This paper introduces the market framing bias (MFB): a framing effect that affects the return-risk tradeoff under different frameworks of aggregate market losses and profits, which is measured by the absolute difference between betas in the rising and falling markets. The paper finds that the MFB can predict lower future stock return on the cross-section. Specifically, after controlling for various firm-specific characteristics, this predictive power of the FMB declines over time. Furthermore, the predictive power of the FMB is stable in the short term even after controlling for various pricing factors and firm-specific characteristics.
The framing effect is an anomaly that extensionally equivalent descriptions lead to different choices by altering the relative salience of different aspects of the problem [[
Unlike the previously mentioned papers, our analysis starts with the concept of market framing bias (MFB) which is the bias of the risk-return trade-off between the up and down markets. Indeed, the framing effect in behavioral finance is a cognitive bias where people make decisions based on whether they are in a loss-making or profit-making environment. The up or down markets are the natural positive framing (profit-making environment) or negative framing (loss-making environment) in the stock market. Thus, this bias of the risk-return trade-off between up and down markets is distinctly inconsistent with traditional CAPM [[
Our paper then shifts to exploring whether MFB can predict the future returns of stocks in the cross-section. Based on the fact that investor irrationality can be translated into abnormal profits [[
Our paper contributes three fresh insights into the framing effect measured by MFB. First, unlike the narrow framework [[
The rest of the paper is organized as follows: Section 2 estimates the MFB. Section 3 describes the data and variables. Section 4 offers the predictive power of MFB for cross-sectional stock returns. Section 5 presents additional robustness tests. Section 6 concludes the study.
Current studies typically discuss the framing effect through experiments and surveys. However, the framing effect is not a laboratory curiosity, but a ubiquitous reality [[
First, the up or down market provides a "natural frame" to characterize the investment climate. Kahneman [[
Second, we use the difference between the betas in the up and down markets to measure the behavioral bias of investors. Glascock and Lu-Andrews [[
Finally, we take the absolute value, because the framing effect, as defined by Kahneman [[
Thus, we define the absolute difference between betas in up and down markets as market framing bias (MFB), which measures a framing effect that influences people's decisions depending on whether they are in a loss-making or profit-making environment. Specifically, we denote δ as a binary variable that indicates up or down markets:
Graph
Following Glascock and Lu-Andrews [[
Graph
Where R
Graph
Graph
Graph
Graph
Graph
Graph
The definition of MFB (MFB) in formula (
We conclude by noting that the MFB reflects how the risk-return trade-off varies between the up and down markets, similar to how investor expectations differ in the gain/loss framework. If the investor is rational for a stock i, then according to CAPM theory, the risk-return trade-offs should be the same under gain/loss framing, that is, MFB
The MFB is a cognitive bias, and the simultaneous existence of investors' rationality and cognitive biases thereby making investors adapt to the changing environment [[
We collect the sample data for all A-shares (traded in the Shanghai Stock Exchange and Shenzhen Stock Exchange, excluding SSE STAR Market, as the official opening of the SSE STAR Market is July 22, 2019, which results in too little available data.). Daily and monthly stock market data used in this paper are from the RESSET database, except for the following data that come from China Stock Market & Accounting Research Database (CSMAR): momentum (UMD; Carhart, [[
We aim to analyze the role of MFB in predicting cross-sectional returns. Thus, we control for various firm-specific characteristics that affect expected stock returns. Specifically, the firm-specific characteristics are defined as follows. 1) Beta, following Bali et al. [[
Graph
The descriptive statistics and correlations with relative firm-specific characteristics used in this study are presented in Table 1, where the statistics are computed as time-series averages of the monthly cross-sectional means. Panel A of Table 1 shows that MFB has a mean equal to 1.12, a median equal to 1.00, and a standard deviation equal to 0.49. It appears that many stocks have quite respectable MFB spreads large enough to conclude that investors are always subject to framing effects and potentially allow for differential pricing effects of MFBs.
Graph
Table 1 Descriptive statistics and correlation matrix of firm-specific variables.
Panel A: Descriptive statistics MFB Beta Size BM STR MOM Illiq Coskew BD VOLDU VaR Mean 1.12 1.11 21.75 0.36 0.01 0.18 0.24 -5.04 1.23 0.83 0.07 St Dev 0.49 0.18 0.99 0.13 0.10 0.52 0.28 11.09 0.34 15.65 0.01 Median 1.00 1.08 22.12 0.34 0.01 0.03 0.12 -2.83 1.16 -0.50 0.07 Min 0.29 0.59 19.75 0.17 -0.29 -0.60 0.01 -88.98 0.25 -49.90 0.05 Max 3.86 2.10 23.33 0.70 0.34 2.34 1.45 56.54 2.96 98.71 0.10 Skew 1.83 1.27 -0.46 0.43 0.28 1.75 1.96 -1.91 1.47 2.69 0.31 Kurt 6.20 4.94 -1.18 -0.75 1.08 3.45 3.63 19.04 5.49 14.88 -0.68 10th Per 0.63 0.95 20.29 0.20 -0.09 -0.29 0.04 -14.89 0.95 -10.92 0.06 90th Per 1.72 1.34 22.93 0.55 0.14 0.80 0.61 1.60 1.61 12.22 0.10 Panel B: Correlation matrix MFB Beta Size BM STR MOM Illiq Coskew BD VOLDU VaR MFB 1.00 Beta 0.27 1.00 Size 0.29 0.27 1.00 BM 0.02 0.17 -0.17 1.00 STR 0.10 -0.09 0.12 -0.21 1.00 MOM 0.01 -0.17 0.30 -0.55 0.37 1.00 Illiq -0.24 -0.07 -0.66 0.25 -0.17 -0.24 1.00 Coskew -0.43 -0.31 -0.20 -0.04 -0.01 0.01 0.16 1.00 BD 0.39 0.72 0.25 0.16 -0.04 -0.10 -0.06 -0.73 1.00 VOLDU 0.05 -0.24 0.07 -0.14 0.45 0.37 0.08 0.01 -0.15 1.00 VaR -0.14 0.06 0.50 -0.13 0.10 0.35 -0.26 0.10 0.02 -0.10 1.00
1 Note: Panel A reports the mean, standard deviation, median, minimum, maximum, skewness, kurtosis, 10th percentile, and 90th percentile for each variable. The statistics are calculated as the time-series averages of monthly cross-sectional means. Panel B reports the time-series average of monthly cross-sectional correlations among the variables. The sample period covers January 2000 to December 2019.
Panel B of Table 1 presents the time-series averages of cross-sectional correlations for all firm-specific characteristics, including the MFB. Generally, the MFB has no strong correlations with any of the firm-specific characteristics. Specifically, the correlation between co-skewness and MFB is -0.43, and the correlation between downside beta and MFB is 0.39, indicating that co-skewness and downside betas are weakly related to the MFB. Meanwhile, other firm-specific characteristics exhibit a weak correlation with the MFB, as indicated by the absolute values of all those correlation coefficients below 0.3. In addition, Panel B shows significant and negative correlations between size and illiquidity, co-skewness, and downside beta.
Since most asset pricing factors are derived from the monthly returns of a portfolio, they are likely to exhibit contemporaneous correlations. Table 2 presents the correlation matrix for monthly asset pricing factors (The asset pricing factors used in the paper include: FF5 factors (the market (MKT), size (SMB), value (HML), investment (CMA) and profitability (RMW) factors of Fama and French, [[
Graph
Table 2 Stationary tests and correlation matrix of pricing factors.
MKT SMB HML RMW CMA UMD SENT PEAD FIN SMBQ IAQ ROEQ MKT 1.00 SMB 0.08 1.00 HML -0.14 -0.50 1.00 RMW -0.28 -0.78 0.27 1.00 CMA 0.14 0.42 0.14 -0.67 1.00 UMD -0.07 -0.23 -0.06 0.38 -0.32 1.00 SENT -0.03 0.05 -0.11 -0.04 0.07 0.03 1.00 PEAD -0.20 -0.06 0.04 0.11 -0.02 -0.04 -0.03 1.00 FIN -0.50 -0.24 0.16 0.31 -0.18 -0.02 0.04 0.26 1.00 SMBQ 0.04 0.92 -0.43 -0.66 0.35 -0.20 0.05 -0.07 -0.17 1.00 IAQ -0.04 0.02 0.33 -0.22 0.46 -0.23 -0.06 -0.03 0.09 0.04 1.00 ROEQ 0.20 0.63 -0.12 -0.80 0.64 -0.52 0.04 -0.01 -0.20 0.48 0.16 1.00
2 Note: This table presents the correlation matrix for the monthly factors. The sample period is from January 2000 to December 2019.
In this subsection, we use univariate portfolio-level analysis based on FF5 factors [[
A univariate and multi-term portfolio-level analysis is performed where deciles are formed every month by ascending sorting stocks based on their MFB values. Excess returns and abnormal returns (α
Table 3 presents the time-series averages of excess returns and alphas calculated after adjusting by the FF5 factors (MKT, SMB, HML, BMW, and CMA; Fama and French, [[
Graph
Table 3 Univariate and multi-term portfolio analysis.
Panel A: Excess returns ( P1 0.76 0.53 0.56 0.58 0.53 0.40 0.44 0.54 0.54 0.45 0.45 0.53 (1) (0.74) (0.75) (0.79) (0.7) (0.54) (0.58) (0.71) (0.71) (0.61) (0.59) (0.67) P10 -0.14 0.26 0.32 0.35 0.34 0.37 0.42 0.43 0.33 0.40 0.30 0.34 (-0.19) (0.34) (0.43) (0.46) (0.45) (0.47) (0.52) (0.54) (0.41) (0.49) (0.38) (0.43) H-L -1.17 -0.53 -0.50 -0.49 -0.45 -0.29 -0.28 -0.37 -0.47 -0.32 -0.41 -0.45 (-8.95) (-3.86) (-4.14) (-3.79) (-3.46) (-2.10) (-2.50) (-3.04) (-3.66) (-2.32) (-3.62) (-3.14) Panel B: Abnormal returns ( P1 -0.39 -0.46 -0.43 -0.39 -0.45 -0.56 -0.50 -0.31 -0.39 -0.42 -0.48 -0.37 (-2.96) (-3.62) (-3.13) (-2.88) (-3.88) (-3.96) (-3.85) (-2.20) (-3.02) (-3.81) (-3.89) (-2.73) P10 -1.18 -0.77 -0.70 -0.66 -0.66 -0.66 -0.54 -0.61 -0.58 -0.57 -0.67 -0.58 (-7.52) (-4.93) (-5.11) (-4.41) (-4.70) (-4.57) (-3.69) (-4.18) (-4.30) (-4.11) (-4.4) (-4.17) H-L -1.05 -0.57 -0.54 -0.53 -0.48 -0.37 -0.30 -0.57 -0.46 -0.42 -0.46 -0.48 (-8.49) (-4.25) (-4.69) (-4.74) (-4.59) (-3.08) (-2.73) (-4.23) (-4.01) (-3.95) (-4.66) (-3.95)
- 3 Note: Panel A reports value-weighted excess returns (in percentage) from 1 to 12 months ahead after portfolio formation. Panel B presents the FF5-adjusted alphas for each decile from 1 to 12 months after portfolio formation. Newey and West [[
33 ]] t-statistics are given in parentheses. - 4
‡ : Significant at 1%. - 5
† : Significant at 5%. - 6 *: Significant at 10%.
Panel B of Table 3 examines whether the FF5 factor can explain the difference in excess returns between the extreme MFB deciles. In general, the abnormal returns (alphas) of the lowest MFB deciles are always greater than the alphas of the highest MFB deciles. For example, 1 month after portfolio formation, the alpha of the lowest and highest MFB deciles are -0.39% and -1.18%, respectively. And the alpha of the portfolio H-L is -1.05% with a significant Newey and West [[
The poor explanation of the FF5 factors for the abnormal returns produced by the MFB presented in Table 3 is observed, possibly for two reasons. One reason is that a firm-specific characteristic correlated with MFB but not captured by the FF5 factors has a significant impact on expected stock returns. Another reason is that the market may be influenced by rational and irrational forces [[
Bivariate two-stage 10×10 dependent sorts are used to do bivariate portfolio analysis. First, we sort stocks into decile portfolios monthly based on various firm-specific characteristics (Beta, Size, BM, STR, MoM, Illiq, Coskew, BD, VOLDU, and VaR). Second, we sort stocks into additional deciles based on the MFB within each firm-specific characteristic decile that is sorted in step one. Then, 100 conditionally double-sorted groups are provided to construct ten portfolios through the above two steps. Portfolio 1 is the combined group of stocks with the lowest MFB in each firm-specific characteristic decile, whereas portfolio 10 is the one with the highest MFB.
Table 4 presents the results of the bivariate portfolio analysis (for brevity, only the results of 1, 3, 6, and 12 months after portfolio formation are reported). Panel A shows the excess returns of the portfolio H-L by controlling for various firm-special characteristics. It is easy to find that all excess returns are significantly smaller than zero, except when controlled by BM (τ = 12), MoM (τ = 12) and VOLDU (τ = 6, 12), respectively. These results simply show that the role of MFB on future stock returns may be consistent with BM, MoM, and VOLDU in the long run, but BMF is not replaced by these firm-special characteristics because the short-term role cannot be replaced by other firm characteristics. Furthermore, we can find that, after controlling for these firm-special characteristics, MFB's fine predictive power diminishes over time. For example, when Beta is the first-stage sorting variable, the excess return of the portfolio H-L changes from -0.89% at τ = 1 to -0.15% at τ = 12.
Graph
Table 4 Bivariate portfolio analysis.
Beta Size BM STR MoM Illiq Coskew BD VOLDU VaR Panel A: Return of H-L -0.89 -0.86 -0.76 -0.88 -0.89 -0.89 -0.50 -0.76 -0.49 -0.76 (-6.92) (-6.47) (-6.28) (-7.22) (-7.13) (-6.82) (-4.98) (-6.76) (-3.53) (-5.98) -0.41 -0.46 -0.31 -0.43 -0.39 -0.40 -0.29 -0.31 -0.20 -0.36 (-4.06) (-4.61) (-3.07) (-4.32) (-4.04) (-3.62) (-4.34) (-4.09) (-1.7) (-4.47) -0.23 -0.34 -0.14 -0.23 -0.21 -0.27 -0.21 -0.16 -0.12 -0.19 (-2.78) (-4.74) (-1.68) (-2.42) (-2.7) (-2.95) (-4.71) (-2.61) (-1.24) (-2.85) -0.15 -0.22 -0.05 -0.16 -0.10 -0.18 -0.12 -0.11 -0.06 -0.13 (-2.36) (-4.09) (-0.78) (-2.13) (-1.41) (-2.64) (-3.21) (-2.48) (-0.82) (-2.34) Panel B: Abnormal return adjusted by FF5 -0.77 -0.68 -0.63 -0.77 -0.79 -0.74 -0.47 -0.62 -0.37 -0.68 (-6.99) (-5.73) (-5.62) (-6.79) (-7.45) (-6.38) (-4.37) (-5.97) (-3.42) (-5.29) -0.51 -0.55 -0.41 -0.53 -0.48 -0.49 -0.34 -0.39 -0.31 -0.45 (-6.00) (-6.80) (-4.93) (-6.32) (-5.75) (-5.42) (-4.30) (-5.39) (-3.30) (-5.38) -0.29 -0.37 -0.21 -0.29 -0.27 -0.32 -0.21 -0.18 -0.16 -0.23 (-4.78) (-7.03) (-3.94) (-4.38) (-4.62) (-4.66) (-4.53) (-3.51) (-2.25) (-4.70) -0.18 -0.23 -0.08 -0.18 -0.11 -0.20 -0.12 -0.12 -0.08 -0.13 (-2.98) (-4.63) (-1.28) (-2.71) (-1.81) (-3.29) (-2.89) (-2.86) (-1.22) (-2.62) Panel C: Abnormal return adjusted by FF5+SL2+UMD+SENT -0.93 -0.86 -0.76 -0.93 -0.91 -0.92 -0.54 -0.75 -0.47 -0.86 (-7.85) (-6.69) (-6.59) (-7.28) (-8.12) (-7.15) (-4.46) (-6.30) (-3.93) (-6.40) -0.58 -0.61 -0.45 -0.60 -0.53 -0.56 -0.39 -0.41 -0.35 -0.48 (-5.48) (-6.07) (-4.28) (-5.60) (-5.36) (-4.81) (-4.35) (-4.85) (-3.05) (-5.17) -0.28 -0.36 -0.21 -0.29 -0.29 -0.31 -0.19 -0.14 -0.16 -0.22 (-4.14) (-5.98) (-3.15) (-3.59) (-4.31) (-4.04) (-3.42) (-2.41) (-1.90) (-3.86) -0.17 -0.22 -0.07 -0.17 -0.11 -0.21 -0.10 -0.10 -0.07 -0.11 (-2.57) (-3.98) (-1.04) (-2.34) (-1.60) (-3.13) (-2.32) (-2.21) (-1.02) (-2.03)
- 7 Note: Reported are the results of value-weighted bivariate portfolio analysis for future 1, 3, 6, and 12 months (τ = 1, τ = 3, τ = 6, and τ = 12). Only the results for the portfolio H-L are reported in the table. Newey and West [[
33 ]] t-statistics are given in parentheses. - 8
‡ : Significant at 1%. - 9
† : Significant at 5%. - 10 *: Significant at 10%.
Panel B shows the abnormal returns, adjusted by FF5 factors, for the portfolio H-L grouped by the MFB at the second-stage. First, all abnormal returns adjusted by FF5 factors are significantly negative in the short-term (τ = 1, 3), which implies that FF5 factors cannot explain the excess returns produced by the bivariate sorts in the short-term. Second, consistent with panel A, we also find that the ability of MFB to generate abnormal returns still fades over time after controlling for these firm-special characteristics. For example, when Size is the first-stage sorting variable, the abnormal returns are -0.86%, -0.46%, -0.34%, and -0.22% at t+1, t+3, t+6, and t+12, respectively.
Concerning the possibility that the FF5 factors are not sufficient to explain the excess returns generated by the FMB, in Panel C we include more factors to explore this issue. Panel C in Table 4 presents the results adjusted by more asset pricing factors, which include FF5 factors (MKT, SMB, HML, RMW, and CMA; Fama and French, [[
In summary, by the bivariate portfolio analysis, we show that a higher FMB still has a lower future return and vice versa, especially in the short term. Moreover, this negative predictive power of the FMB weakens over time, but cannot fade by controlling for various firm-specific characteristics and asset pricing factors.
In this subsection, firm-level Fama-MacBeth regressions (Fama-MacBeth, [[
In the first stage, monthly cross-sectional regressions of excess stock returns (R
Graph
The univariate regression or the full regression specification of the model (
In the second stage, the cross-sectional regression coefficients are estimated by the time-series averages where Newey and West [[
Graph
Table 5 Firm-level Fama-MacBeth regression.
(1) (2) (1) (2) (1) (2) (1) (2) MFB -2.23 -2.22 -2.21 -1.90 -1.91 -1.65 -1.58 -1.04 (-6.81) (-5.59) (-5.59) (-4.87) (-4.12) (-4.01) (-3.17) (-2.67) Beta -3.84 -6.33 -9.33 -12.67 (-3.78) (-3.58) (-4.03) (-4.53) Size 1.40 1.73 1.60 1.75 (2.28) (1.78) (1.37) (1.34) BM -0.67 -0.22 0.41 0.66 (-1.43) (-0.51) (0.98) (1.61) STR 1.90 1.87 0.66 -1.34 (2.77) (1.88) (0.56) (-1.17) MoM 4.83 9.91 9.38 10.45 (1.60) (2.00) (1.95) (2.05) Illiq 0.44 1.53 1.26 0.77 (0.51) (1.72) (1.23) (1.02) Coskew 0.18 1.46 1.37 0.86 (0.22) (1.64) (1.20) (1.17) Betadown -3.37 -3.22 -2.86 -2.17 (-8.04) (-6.32) (-5.06) (-3.62) Voldu -2.86 -3.31 -3.49 -3.01 (-7.23) (-5.32) (-4.30) (-3.30) VaR 3.24 1.16 -0.57 -1.55 (4.78) (1.57) (-0.68) (-1.69) Adj. R2 0.47 9.19 0.30 8.61 0.22 8.49 0.18 9.45
- 11 Note: This table reports the results of the Fama-MacBeth cross-sectional regressions of individual firms for future 1, 3, 6, and 12 months (τ = 1, τ = 3, τ = 6, and τ = 12) on the control variables measured in month t. The column labeled "(
1 )" or "(2 )" presents the average coefficient and the adjusted R2 (in percentage) of the univariate regression or the full regression specification, respectively. Newey and West [[33 ]] t-statistics are given in parentheses. - 12
‡ : Significant at 1%. - 13
†: Significant at 5%. - 14 *: Significant at 10%.
In the short-term τ = 1 case, the average slope coefficient from the univariate regression in column (
We also find that, from the full regression, the variables that have a stable effect on future returns, both in the short and long term, are Beta, BD, and Voldu. And in the short-term (τ = 1), different from the results of Atilgan et. al [[
The in-sample tests in the above discussion may present a look-ahead bias. Following Goyal and Welch [[
Our test data are the value-weighted MFB-sorted quintile portfolios. At first, in the 100 months rolling (expanding) window, the t + τ month's return of a portfolio p (R
Graph
where Z
Graph
where MFB
Graph
Graph
where R
Graph
Graph
Graph
Graph
Graph
Graph
Graph
where u
Graph
Table 6 Out-of-sample predictability.
Rolling approach Expanding approach MFB1 MFB2 MFB3 MFB4 MFB5 MFB1 MFB2 MFB3 MFB4 MFB5 Panel A: 1-month ahead ( FF5 6.12 19.32 16.71 13.62 29.05 3.99 15.26 10.90 7.52 27.08 (4.27) (6.50) (4.99) (3.57) (6.94) (4.46) (6.27) (4.79) (2.57) (7.13) Q4 8.09 15.54 12.14 5.73 17.13 7.01 14.58 11.73 6.33 17.13 (2.99) (3.53) (3.21) (2.40) (4.31) (3.30) (3.87) (3.47) (2.22) (4.82) All -7.34 10.05 9.74 9.36 23.37 -8.43 5.37 1.98 3.08 21.38 (1.66) (4.74) (4.06) (3.64) (4.64) (1.94) (5.09) (4.07) (2.31) (5.67) Panel B: 3-month ahead ( FF5 5.57 18.86 16.12 13.22 28.13 3.47 15.29 10.88 7.48 26.35 (4.15) (6.41) (4.89) (3.46) (6.70) (4.32) (6.19) (4.70) (2.51) (6.87) Q4 7.93 15.84 12.39 6.23 15.83 6.86 14.93 12.11 6.63 15.93 (2.94) (3.50) (3.18) (2.40) (4.10) (3.22) (3.83) (3.44) (2.23) (4.58) All -8.12 9.66 9.16 9.14 22.46 -9.40 5.08 1.61 2.93 20.62 (1.59) (4.65) (3.97) (3.55) (4.46) (1.81) (4.99) (3.94) (2.24) (5.43) Panel C: 6-month ahead ( FF5 4.12 16.74 14.43 12.17 27.16 2.79 14.21 10.54 7.85 25.67 (3.95) (6.13) (4.67) (3.28) (6.56) (4.17) (6.00) (4.60) (2.52) (6.72) Q4 7.36 14.89 11.80 5.76 15.77 6.20 14.00 11.56 6.34 16.08 (2.84) (3.37) (3.09) (2.29) (4.06) (3.08) (3.67) (3.33) (2.15) (4.56) All -9.80 7.35 7.59 8.29 21.81 -10.46 3.76 1.17 3.17 20.04 (1.47) (4.42) (3.79) (3.38) (4.34) (1.70) (4.82) (3.85) (2.24) (5.31) Panel D: 12-month ahead ( FF5 1.61 12.67 10.58 10.18 24.79 1.37 12.29 8.23 8.12 24.35 (3.58) (5.50) (4.12) (2.83) (6.16) (3.88) (5.56) (4.20) (2.40) (6.39) Q4 4.57 10.8 7.99 3.55 14.41 4.71 11.99 9.41 5.26 16.29 (2.46) (2.96) (2.65) (1.82) (3.87) (2.75) (3.29) (2.94) (1.84) (4.44) All -15.27 -1.37 0.26 6.22 16.69 -12.11 1.46 -1.22 3.61 18.29 (1.11) (4.10) (3.30) (2.82) (4.66) (1.49) (4.37) (3.44) (2.19) (5.27)
- 15 Note: We test the forecasting power of the MFB for the τ-month-ahead excess return of the value-weighted MFB-sorted quintile portfolios (τ = 1, τ = 3, τ = 6, and τ = 12). The benchmark models are FF5, Q4, and the factors model, respectively. Panel A, B, C, and D show the 1, 3, 6, and 12-month ahead forecasting ability of MFB, respectively.
- 16 *p <.1;
- 17
† p <.05; - 18
‡ p <.01 (the significance levels for two-sided tests are indicated).
Panel A of Table 6 shows the 1-month ahead forecasting ability of MFB. When a 100-month rolling window is used,
Graph
Graph
Graph
Panel B and Panel C of Table 6 also have similar patterns to Panel A, and the
Graph
Graph
Graph
Taken together, the results of the out-of-sample tests show that our predictive model with MFB outperforms the benchmark pricing model in predicting equity premiums, and in agreement with the in-sample analysis, the predictive power of MFB for future returns weakens as the prediction time increases.
The utilization of publicly traded data to define and quantify framework deviations for the first time is expected to generate considerable debate. Firstly, as we are aware, stocks tend to exhibit stronger co-movement during market downturns [[
In this section, we divide the full sample period into two based on several macroeconomic indicators that signal whether the economy is experiencing recessions or booms. To measure the state of China's economy, we choose the following three macroeconomic indicators from the National Bureau of Statistics: Purchasing manager's index (PMI), an important evaluation index of economic activities and a barometer of economic changes; Economic sentiment leading index (ESLI), which is composed of a set of leading indicators that lead the consensus index and is used to predict the future trend of the economy; Economic sentiment consistent index (ESCI), which reflects the basic trend of the current economy.
Our sample from January 2005 to December 2019 (Since China's PMI index started in 2005) is split into two subsamples by the three indicators, and univariate analysis is made for them. Specifically, after splitting the sample one at a time, quintiles are formed every month by sorting stocks based on ascending MFB, and the monthly value-weighted excess returns (ER) are calculated for each quintile. The alphas (α) to all quintiles, including zero-investment portfolio (H-L) which is long (short) in equities with high (low) framing effect, are obtained by regression from FF5+IVA
Graph
Table 7 Subsample analysis.
Panel A: Subsample divided by PMI Booms: PMI>50 Recessions: PMI< = 50 Port 1 Port 2 Port 3 Port 4 Port 5 H-L Port 1 Port 2 Port 3 Port 4 Port 5 H-L ER 1.32 1.25 1.51 1.05 0.70 -0.90 -0.90 -1.75 -1.20 -1.40 -1.50 -0.87 (1.50) (1.46) (1.43) (1.17) (0.78) (-3.16) (-0.89) (-1.36) (-0.96) (-1.17) (-1.07) (-1.61) -0.19 -0.08 -0.10 -0.49 -1.10 -1.20 -0.05 -0.57 -0.08 -0.88 -1.15 -1.36 (-1.02) (-0.47) (-0.45) (-2.99) (-4.58) (-4.73) (-0.17) (-2.09) (-0.27) (-3.49) (-1.47) (-1.64) Panel B: Subsample divided by ESLI Booms: ESLI>100 Recessions: ESLI< = 100 Port 1 Port 2 Port 3 Port 4 Port 5 H-L Port 1 Port 2 Port 3 Port 4 Port 5 H-L ER 0.87 0.70 1.04 0.42 -0.07 -1.21 1.02 0.72 0.98 1.03 1.15 -0.19 (0.92) (0.76) (0.87) (0.44) (-0.07) (-3.84) (0.87) (0.62) (0.80) (0.77) (0.91) (-0.41) -0.14 -0.15 -0.07 -0.60 -1.05 -1.19 0.05 -0.41 -0.12 -0.24 -0.91 -1.27 (-0.80) (-0.90) (-0.28) (-3.50) (-3.94) (-4.20) (0.13) (-2.30) (-0.52) (-0.94) (-2.02) (-2.42) Panel C: Subsample divided by ESCI Booms: ESCI>100 Recessions: ESCI< = 100 Port 1 Port 2 Port 3 Port 4 Port 5 H-L Port 1 Port 2 Port 3 Port 4 Port 5 H-L ER 0.50 0.24 0.76 -0.02 -0.58 -1.36 1.36 1.20 1.30 1.28 1.26 -0.40 (0.42) (0.20) (0.48) (-0.02) (-0.47) (-4.01) (1.49) (1.40) (1.44) (1.35) (1.33) (-1.28) -0.25 -0.13 0.12 -0.48 -1.33 -1.37 -0.04 -0.28 -0.34 -0.56 -0.85 -1.09 (-1.27) (-0.67) (0.43) (-2.45) (-4.62) (-4.58) (-0.20) (-1.8) (-1.99) (-2.57) (-3.17) (-3.11)
- 19 Note: Newey and West [[
33 ]] t-statistics are given in parentheses. - 20
‡ : Significant at 1%. - 21
† : Significant at 5%. - 22 *: Significant at 10%.
Panel A presents the results for PMI which is used to split the sample. When PMI>50, macroeconomic conditions are booming, the left columns of Panel A show that the excess return decreased from 0.5 for the lowest MFB quintile to -0.58 for the highest MFB quintile. And the excess return of the portfolio H-L is -0.90 with a t-statistic of -3.16. However, when PMI< = 50, macroeconomic conditions are recessions, the right columns of Panel A show that the excess return of zero-investment portfolio is -0.87% with an insignificant t-statistic of -1.61. The results imply that the predictive ability of MFB in the booms subsample (PMI>50) are significant, but not during market downturns (PMI< = 50).
This conclusion is confirmed by Panel B and Panel C in Table 7. For example, when the ESCI< = 100 and the market declines, the excess return of the portfolio H-L is -0.40 with a t-statistic of -1.28 (At this time, although alpha is -1.09 with a t-statistic of -3.11, it only shows that FF5, Q4, SL2, momentum, and investor sentiment factors have weak explanatory ability in China.). In conclusion, the sub-sample results in Table 7 show that, markets co-move strongly during the periods of market downturns (Das et al. 2018), and MFB's forecasting ability fades.
The MFB that we define from an investor psychology perspective differs from the difference in downside and upside beta (Delta Beta) defined by Ang et al. [[
Firstly, in Table 8, we present a simple descriptive analysis of the delta beta statistic. The mean value is 0.23 and the variance is 0.39, with a correlation coefficient of 0.36 to MFB. Results show that most of the monthly downside betas are larger than their upside counterparts. However, the correlation coefficient is less than 0.5, suggesting that there is still a pronounced deviation between MFB and delta beta.
Graph
Table 8 Additional statistical description.
MFB Delta Beta MFB Zero MFB Mu Mean STD MFB 1.00 0.36 >0.99 0.96 1.12 0.49 Delta Beta 0.36 1.00 0.36 0.35 0.23 0.39 MFB Zero >0.99 0.36 1.00 0.96 1.12 0.49 MFB Mu 0.96 0.35 0.96 1.00 1.10 0.47
Furthermore, we conducted firm-level Fama-MacBeth regressions [[
Graph
Table 9 Additional statistical description.
OLS WLS Delta Beta -0.42 -0.14 (-0.26) (-1.21) MFB Zero -2.22 -0.22 (-5.56) (-6.74) MFB Mu -2.30 -0.24 (-5.86) (-6.45) Beta -3.84 -3.84 -3.84 0.36 0.37 0.36 (-3.78) (-3.78) (-3.78) (3.89) (4.27) (4.28) Size 1.54 1.41 1.39 -0.45 -0.45 -0.45 (2.49) (2.28) (2.25) (-3.65) (-3.61) (-3.62) BM -0.7 -0.67 -0.66 0.1 0.09 0.09 (-1.48) (-1.43) (-1.41) (1.39) (1.22) (1.18) Str 1.83 1.90 1.90 -0.07 -0.06 -0.06 (2.69) (2.77) (2.78) (-1.35) (-1.3) (-1.27) MoM 4.50 4.83 4.80 0.16 0.17 0.17 (1.59) (1.60) (1.60) (2.47) (2.57) (2.55) Illiq 0.52 0.44 0.37 0.24 0.24 0.24 (0.38) (0.51) (0.42) (3.1) (3.17) (3.16) Coskew 0.79 0.17 0.18 -0.03 0.07 0.06 (0.75) (0.22) (0.23) (-0.19) (0.68) (0.62) Betadown -3.61 -3.37 -3.35 0.11 0.04 0.04 (-8.81) (-8.04) (-8.04) (0.93) (0.37) (0.47) Voldu -3.12 -2.86 -2.86 -0.35 -0.33 -0.32 (-8.00) (-7.23) (-7.18) (-7.12) (-6.51) (-6.51) VaR 3.12 3.24 3.19 -0.34 -0.31 -0.31 (4.01) (4.78) (4.78) (-6.38) (-5.89) (-5.82) Adj. R2 0.09 0.09 0.09 0.09 0.09 0.09
- 23 Note: This table reports the results of the Fama-MacBeth cross-sectional regressions of individual firms for future 1 month on the control variables measured in month t. The columns on the left show the results of ordinary least squares. The columns on the right present results estimated for one-month-ahead returns using the weighted least squares (WLS) methodology of Asparouhova et al. [[
41 ]], where each observed return is weighted by one plus the observed prior return on the stock. Newey and West [[33 ]] t-statistics are given in parentheses. - 24
‡ : Significant at 1%. - 25
† : Significant at 5%. - 26 *: Significant at 10%.
According to Junior et al. [[
In fact, due to the small size of the riskless rate on a daily basis, MFB and MFB Zero should be highly correlated. As confirmed by the results in Table 8 from the previous section, the correlation coefficient between MFB and MFB Zero is greater than 0.99, indicating that the choice of reference point between riskless rate and zero rate of return is almost identical. Additionally, the results in Table 8 also show a high correlation coefficient of 0.96 between MFB and MFB Mu, indicating that using average market returns as the reference point has a strong correlation with using riskless rate as the reference point. Our results are almost identical to Ang et al.'s conclusions: These betas calculated by different reference points exhibit a correlation greater than 0.96.
Finally, we conducted firm-level Fama-MacBeth regressions based on ordinary least squares (OLS) and weighted least squares (WLS, Asparouhova et al. [[
The results in Table 9 show that regardless of whether the zero rate of return or average market returns are used as the reference point, and whether OLS or WLS is used, MFB still has a significant predictive power for stock returns in the next month. For instance, when using the zero rate of return as the reference point, the Fama-MacBeth regression coefficient of MFB Zero based on OLS is -2.22, with a t-statistic of -5.56, while that based on WLS is -0.22, with a t-statistic of -6.74. Moreover, we also found an interesting phenomenon: the sign of the Fama-MacBeth regression coefficients of Beta and Betadown based on OLS or WLS differs, while our MFB has a consistent sign. This further suggests that our MFB is more stable than Beta and Betadown in predicting stock returns when the prior return on the stock is taken into account. In summary, the predictive power of MFB based on different reference points is robust.
In this paper, we propose that investors may exhibit a framing effect in the return-risk tradeoff under different frameworks of aggregate market losses and profits, which is defined as the market framing bias (MFB). We measure the MFB of an individual stock using the absolute difference between the betas in the up and down markets, and we also explore the predictive power of the MFB on future stock returns in the cross-section. More specifically, we find the following:
First and foremost, the paper finds that MFB is able to predict future stock returns in the cross-section. By going long a portfolio of the stocks with the lowest MFB and short selling a portfolio of the stocks with the highest MFB, we find that there is a significant negative τ-month-ahead excess return, where τ ranges from 1 month to 12 months. In other words, the MFB is able to predict the excess return from 1 month to 12 months. Moreover, we show that, after controlling for various firm-specific characteristics, the predictive power of MFB diminishes as future returns are predicted from 1 month to 12 months. Last but not least, various asset pricing factors (FF5, SL3, UMD, and SENT) do not eliminate the MFB's predictive power. We confirm our conclusions by replacing the seminal FF5 factor with the popular Q4 asset pricing factor and performing a bivariate portfolio analysis, as well as by using out-of-sample tests.
Our findings exhibit robustness across the following dimensions: Subsample analysis conducted on boom and recession periods, comparison of Delta Beta's predictive ability from a risk perspective, and varying reference points utilized in MFB construction. Ongoing changes in the stock market structure also present new challenges for researchers. For example, we may construct an asset pricing factor based on the MFB and study if this factor exhibits explanatory power on the cross-section of stock returns.
S1 Data.
(ZIP)
By Jun Xie; Baohua Zhang and Bin Gao
Reported by Author; Author; Author