Measuring Attention: Search Volume Internet users commonly use a search engine to gather information. If you search for a term, you have to be paying attention to it! Google tracks the amount (and geographical origin) of searches for most common terms since Negativity Bias in Attention Allocation 5 / 28

7 Measuring Attention: Search Volume Search Volume Index (SVI) from Google, SVIregion,t term Searchesregion,t term = Total Searches t Constant diet twitter 0.8 Search Volume Index Year Negativity Bias in Attention Allocation 6 / 28

8 Measuring Attention: Proxies We use three aggregate measures of attention: All Retail Investors Attention (AllInv) measures general interest in the stock market and price movements; New Investors Attention (NewInv) relates to people searching information to open a brokerage account; Old Investors Attention (OldInv) concerns to retail investors who already own a brokerage account and use Google to access it. We also use a measure of attention for specific stocks. Negativity Bias in Attention Allocation 7 / 28

9 Measuring Attention: Search Terms AllInv NewInv OldInv best stocks stock prices best online trading ameritrade dow jones stock quotes discount broker charles schwab good stocks yahoo finance discount brokers etrade google finance online broker scottrade hot stocks online brokerage sharebuilder market watch online brokers nasdaq online investing stock market online stock trading stock market news online trading stock market today stock broker Where are these terms coming from? Start with a small set of terms. Obtain related searches recommended by Google. Drop irrelevant terms. Iterate one last time with the remaining terms. Negativity Bias in Attention Allocation 8 / 28

10 Measuring Attention: Aggregate and Abnormal SVI Aggregate SVI across terms, NewInv r,t = j NewInv terms SVI j r,t, r, t. Abnormal attention, ( ) NewInv r,t ANewInv r,t = log q=1 NewInv. r,t q r represents a geographical region, which can be the U.S. or a state. We use eight weeks to compute the base level of attention; our main results are robust to different window lengths. Negativity Bias in Attention Allocation 9 / 28

11 Measuring Attention: Validation TD Ameritrade reports the number of new accounts opened each quarter on forms 10-K/Q since Negativity Bias in Attention Allocation 10 / 28

12 Measuring Attention: Validation Correlation with alternative proxies Abnormal Trading Volume (Barber and Odean (2008); Gervais et al. (2001); Hou et al. (2008)). AVlm t = log(vlm t ) log[mean(vlm t 1,..., Vlm t 8 )] Vlm t is trading volume in the market portfolio. Abnormal VIX, the CBOE market volatility index (De Long et al. (1990)). AVIX t = log(vix t ) log[mean(vix t 1,..., VIX t 8 )] AStockMarket t AOnlineTrading t AEtrade t AVlm t AVIX t AStockMarket t 1 AOnlineTrading t AEtrade t AVlm t AVIX t Negativity Bias in Attention Allocation 11 / 28

13 SVI and Stock Returns: U.S. Level We sort weekly returns of a portfolio of high market capitalization firms (Ret mcap,t, highest quartile) into quintiles (QU i ), I i Ret mcap,t = 1{Ret mcap,t QU i } i {1(low),..., 5(high)} Ret mcap,t sorted by QU i Ret mcap,t -4% -3% -2% -1% 0% 1% 2% 3% 4% 5% Quantile(QU) I 1 Ret mcap,t I 2 Ret mcap,t I 3 Ret mcap,t I 4 Ret mcap,t I 5 Ret mcap,t Ret mcap,t sorted by time Week(t) Ret mcap,t 3% 1% 2% 4% -2% -1% -3% 0% 5% -4% I 1 Ret mcap,t I 2 Ret mcap,t I 3 Ret mcap,t I 4 Ret mcap,t I 5 Ret mcap,t Negativity Bias in Attention Allocation 12 / 28

14 SVI and Stock Returns: U.S. Level Similarly, we define P i Ret mcap,t = Ret mcap,t I i Ret mcap,t Ret mcap,t 3% 1% 2% 4% -2% -1% -3% 0% 5% -4% P 1 Ret mcap,t % 0 0-4% P 2 Ret mcap,t % -1% P 3 Ret mcap,t 0 1% % 0 0 P 4 Ret mcap,t 3% 0 2% P 5 Ret mcap,t % % 0 Regression specification, DepVar t = α + DepVar t = α + 5 β i P i Ret mcap,t 1 + δq FE + ε t ; i=1 5 β i I i Ret mcap,t 1 + δq FE + ε t. i=1 Q FE are quarter dummies. t-statistics are in parenthesis; standard errors are computed using Newey and West (1987). Variables are divided by their standard deviations. Negativity Bias in Attention Allocation 13 / 28

15 SVI and Stock Returns: U.S. Level (1) (2) (3) AAllInv t ANewInv t AOldInv t P 1Ret mcap,t (-0.99) (-0.58) (0.35) P 2Ret mcap,t (0.91) (0.35) (2.04) P 3Ret mcap,t (0.29) (1.00) (-0.72) P 4Ret mcap,t (-0.03) (1.21) (-0.53) P 5Ret mcap,t (-0.51) (1.23) (-0.27) Q FE Yes Yes Yes Adj R-Squared Observations (1) (2) (3) AAllInv t ANewInv t AOldInv t I 1Ret mcap,t (1.38) (-0.14) (-0.00) I 2Ret mcap,t (-0.24) (-1.78) (-0.53) I 3Ret mcap,t (0.27) (-0.29) (0.67) I 4Ret mcap,t (0.22) (0.02) (0.26) Q FE Yes Yes Yes Adj R-Squared Observations t-statistics are in parenthesis; standard errors are computed using double clustering. Negativity Bias in Attention Allocation 14 / 28

16 SVI and Stock Returns: State Level We sort companies by state using location codes from Compustat. For each state and week, we construct a portfolio of high market capitalization firms located within the state (Ret in mcap,t), and another with firms located outside the state and analogous characteristics (Ret out mcap,t). Regression specification, DepVar s,t = α + 5 β i P i Retmcap,s,t 1 in + i=1 + δ 1M FE + δ 2S FE + ε s,t. 5 i=1 β i P i Ret out mcap,s,t 1 + γctrls Monthly state controls are: (i) Coincident Economic Activity Index; (ii) Leading Index; and (iii) Unemployment Rate. M FE and S FE are month and state dummies; standard errors are estimated using double clustering. Negativity Bias in Attention Allocation 15 / 28

17 SVI and Stock Returns: State Level DepVar s,t = α + 5 i=1 β in i P i Ret in mcap,s,t i=1 βi out P i Retmcap,s,t 1 out + γctrls + δ 1M FE + δ 2S FE + ε s,t (1) (2) (3) AAllInv t ANewInv t AOldInv t P 1 Retmcap,t 1 in (-0.49) (-0.52) (-1.19) P 2 Retmcap,t 1 in (-0.52) (1.19) (-0.83) P 3 Retmcap,t 1 in (-0.20) (0.07) (0.58) P 4 Retmcap,t 1 in (-1.04) (0.01) (1.36) P 5 Retmcap,t 1 in (-1.69) (0.35) (0.29) P 1 Retmcap,t 1 out (-0.50) (-0.25) (1.86) P 2 Retmcap,t 1 out (0.03) (-1.30) (2.11) P 3 Retmcap,t 1 out (0.76) (0.21) (0.81) P 4 Retmcap,t 1 out (0.74) (-0.21) (-0.13) P 5 Retmcap,t 1 out (0.36) (-0.50) (-0.20) Controls Yes Yes Yes Q FE Yes Yes Yes S FE Yes Yes Yes Adj R-Squared Observations t-statistics are in parenthesis; standard errors are computed using double clustering. Negativity Bias in Attention Allocation 16 / 28

18 SVI and Stock Returns: State Level DepVar s,t = α + 5 i=1 β in i I i Ret in mcap,s,t 1 + γctrls + δ 1M FE + δ 2S FE + ε s,t (1) (2) (3) AAllInv t ANewInv t AOldInv t I 1 Retmcap,t 1 in (0.80) (0.35) (1.51) I 2 Retmcap,t 1 in (0.55) (-1.08) (1.00) I 4 Retmcap,t 1 in (-1.59) (-0.35) (1.30) I 5 Retmcap,t 1 in (-2.43) (-0.33) (0.11) I 1 Retmcap,t 1 out (-0.88) (-0.51) (-2.92) I 2 Retmcap,t 1 out (-1.36) (1.16) (-2.64) I 4 Retmcap,t 1 out (-0.21) (-0.50) (-0.66) I 5 Retmcap,t 1 out (-0.28) (0.10) (-0.78) Controls Yes Yes Yes Q FE Yes Yes Yes S FE Yes Yes Yes Adj R-Squared Observations t-statistics are in parenthesis; standard errors are computed using double clustering. Negativity Bias in Attention Allocation 17 / 28

19 SVI and Stock Returns: Company Level We test if the negativity bias also exists at the company level using the 100 largest companies in the S&P. Regression specification, 5 ATicker c,t = α + β i P i Ret c,t 1 + γctrls + δ 1 M FE + δ 2 C FE + ε c,t ; ATicker c,t = α + i=1 5 β i I i Ret c,t 1 + γctrls + δ 1 M FE + δ 2 C FE + ε c,t. i=1 ATicker c,t is abnormal SVI for the ticker of company c (Da et al. (2009)); Ret c,t are weekly company returns. Controls are: (i) AVlm, Abnormal Trading Volume (from NYSE); (ii) LMcap, log of the firms market capitalization, and (iii) P i VWRet, quintiles of the value-weighed return of all stocks in CRSP. M FE and C FE are month and company dummies; standard errors are computed using double clustering. Negativity Bias in Attention Allocation 18 / 28

20 SVI and Stock Returns: Company Level 5 ATicker c,t = α + β i P i Ret c,t 1 + γctrls + δ 1M FE + δ 2C FE + ε c,t i=1 t-statistics are in parenthesis; standard errors are computed using double clustering. Negativity Bias in Attention Allocation 19 / 28

21 SVI and Stock Returns: Company Level We test if the negativity bias also exists at the company level using the 100 largest companies in the S&P. Regression specification, 5 ATicker c,t = α + β i P i Ret c,t 1 + γctrls + δ 1 M FE + δ 2 C FE + ε c,t ; ATicker c,t = α + i=1 5 β i I i Ret c,t 1 + γctrls + δ 1 M FE + δ 2 C FE + ε c,t. i=1 ATicker c,t is abnormal SVI for the ticker of company c (Da et al. (2009)); Ret c,t are weekly company returns. Controls are: (i) AVlm, Abnormal Trading Volume (from NYSE); (ii) LMcap, log of the firms market capitalization, and (iii) P i VWRet, quintiles of the value-weighed return of all stocks in CRSP. M FE and C FE are month and company dummies; standard errors are computed using double clustering. Negativity Bias in Attention Allocation 20 / 28

22 SVI and Stock Returns: Company Level 5 ATicker c,t = α + β i P i Ret c,t 1 + γctrls + δ 1M FE + δ 2C FE + ε c,t i=1 t-statistics are in parenthesis; standard errors are computed using double clustering. Negativity Bias in Attention Allocation 21 / 28

23 SVI and Stock Returns: Company Level ATicker c,t = α + 5 β i I i Ret c,t 1 + γctrls + δ 1M FE + δ 2C FE + ε c,t i=1 (1) (2) (t-1) (0.50) (0.43) (t-1) (-0.16) (0.17) (t-1) (0.19) (0.33) (t-1) (0.47) (0.45) (t-1) (0.07) (0.19) AVlm t (2.57) LMcap t (0.02) VWRet t 1 No Yes S FE Yes Yes M FE Yes Yes Adj R-Squared Observations t-statistics are in parenthesis; standard errors are computed using double clustering. Negativity Bias in Attention Allocation 22 / 28

24 Robustness Checks: Distribution of Returns Larger negative coefficient may be driven by skewness and/or asymmetric outliers Density Density Original Return Transformed Return Negativity Bias in Attention Allocation 23 / 28

25 Robustness Checks: Distribution of Returns Redistribute negative returns to replicate the distribution of positive returns. For each week and portfolio (or stock): If the return is positive, keep it without modification. If the return is negative: (a) get the time-series returns for the past 5 years associated with that portfolio (or stock); (b) split the previous time-series into positive and negate returns; (c) find the percentile rank of the current (negative) return within the group of negative returns; (d) find a (positive) return, within the group of positive returns, with percentile rank closest to the one in (c); (e) replace the current (negative) return with the negative of the positive return from (d). Negativity Bias in Attention Allocation 24 / 28

26 Robustness Checks: Distribution of Returns Week Ret ( )Ret ( )Pct (+)Ret (+)Pct Ret T t 260 2% 2% 3 t % 6 t 258 1% 1% 0 t % 6 t % 70% t 2 80% 80% 100 t 1 4% 4% 3 t 8% 5% t + 1 8% 8% t + 2 2% 2% t + 3 9% Negativity Bias in Attention Allocation 25 / 28

27 Robustness Checks: Distribution of Returns SVI and Stock Returns: Company Level ATicker c,t = α + 5 β i P i Retc,t 1 T + γctrls + δ 1M FE + δ 2C FE + ε c,t i=1 Ret T are transformed returns. (1) (2) P1 Ret (t-1) (-1.04) (-0.91) P2 Ret (t-1) (-0.78) (-1.65) P3 Ret (t-1) (-0.80) (-1.12) P4 Ret (t-1) (1.04) (0.47) P5 Ret (t-1) (-0.05) (-2.58) AVlm t (2.90) LMcap t (0.67) VWRet t 1 No Yes S FE Yes Yes M FE Yes Yes Adj R-Squared Observations Negativity Bias in Attention Allocation 26 / 28

28 Conclusion We find that individual investors display a negativity bias in attention allocation with respect to stock returns. At the U.S., state, and company level, lagged negative extreme returns are stronger predictors of retail investors attention than positive extreme returns. By redistributing returns, we rule out that negative returns are stronger simply because negative and positive returns are not symmetric events to stockholders. Negativity Bias in Attention Allocation 27 / 28

29 Robustness Checks: Distribution of Returns SVI and Stock Returns: Company Level 5 ATicker c,t = α + β i I i Retc,t 1 T + γctrls + δ 1M FE + δ 2C FE + ε c,t i=1 Ret T are transformed returns. Negativity Bias in Attention Allocation 28 / 28

30 References I Brad M. Barber and Terrance Odean. All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors. Review of Financial Studies, 21(2): , R F Baumeister, Ellen Bratslavsky, and Kathleen D Vohs. Bad is stronger than good. Review of General Psychology, 5(4): , Colin A. Cameron, Jonah B. Gelbach, and Douglas L. Miller. Robust Inference with Multi-way Clustering. Working Paper 327, National Bureau of Economic Research, September D. Cochrane and G. H. Orcutt. Application of least squares regression to relationships containing auto- correlated error terms. Journal of the American Statistical Association, 44(245): pp , ISSN Negativity Bias in Attention Allocation 29 / 28

31 References II Jennifer Conrad and Gautam Kaul. Time-variation in expected returns. The Journal of Business, 61(4):pp , ISSN Zhi Da, Joseph Engelberg, and Pengjie Gao. In Search of Attention. SSRN elibrary, J B De Long, Andrei Shleifer, L H Summers, and R J Waldmann. Noise trader risk in financial markets. Journal of Political Economy, 98(4): , Simon Gervais, Ron Kaniel, and Dan H. Mingelgrin. The high-volume return premium. The Journal of Finance, 56(3): , ISSN David Hirshleifer, James N. Myers, Linda A. Myers, and Siew Hong Teoh. Do individual investors drive post-earnings announcement drift? direct evidence from personal trades. Finance , EconWPA, December Negativity Bias in Attention Allocation 30 / 28

32 References III Kewei Hou, Lin Peng, and Wei Xiong. A Tale of Two Anomalies: The Implications of Investor Attention for Price and Earnings Momentum. SSRN elibrary, AW Lo and AC MacKinlay. Stock market prices do not follow random walks: evidence from a simple specification test. Review of Financial Studies, 1(1):41 66, Whitney K. Newey and Kenneth D. West. Hypothesis testing with efficient method of moments estimation. International Economic Review, 28(3):pp , ISSN F Pratto and O P John. Automatic vigilance: the attention-grabbing power of negative social information. Journal of Personality and Social Psychology, 61(3): , Hal R. Varian and Hyunyoung Choi. Predicting the Present with Google Trends. SSRN elibrary, Negativity Bias in Attention Allocation 31 / 28

33 Robustness Checks: Modeling Autocorrelation Stock returns are autocorrelated (Conrad and Kaul (1988); Lo and MacKinlay (1988)). Up to this point we have used: Newey and West (1987) in time-series regressions; Cameron et al. (2006) s double clustering in panel regressions. Alternatives: Cochrane and Orcutt (1949) estimation; Autoregressive distributive lag (ADL) model, ASVI t = α + βret t 1 + ε t, Ret t = c + ρret t 1 + ν t, Ret t 1 ASVIt {}}{{}}{ ASVI t ρasvi t 1 = α + β( Ret t 1 ρret t 2) + µ t, µ t = ε t ρε t 1. Negativity Bias in Attention Allocation 32 / 28

34 Robustness Checks: Modeling Autocorrelation SVI and Stock Returns: Company Level ATicker c,t {}}{ ATicker c,t ρaticker c,t = α + 5 {}}{ β i ( P i Ret c,t 1 ρp i Ret c,t 2) i=1 P i Ret c,t 1 +γctrls + δ 1M FE + δ 2C FE + µ c,t Negativity Bias in Attention Allocation 33 / 28

35 SVI and Stock Returns: State Level Available data for AllInv Google returns no data for the states in white. For the rest, the percentage of weeks with available data is in parenthesis; the maximum number of weeks is 354. States with warmer colors have more data. Negativity Bias in Attention Allocation 34 / 28