The purpose of this paper is to answer a fundamental question – are individual stock picks by a particular internet investment community informative enough to beat the market? The author observes that the stock picks by the CAPS community are reflective of existing information and portfolios based upon CAPS community stock rankings do not generate abnormal returns. The CAPS community is good at tracking existing performance but, it lacks predictive ability.
The study uses a unique data set of stock ratings from Motley Fools CAPS community to determine the information content embedded in these ratings. Observing predictive ability of this web-based stock ratings forum will raise questions about the efficiency of the financial markets. The author forms stock portfolios based on stocks’ star ratings, and star rating changes, and test if the long-short portfolio strategy generates significant α after controlling for single, and multi-factor asset pricing models, such as Fama-French three-factor model and Carhart four-factor model.
The paper finds no evidence that the CAPS community ratings contain “information content,” which can be exploited to generate abnormal returns. CAPS community ratings are good at tracking existing stock performance, but cannot be used to make superior forecasts to generate abnormal returns. The findings are consistent with the efficient market hypothesis. Furthermore, the author provides evidence that CAPS community ratings are themselves determined by stock performance rather than the other way around.
The study employs a unique data set capturing the stock ratings of a very popular web-based investment community to evaluate its ability to make better than random forecasts. Besides applying well-accepted asset pricing models to generate α, the study conducts causality tests to discern a causal relation between stock ratings and stock performance.
Mahajan, A. (2018), "Information content of web-based stock ratings: the case of Motley fool CAPS data", Journal of Advances in Management Research, Vol. 15 No. 3, pp. 393-410. https://doi.org/10.1108/JAMR-02-2018-0025Download as .RIS
Emerald Publishing Limited
Copyright © 2018, Emerald Publishing Limited
Using data from Motley Fool CAPS community’s (www.caps.fool.com) stock picks records, Avery et al. (2011) document that the individual predictions of stock performance on the Motley Fool CAPS website are informative. They find that shorting stocks with high negative picks, and buying stocks with high positive picks, generate an annual (raw) return of nine percent over the sample period from November 1, 2006 to December 31, 2008. They also find that the rating based on individual negative picks drives the results. Using Fama-French 3-factor model with momentum factor, they conclude that these results are largely due to stock-picking rather than style factors.
However, contrary to Avery et al. (2011), we find little evidence of informative-ness in Motley Fool CAPS data. Our findings suggest that CAPS star ratings follow same-day stock performance and are not a source of new information content. Because of data availability, our tests rely on the star ratings given by the CAPS community. We do find a monotonically increasing trend for raw returns – stocks with 1(5)-star rating have the lowest (highest) return. However, after controlling for market risk and style factors, a strategy of shorting 1-star stocks and buying 5-star stocks generates a statistically insignificant Jensen’s α, i.e. we find no evidence of abnormal returns, regardless of daily, or monthly, portfolio sorts. For daily sorted portfolios, a sort by the star rating on the same day, instead of previous trading day, produces a monotonically increasing Jensen’s α from 1-star to 5-star portfolios, and a statistically significant positive Jensen’s α for the long-short strategy. We interpret these results as the CAPS community doing well on tracking current day stock performance, but not at predicting the future stock price. In support of this implication, we form portfolios by star rating changes, rather than level values and find results supporting our hypothesis.
This paper is related to at least three strands of literature. First, it relates to how the media interact with the stock market. Empirical studies have found that media outlets can have a temporary influence on prices. For instance, Tetlock (2007) uses the daily content in the Wall Street Journal’s “Abreast of the Market” column to form a pessimism factor and finds that high media pessimism predicts downward pressure on market prices, but is followed by a reversion to fundamentals. He also finds that unusually high or low pessimism predicts high market trading volume. Similarly, Barber and Odean (2001) find that the spread of information on the internet may exacerbate behavioral biases, which lead to suboptimal investments. Other studies within this strand of literature recognize how message boards, or other internet/user-based media, can serve as a source of information about future asset prices. Antweiler and Frank (2004) identify that a high volume of activity on message boards tends to predict higher future trading volumes and price volatility. Using Google search trend data, Da et al. (2011) document that an increase in SVI (an index measuring attention of retail investors) predicts higher stock prices in the next two weeks, and an eventual price reversal within the year. This line of research, which focuses on individual measures of investor reactions (to media, message boards, etc.) or on individual investor signals (Google search trends), parallels more broadly with market-wide measures of sentiment. A notable example of this is Baker and Wurgler’s (2006) aggregate investor sentiment index (BW index), which is constructed from various sentiment proxies such as the closed-end fund discount, and new-equity issuance share valuation, among others. While Sibley et al. (2016) find that the BW index is largely explained by macroeconomic considerations, Huang et al. (2015) find that an econometrically improved version of the index is not adequately explained. This paper finds that CAPS sentiment, as measured by daily cross-sectional dispersion in rating changes, tends to be correlated with high and low periods of Baker-Wurgler sentiment, suggesting that CAPS may serve as a relatively high-frequency measure of investor disposition. Given that CAPS sentiment also exhibits a strong correlation to Huang et al. (2015), our findings suggest that there may be a non-risk based explanation for rating changes as well.
Second, this paper relates to the literature that asks the question – How do individual investors perform? Barber and Odean (2000) demonstrate that the stock choices of individual investors underperform market indices; while Barber et al. (2009) document a short-lived better performance of stocks purchased heavily by individual investors than other stocks. Thus, it will be interesting to see how aggregated individual opinions predict stock performance. Moreover, do some investors have superior information over others? Coval et al. (2005) show a strong persistence in the performance of individual investors’ trades. The question is, do investors with superior information share their opinions on message boards or a community like CAPS, or do they use the information for their own trading?
Finally, this study relates to the literature that asks whether some investors outperform others. Many studies, like Chevalier and Ellison (1999), Metrick (1999), Mikhail et al. (2004), etc., document that a small percentage of investment professionals consistently “beat the market.” Since the players in the CAPS community are not necessarily professionals, will we see them “beating” the market?
The remainder of this paper is arranged as follows – Section 2 describes the data used in the analysis. Section 3 describes the empirical methods and the results. Section 4 contains robustness tests. Section 5 concludes the paper.
The Motley Fool is a multimedia financial-services company. In September 2006, the company unveiled Motley Fool CAPS, a service that monitors and ranks the most successful stock pickers amongst its members. The players in the CAPS community predict whether a stock would underperform, or outperform, S&P 500 by giving a thumbs-down or thumbs-up, respectively. The players also specify the time horizon for which their predictions are valid. CAPS tracks these predictions, and related stock activities, to rate the stocks and the players. Based on the accuracy of their predictions, players receive a percentile score from 1 to 100. Players with a score of 80 or above are dubbed as “all-star” players. Stocks are rated based on the picks received, and the players who assigned the picks. Stocks get a star rating from 1 (lowest) to 5 (highest) in increments of 1. Stocks with 1 star are described as “ominous,” 2 stars “unattractive,” 3 stars “appealing,” 4 stars “favorite,” and 5 stars “the best.” Although the algorithm to calculate this star rating is proprietary, it is stressed that the star rating is an aggregation of the positive and negative picks, with higher rated players carrying more weight. Moreover, for a stock to receive a star rating, it must have no less than 10 active picks, including at least one pick from an “all-star” player. If a stock drops below this requirement, it loses its rating.
These unique data were provided to us by Motley Fool CAPS community. The ratings are updated daily, and the sample period is from January 2006 to December 2010. Table I describes the data set. From panel A, we can see that the total firm-day observations in each year do not change much, meaning that we have a relatively balanced panel data set. Since the community officially launched in November 2006, we limit our primary sample period to November 1, 2006 through December 31, 2010. Panel B lists the variables included in the CAPS data set. Ticker ID is the special identification used by the CAPS community. Ticker symbol is the same as that used on the Center of Research on Stock Prices (CRSP), thus we use this identifier to merge our data with CRSP daily data set. Two variables used in the study are worth mentioning – “total outperform picks” and “total underperform picks.” These two variables are different from the “positive picks” and “negative picks” used in Avery et al. (2011). The variables, listed on one date for a stock, indicate the cumulative number of “positive picks” and “negative picks” the stock has received up to that specific date. Since Avery et al. (2011) did not use cumulative picks data, their results are not directly comparable to those reported here.
The CAPS data set is merged with the CRSP daily, and monthly stock files, using date and ticker symbol. We obtain information on stock price, holding period return, shares outstanding, and trading volume, from CRSP. The market risk and style factor data are obtained from Ken French’s Data Library. Using the merged data set, we execute a partial autocorrelation test on the average star rating time series. Figure 1 shows the results of the test. As we can see, the first order partial correlation is very high, indicating the comprehensive star ratings are sticky.
3. Empirical methodology and results
Table II gives the mean and standard deviation of daily stock returns, size measured by market capitalization, and trading volume by star ratings and by stocks. It is interesting to notice that the trading volume for stocks with the highest and lowest star ratings are similar, and both are smaller than that of other stocks. According to previous studies, like Tetlock (2007), we would expect the extreme star ratings to predict higher trading volumes, but the opposite holds true here. We can see a weak increasing trend in the daily stock returns as the star rating increases. There is no monotonic trend in the market capitalization.
Given the difference in market capitalization, and the well-documented fact that smaller firms’ stocks often generate larger returns, we would like to separate the “size effect” from the relationship between star rating and stock performance. For this reason, we group the stocks into deciles with group 1 having the smallest size and group 10 the largest. Table III presents the average daily raw returns of independently double-sorted portfolios. Compared to the rest of the sample, smaller firm stocks have larger daily returns. A strategy of shorting all 1-star stocks and longing all 5-star stocks would generate approximately a 5 percent raw return per annum. However, within a single size group, the trend from 1-star to 5-star stocks does not always exist. The size trend in returns is relatively persistent within each star rating group, declining with size. Contrary to the initial impression, these two observations indicate that CAPS community ratings may not have higher information content.
In the next two subsections, we analyze the performance of portfolios using different strategies. We focus on Jensen’s α, which is generally used to evaluate the performance of investment strategies. We run the following three regressions on different portfolios and pay attention to the constant term, α, where significant positive (negative) α represents abnormally good (bad) performance:
Performance of stock portfolios sorted by star ratings
We first examine the performance of portfolios sorted monthly. Over the period from December 2006 to November 2010, the stocks are sorted by their star rating on the last trading day of each month. These portfolios are held for the following month and the value-weighted returns are calculated. Table IV shows the regression results for value-weighted portfolios. Columns (1) to (5) include the regression results for 1-star to 5-star portfolios, and column (5-1, i.e. 5 minus 1) reports the regression result for the hedge portfolio, which shorts the 1-star portfolio and goes long on the 5-star portfolio. The hedge portfolio neither “beats” the market nor outperforms any of the other portfolios. Note that, as expected, β1 of the hedge portfolio is statistically indifferent from zero for 1 and 3 factor models. We observe no supportive evidence for the “information content” argument. Notice that even stocks marked as “ominous” i.e. the 1-star stocks, do not perform worse than the market. This fact is interesting because Avery et al. (2011) observe that the players in the CAPS community are relatively bullish and claim that the negative picks contribute the most to predictability.
It is possible that the “information content” may exist but is short-lived. To mitigate this concern, we form the portfolios on a daily basis and repeat the regressions. On every trading day from December 29, 2006, to December 30, 2010, the stocks are sorted into portfolios by their star rating on that day. These portfolios are held for the following trading day and the value-weighted portfolio returns are calculated. Table V presents the regression results for these portfolios. Similar to the monthly analysis, the hedge portfolio does not generate abnormal returns which are significantly different from zero. Irrespective of data frequency, and whether we use a single or multi-factor asset pricing model, we observe, contrary to Avery et al. (2011), no evidence of the CAPS community having superior predicting ability. Returns based on previous month’s (or day’s) rating contain no positive abnormal component.
We repeat the above analysis by forming portfolios based on the same-day star rating. That is, on every trading day from January 1, 2007 to December 31, 2010, the stocks are sorted into portfolios by their star rating on that same day. These portfolios’ returns are calculated on the same trading day. If CAPS ratings themselves are determined by contemporaneous returns, we should observe a significant intercept in the regressions. Given our previous results, this will suggest that these ratings do not contain information to forecast returns better than randomly. Instead, these ratings merely reflect return information which the market already knows and merely track this return. We run the three asset pricing regression models. The results are positive for the “stock performance tracking” hypothesis every single hedge portfolio generates an economically and statistically significant positive daily return of no less than 0.09 percent per day, which translates to more than 23 percent per annum. Additionally, the strong monotonicity in returns of star portfolios can be observed in 4 of the 6 regression results. This suggests that stock ratings are a function of contemporaneous returns but cannot forecast future returns.
The difference between Tables V and VI is surprising. If there was information content, we would not expect such a sharp drop in the abnormal returns from the star rating from one day to another, since star ratings are quite stable. If there is no relation between the star rating and the stock returns at all, the patterns in Table VI should not be observed. To test the robustness of these results, we form the portfolios based on star rating changes in the next subsection.
Performance of stock portfolios sorted by star rating changes
Out of the whole sample, about 5 percent of observations have star rating changes. It is possible that change contains more information. Therefore, we sort stocks into three portfolios: star ratings that have increased, star ratings that have decreased, and star ratings that have not changed, all in comparison with their star ratings on the previous trading day. With these, we form a fourth portfolio, the hedge portfolio, longing stocks with increased star ratings, and shorting those with decreased star ratings.
We replicate the tests in section 3A for the three portfolios formed according to the star rating changes. We first look at portfolio performance on the following trading day. Table VII presents the regression results. Following previous results, the hedge portfolio does not generate a statistically significant abnormal return under any case. This strengthens the previous inference that there is no information content in either the level of star ratings or changes in star ratings.
Next, we look at portfolio performance on the same trading day when the star rating changed. Regression results are reported in Table VIII. Here we observe a very strong pattern: stocks without star rating changes generate a modest abnormal return; stocks with an increased star rating generate economically and statistically significant positive abnormal returns; and those whose star rating decreased generate significant negative abnormal returns. Further, the hedge portfolio generates a positive return of more than 1 percent every day. This result largely supports the hypothesis that the star ratings on CAPS community website are actually tracking the market performance.
In this section, we perform additional tests to show that CAPS community is tracking stock performance and does not have predictive ability.
First, we perform Granger-causality tests for whether or not firm stock returns Granger-cause rating changes. Rating changes may result either from market condition changes or from individual stock information content changes perceived by users on Motley Fool CAPS. We separately check if either of those affects stock rating changes. First, we try to determine if the investors on Motley Fool CAPS follow performance changes on individual stocks and adjust their ratings accordingly. As rating changes are infrequent, we test whether the dispersion of stock returns on CAPS will deliver Granger-cause dispersion of rating changes. Larger dispersion of rating changes may carry more information on disagreement of stock performance or it may carry more information on dispersion of rating changes. For each trading day, we calculate the standard deviations of both returns and rating changes for all stocks on CAPS. The Granger-causality test result for whether CAPS return dispersion Granger-cause CAPS ratings change dispersion is shown in Table IX column 2.
The null hypothesis is dispersion on returns do not Granger-cause dispersion in rating changes. The result shows that stock return information does not percolate to stock rating changes until eight days later. In other words, stock performance information that is more than eight days old will start to affect the investors’ decisions on stock rating changes. We also conduct a Granger-causality test on whether rating changes will Granger-cause stock returns. The result is shown in Table IX column 3. From this result, we can conclude that stock rating changes do not infer any information on stock performance until at least 21 days later, except for the subsequent day’s stock return.
Combining two Granger-causality test results, we conclude that individual stock rating changes mostly follow past stock performance and not the other way around.
Next, we conduct a test as to whether value-weighted market returns Granger-cause cross-sectional standard deviation in rating changes. If participants are updating their rating changes based on recent past performance of the market as a whole, this would lend further support to the idea that star ratings are more likely to track performance than to predict it. The Granger-causality tests reported in Table X show that it takes about three weeks for aggregate market returns to percolate into cross-sectional dispersion in ratings and this influence continues for some time. There are signs that this market effect begins to occur earlier—by as much as a week—but the p-values rejecting the null hypothesis of no Granger-causality do not consistently stay below the conventional 5 percent level of significance until the third week.
As an additional test, we examined the possibility that Fama and French (1993) value portfolio returns Granger-cause cross-sectional standard deviation in rating changes. The results are presented in Table XI. The results indicate that while there is only weak evidence that value movements can Granger-cause dispersion in rating changes at a short frequency (e.g. marginal significance at day 4 and 9), it does appear that value return fluctuations can lead to cross-sectional dispersion over a longer horizon. Notably, value portfolio stock returns Granger-cause CAPS rating dispersion by around the fifth week. This, when compared to the shorter timeline for an effect observed from market returns, indicates that value returns may be a slower moving source of influence in determining rating changes. Overall, it appears that rating changes follow prior movements in individual and market-wide movements.
Relation with other sentiment measures
Given that user-submitted rating recommendations might act as a proxy for market-participant sentiment, we test how closely aligned the measure of CAPS sentiment is with other investor sentiment proxies. We start by examining pairwise correlations of monthly CAPS sentiment (as measured by the monthly average of the cross-sectional dispersion of daily rating changes), with the monthly Baker and Wurgler (2006) and modified Baker and Wurgler (Huang et al., 2015) sentiment indices. To test the extent to which cross-sectional dispersion is correlated with the level of macroeconomic uncertainty, as opposed to a more abstract notion of sentiment, we also estimate pairwise correlations between CAPS dispersion and the Bali et al. (2014) macroeconomic uncertainty index.
The Baker and Wurgler (2006) index is constructed by extracting the first principal component from a set of six variables: the closed-end fund discount (i.e. the average difference in NAV and market prices for closed-end mutual funds); NYSE share turnover; the number and first-day returns of IPOs; aggregate seasoned equity issuance as a percentage of existing equity; and the relative spread in book-to-market ratios between dividend payers and non-dividend payers. One of the drawbacks of the resulting index is its limited power to explain subsequent market returns. As a response, Huang et al. (2015) reconstruct the index using the partial-least squares methodology of Kelly and Pruitt (2013), with the effect that the underlying components of the index are better calibrated toward forecasting market returns.
By contrast, the Bali et al. (2014) measure is relatively more focused on macroeconomic variables than on sentiment alone. Here, the methodology is to first estimate the time-varying conditional volatility of the default spread, term spread, inflation, and a number of similar proxies for macroeconomic risk. Then, in order to produce an index, the first principal component is extracted from these resulting conditional volatilities.
Table XII shows that CAPS dispersion is correlated with both Baker and Wurgler (2006) sentiment, as well as Bali et al. (2014) macroeconomic risk. When sentiment is high (i.e. positive), ratings dispersion also tends to be high. Likewise, when macroeconomic risk is low, cross-sectional dispersion also tends to be high. In other words, ratings dispersion likely tends to occur amidst positive sentiment and low macroeconomic uncertainty. During times of stress, dispersion is more compact (i.e. opinions are more concentrated).
Interestingly, however, little relationship appears to exist between CAPS dispersion and the modified Baker-Wurgler sentiment index. This might be because the modified index is constructed so that it better predicts market returns, rather than as an attempt to describe sentiment. Consequently, while we can infer that CAPS may not serve as a useful indicator for predicting future market returns and simply appears to follow market movements, the relatively small correlation between CAPS dispersion and the modified Baker-Wurgler index does not necessarily rule out CAPS’ usefulness as a proxy for investor sentiment.
Further results are presented graphically in Figure 2. Here, monthly sentiment and uncertainty indices are divided into tercile over the full-sample period. Within each of these regimes (low, middle, and high sentiment or uncertainty), a kernel density estimate shows the probability distribution of the cross-sectional dispersion of CAPS rating changes. The results confirm that the lowest dispersion tends to be observed during “bad” times—when sentiment is low, or macroeconomic risk is high.
We test the efficient market hypothesis using the Motley Fool CAPS data set, mainly utilizing the provided star ratings. We find little evidence of “information content” in the Motley Fool CAPS community, and strong justification for the hypothesis that CAPS community is good at tracking existing stock performance. Portfolios sorted on the star ratings fail to generate a significant α. Results are similar if the portfolios are sorted by star rating changes at monthly or daily frequency. There is an increase in abnormal returns from 1-star rated to 5-star rated stocks on the day when the star rating is given but the returns vanish the next day. Our results support the efficient market hypothesis.
Description of motley fool CAPS data set
|Panel A: The number of observations in the Motley Fool CAPS data set by year|
|Year||Number of observations|
|Panel B: List of variables included in the Motley Fool CAPS data set|
|Ticker ID Ticker symbol|
|Date on which the stock received a particular rating/pick|
|Top Fool in the ticker on that date|
|Numerical rank of the ticker across the rated universe|
|Adjusted close price of the stock|
|Total outperform picks|
|Total underperform picks|
|All Star outperform picks|
|All Star underperform picks|
|Wall Street outperform picks|
|Wall Street underperform picks|
|Volume of trade|
Notes: This table presents the Motley Fool CAPS data set before it is merged with CRSP data set. Panel A summarizes the raw number of daily observations (firm-day observations) by year. Panel B lists the names of variables included in the Motley Fool CAPS data set
Notes: This table gives the mean and standard deviation of daily return, size, and trading volume by star rating. Returns are in percentage, size is in million dollars, and trading volume is in million shares. Size is calculated as shares outstanding multiplied by share price
Daily portfolio returns by star rating and size
Notes: This table displays the average daily portfolio returns. Portfolios are formed by double-sorting stocks independently by size and by star rating on the previous trading day. The stocks are grouped into deciles according to the market capitalization with 1 representing the smallest stocks and 10 the largest. All returns are in percentages
Regression results for value-weighted portfolios, sorted monthly
|Portfolios based on star ratings from 1-star to 5-star|
|Panel A: 1-factor value weighted portfolios–monthly|
|MKTEX||1.17569*** (14.83)||1.27978*** (23.19)||1.10944*** (39.38)||0.93939*** (36.82)||0.99589*** (17.41)||−0.17735 (−1.47)|
|Constant||−0.36418 (−0.80)||−0.07041 (−0.22)||0.00030 (0.00)||0.17683 (1.21)||0.21015 (0.64)||0.44046 (0.64)|
|Panel B: 3-factor value weighted portfolios–monthly|
|MKTEX||0.93637*** (14.28)||1.17794*** (19.50)||1.07061*** (33.67)||1.00957*** (42.44)||1.10953*** (18.05)||0.17343 (1.64)|
|SMB||0.25968* (1.74)||0.24380* (1.77)||0.00738 (0.10)||−0.14266** (−2.64)||−0.27468* (−1.97)||−0.51946** (−2.16)|
|HML||0.74305*** (6.18)||0.22671** (2.05)||0.14389** (2.47)||−0.17328*** (−3.98)||−0.25121** (−2.23)||−0.99593*** (−5.13)|
|Constant||−0.27119 (−0.83)||−0.10748 (−0.36)||0.03537 (0.22)||0.18780 (1.59)||0.25302 (0.83)||0.38397 (0.73)|
|Panel C: 4-factor value weighted portfolios–monthly|
|MKTEX||0.89136*** (14.20)||1.12058*** (21.34)||1.04484*** (35.70)||1.02729*** (45.90)||1.11496*** (17.38)||0.22240** (2.09)|
|SMB||0.25989* (1.88)||0.24406** (2.11)||0.00750 (0.12)||−0.14274*** (−2.90)||−0.27470* (−1.95)||−0.51968** (−2.21)|
|HML||0.63609*** (5.42)||0.09042 (0.92)||0.08265 (1.51)||−0.13117*** (−3.13)||−0.23831* (−1.99)||−0.87959*** (−4.41)|
|UMD||−0.14085*** (−2.87)||−0.17948*** (−4.37)||−0.08064*** (−3.52)||0.05545*** (3.17)||0.01700 (0.34)||0.15321* (1.84)|
|Constant||−0.34893 (−1.15)||−0.20653 (−0.81)||−0.00913 (−0.06)||0.21840** (2.02)||0.26240 (0.85)||0.46853 (0.91)|
Notes: t statistics in parentheses. At the end of every month from December 2006 to November 2010, the stocks are sorted by their star rating on the last trading day of that month. These portfolios are held for the following month, and the value-weighted portfolio returns are calculated. Results in columns are reported for 1-star to 5-star portfolios. Panel A presents the regression results of rit−rif=α+βmktext+eit; Panel B presents the regression results of rit−rif=α+β1mktext+β2smbt+β3hmlt+eit; and Panel C presents the regression results of rit−rif=α+β1mktext+β2smbt+β3hmlt+β4umdt+eit. MKTEX is excess market return, SMB is small minus big, HML is high minus low, and t-statistics in parentheses. *p<0.10; **p<0.05; ***p<0.01
Regression results for value-weighted portfolios, sorted daily
|Panel A: 1-factor value weighted portfolios by star ratings on the previous trading day|
|MKTEX||1.22227*** (77.35)||1.24499*** (104.97)||1.07336*** (207.32)||0.96017*** (184.38)||0.95005*** (109.47)||−0.27215*** (−12.27)|
|Constant||−0.00931 (−0.35)||−0.00239 (−0.12)||0.00163 (0.18)||0.00721 (0.81)||0.00906 (0.61)||0.01193 (0.32)|
|Panel B: 3-factor value weighted portfolios by star ratings on the previous trading day|
|MKTEX||1.03394*** (75.96)||1.09045*** (105.24)||1.04636*** (176.16)||1.01885*** (207.72)||1.01136*** (106.07)||−0.02253 (−1.11)|
|SMB||0.46495*** (15.90)||0.20904*** (9.39)||0.00332 (0.26)||−0.10667*** (−10.12)||−0.10728*** (−5.24)||−0.57183*** (−13.12)|
|HML||0.74245*** (25.29)||0.61394*** (27.47)||0.10816*** (8.44)||−0.23237*** (−21.96)||−0.24290*** (−11.81)||−0.98526*** (−22.52)|
|Constant||−0.00974 (−0.50)||−0.00000 (−0.00)||0.00258 (0.30)||0.00673 (0.96)||0.00850 (0.62)||0.01179 (0.40)|
|Panel C: 4-factor value weighted portfolios by star ratings on the previous trading day|
|MKTEX||0.97628*** (76.45)||1.03843*** (112.46)||1.03124*** (172.22)||1.03731*** (219.87)||1.03118*** (105.95)||0.05487*** (2.82)|
|SMB||0.48117*** (18.31)||0.22367*** (11.77)||0.00757 (0.61)||−0.11186*** (−11.52)||−0.11285*** (−5.64)||−0.59359*** (−14.82)|
|HML||0.52334*** (17.51)||0.41630*** (19.27)||0.05072*** (3.62)||−0.16223*** (−14.69)||−0.16759*** (−7.36)||−0.69116*** (−15.18)|
|UMD||−0.25492*** (−15.56)||−0.22994*** (−19.41)||−0.06684*** (−8.70)||0.08160*** (13.48)||0.08761*** (7.02)||0.34217*** (13.70)|
|Constant||−0.01602 (−0.91)||−0.00567 (−0.45)||0.00093 (0.11)||0.00875 (1.35)||0.01066 (0.80)||0.02023 (0.76)|
Notes: t-statistics in parentheses. At the end of every trading day from December 29, 2006, to December 30, 2010, the stocks are sorted by their star rating on that day. These portfolios are held for the following trading day and the value-weighted portfolio returns are calculated. Panel A presents the regression results of rit−rif=α+βmktext+eit; Panel B presents the regression results of rit−rif=α+β1mktext+β2smbt+β3hmlt+eit; and Panel C presents the regression results of rit−rif=α+β1mktext+β2smbt+β3hmlt+β4umdt+eit. MKTEX is excess market return, SMB is small minus big, HML is high minus low, and t-statistics in parentheses. ***p<0.01
Regression results for value-weighted portfolios, sorted daily by contemporaneous star rating
|Panel A: 1-factor value weighted portfolios by star ratings on the same trading day|
|MKTEX||1.21935*** (77.78)||1.24522*** (110.10)||1.07559*** (200.13)||0.95852*** (182.25)||0.95085*** (112.60)||−0.26843*** (−12.32)|
|Constant||−0.04840* (−1.81)||−0.04195** (−2.18)||−0.00875 (−0.96)||0.00908 (1.01)||0.05078*** (3.53)||0.09274** (2.50)|
|Panel B: 3-factor value weighted portfolios by star ratings on the same trading day|
|MKTEX||1.03260*** (76.83)||1.09820*** (110.62)||1.04537*** (170.55)||1.01852*** (206.45)||1.00902*** (108.43)||−0.02353 (−1.19)|
|SMB||0.47026*** (16.28)||0.19196*** (9.00)||0.00743 (0.56)||−0.10549*** (−9.95)||−0.10631*** (−5.32)||−0.57617*** (−13.51)|
|HML||0.73595*** (25.39)||0.58426*** (27.29)||0.12099*** (9.15)||−0.23769*** (−22.34)||−0.23036*** (−11.48)||−0.96623*** (−22.57)|
|Constant||−0.04897** (−2.54)||−0.03956*** (−2.78)||−0.00775 (−0.88)||0.00854 (1.21)||0.05032*** (3.77)||0.09284*** (3.26)|
|Panel C: 4-factor value weighted portfolios by star ratings on the same trading day|
|MKTEX||0.97602*** (77.28)||1.04934*** (117.73)||1.02992*** (166.56)||1.03652*** (217.25)||1.02778*** (108.05)||0.05172*** (2.71)|
|SMB||0.48617*** (18.71)||0.20570*** (11.22)||0.01177 (0.93)||−0.11055*** (−11.26)||−0.11158*** (−5.70)||−0.59733*** (−15.23)|
|HML||0.52096*** (17.63)||0.39864*** (19.11)||0.06227*** (4.30)||−0.16928*** (−15.16)||−0.15909*** (−7.15)||−0.68028*** (−15.25)|
|UMD||−0.25013*** (−15.44)||−0.21596*** (−18.88)||−0.06831*** (−8.61)||0.07958*** (13.00)||0.08292*** (6.79)||0.33269*** (13.60)|
|Constant||−0.05514*** (−3.18)||−0.04488*** (−3.67)||−0.00943 (−1.11)||0.01050 (1.60)||0.05236*** (4.01)||0.10105*** (3.86)|
Notes: t-statistics in parentheses. On every trading day from January 3, 2007, to December 31, 2010, the stocks are sorted into portfolios by their star rating on that day. The value-weighted portfolio returns are calculated on the same day. Panel A presents the regression results of rit−rif=α+βmktext+eit; Panel B presents the regression results of rit−rif=α+β1mktext+β2smbt+β3hmlt+eit; and Panel C presents the regression results of rit−rif=α+β1mktext+β2smbt+β3hmlt+β4umdt+eit. MKTEX is excess market return, SMB is small minus big, HML is high minus low, and t-statistics in parentheses. *p<0.10; **p<0.05; ***p<0.01
Regression results for value-weighted portfolios, sorted daily by star rating changes
|Panel A: 1-factor value weighted portfolios by star rating changes on previous trading day|
|MKTEX||1.02781*** (460.62)||1.00693*** (57.12)||1.01645*** (63.91)||−0.02943 (−1.26)|
|Constant||0.00476 (1.26)||0.02171 (0.73)||−0.02161 (−0.80)||0.04233 (1.07)|
|Panel B: 3-factor value weighted portfolios by star rating changes on previous trading day|
|MKTEX||1.03174*** (395.10)||1.03791*** (50.31)||1.01653*** (54.75)||−0.01299 (−0.47)|
|SMB||−0.02191*** (−3.90)||0.11360** (2.57)||0.19145*** (4.81)||−0.07951 (−1.35)|
|HML||−0.01559*** (−2.74)||−0.13094*** (−2.88)||−0.00299 (−0.07)||−0.06666 (−1.10)|
|Constant||0.00488 (1.30)||0.01850 (0.62)||−0.02462 (−0.92)||0.04283 (1.08)|
|Panel C: 4-factor value weighted portfolios by star rating changes on previous trading day|
|MKTEX||1.03308*** (378.94)||1.03400*** (48.02)||1.03679*** (53.79)||−0.02645 (−0.92)|
|SMB||−0.02229*** (−3.96)||0.11476*** (2.59)||0.18522*** (4.68)||−0.07545 (−1.28)|
|HML||−0.01049 (−1.63)||−0.14581*** (−2.85)||0.07283 (1.59)||−0.11886* (−1.74)|
|UMD||0.00595* (1.70)||−0.01765 (−0.64)||0.09012*** (3.64)||−0.06161* (−1.66)|
|Constant||0.00502 (1.34)||0.01816 (0.61)||−0.02271 (−0.86)||0.04171 (1.05)|
Notes: t statistics in parentheses. At the end of every trading day from January 3, 2006, to December 30, 2010, the stocks are sorted by their star rating changes from the previous trading day to that day. These portfolios are held for the following trading day and the value-weighted portfolio returns are calculated. Panel A presents the regression results of rit−rif=α+βmktext+eit; Panel B presents the regression results of rit−rif=α+β1mktext+β2smbt+β3hmlt+eit; and Panel C presents the regression results of rit−rif=α+β1mktext+β2smbt+β3hmlt+β4umdt+eit. MKTEX is excess market return, SMB is small minus big, HML is high minus low, and t-statistics in parentheses. *p<0.10; **p<0.05; ***p<0.01
Regression results for value-weighted portfolios, sorted daily by contemporaneous star rating changes
|Panel A: 1-factor value weighted portfolios by star rating changes on the same trading day|
|MKTEX||1.02401*** (462.36)||1.04063*** (60.87)||1.02816*** (56.30)||0.01148 (0.43)|
|Constant||0.00455 (1.21)||0.67215*** (23.10)||−0.69495*** (−22.30)||1.36436*** (29.67)|
|Panel B: 3-factor value weighted portfolios by star rating changes on the same trading day|
|MKTEX||1.02820*** (393.56)||0.97755*** (49.46)||1.03533*** (47.76)||−0.06321** (−1.99)|
|SMB||−0.01209** (−2.15)||0.20296*** (4.79)||0.07371 (1.58)||0.12723* (1.87)|
|HML||−0.01646*** (−2.92)||0.24596*** (5.79)||−0.03085 (−0.66)||0.29429*** (4.30)|
|Constant||0.00459 (1.22)||0.67160*** (23.71)||−0.69630*** (−22.34)||1.36577*** (30.00)|
|Panel C: 4-factor value weighted portfolios by star rating changes on the same trading day|
|MKTEX||1.02975*** (377.72)||0.96093*** (46.72)||1.05163*** (46.55)||−0.09677*** (−2.94)|
|SMB||−0.01253** (−2.23)||0.20760*** (4.92)||0.06922 (1.49)||0.13652** (2.02)|
|HML||−0.01056* (−1.66)||0.18254*** (3.80)||0.03070 (0.58)||0.16687** (2.17)|
|UMD||0.00687** (1.96)||−0.07364*** (−2.81)||0.07190** (2.49)||−0.14864*** (−3.54)|
|Constant||0.00476 (1.27)||0.66990*** (23.73)||−0.69479*** (−22.35)||1.36261*** (30.10)|
Notes: t statistics in parentheses. On every trading day from January 3, 2007, to December 31, 2010, the stocks are sorted into portfolios by their star rating changes from the previous trading day to that day. The value-weighted portfolio returns are calculated over the same time period. Panel A presents the regression results of rit−rif=α+βmktext+eit; Panel B presents the regression results of rit−rif=α+β1mktext+β2smbt+β3hmlt+eit; and Panel C presents the regression results of rit−rif=α+β1mktext+β2smbt+β3hmlt+β4umdt+eit. MKTEX is excess market return, SMB is small minus big, HML is high minus low, and t-statistics in parentheses. *p<0.10; **p<0.05; ***p<0.01
Granger-causality (GC) test
|Day||P-value of GC tests on returns||P-value of GC tests on ratings changes|
Notes: Column (1) reports the p-values from Granger-causality test with the null hypothesis that CAPS cross-sectional dispersion on returns do not Granger-cause CAPS cross-sectional dispersion in rating changes. Column (2) reports the p-values from Granger-causality test with the null hypothesis that CAPS cross-sectional dispersion on rating changes do not Granger-cause CAPS cross-sectional dispersion in stock returns. Lags of 1 to 30 days are used in the test
Granger-causality test for whether market returns Granger-cause CAPS rating dispersion
|Day||Likelihood ratio test||SSR based χ2||Day||Likelihood ratio test||SSR based χ2|
Notes: This table reports the p-values from Granger-causality test with the null hypothesis that daily CRSP value-weighted market index returns do not Granger-cause CAPS cross-sectional dispersion in rating changes. Column 1 displays the number of days measured with respect to lagged market returns, and columns 2 and 3 display p-values from likelihood ratio and sum-of-squares χ2 tests, respectively
Granger-causality test for whether HML returns Granger-cause CAPS rating dispersion
|Day||Likelihood ratio test||SSR based χ2||Day||Likelihood ratio test||SSR based χ2|
Notes: This table reports the p-values from Granger-causality test with the null hypothesis that daily Fama-French (1993) value portfolio returns do not Granger-cause CAPS cross-sectional dispersion in rating changes. Column 1 displays the number of days measured with respect to lagged HML returns, and columns 2 and 3 display p-values from likelihood ratio and sum-of-squares χ2 tests, respectively
Summary statistics and correlations between CAPS rating dispersion, investor sentiment (Baker-Wurgler (2007)), and macroeconomic uncertainty (Bali, 2014; Huang et al., 2015)
|Panel A: Monthly Summary Statistic|
|Panel B: Means by year|
|Year||Mean Rating||Rating Disp.||Econ Uncert||Baker-Wurgler||Modified BW|
|Panel C: Correlation table|
Section 2 explains the rating system and the justification of relying upon star ratings.
All empirical tests reported in this section are based on value-weighted portfolios. These tests were repeated using equally weighted portfolios for robustness purposes. Results so obtained were similar to the value-weighted results and are available from the author upon request.
Since our test period overlaps with the subprime crises, any violation of the various asset pricing models utilized in this study due to the crises may have generated significant alphas. Not finding significant alphas in our study shows the robustness of the models used and the lack of any information content in the Motley Fool CAPS community.
It would be ideal to distinguish between stocks whose star ratings drop from 5 to 1 from those whose star ratings drop from 4 to 3, but the first case is so uncommon that we do not have enough observations to perform any meaningful test.
Antweiler, W. and Frank, M.Z. (2004), “Is all that talk just noise? The information content of internet stock message boards”, The Journal of Finance, Vol. 59 No. 3, pp. 1259-1294.
Avery, C., Chevalier, J.A. and Zeckhauser, R.J. (2011), The “Caps” Prediction System and Stock Market Returns, National Bureau of Economic Research, Cambridge, MA.
Baker, M. and Wurgler, J. (2006), “Investor sentiment and the cross‐section of stock returns”, The Journal of Finance, Vol. 61 No. 4, pp. 1645-1680.
Bali, T.G., Brown, S.J. and Caglayan, M.O. (2014), “Macroeconomic risk and hedge fund returns”, Journal of Financial Economics, Vol. 114, pp. 1-19.
Barber, B.M. and Odean, T. (2000), “Trading is hazardous to your wealth: the common stock investment performance of individual investors”, Journal of Finance, Vol. 55 No. 2, pp. 773-806.
Barber, B.M. and Odean, T. (2001), “The internet and the investor”, Journal of Economic Perspectives, Vol. 15 No. 1, pp. 41-54.
Barber, B.M., Odean, T. and Zhu, N. (2009), “Do retail trades move markets?”, Review of Financial Studies, Vol. 22, pp. 151-186.
Chevalier, J. and Ellison, G. (1999), “Are some mutual funds managers better than others? Cross-Sectional patterns in behavior and performance”, The Journal of Finance, Vol. 54 No. 3, pp. 875-899.
Coval, J.D., Hirshleifer, D.A. and Shumway, T. (2005), “Can individual investors beat the market?”, HBS Finance Working Paper No. 04-025, Boston, MA.
Da, Z., Engelberg, J. and Gao, P. (2011), “In search of attention”, The Journal of Finance, Vol. 66 No. 5, pp. 1461-1499.
Fama, E.F. and French, K.R. (1993), “Common risk factors in the returns on stocks and bonds”, Journal of Financial Economics, Vol. 33 No. 1, pp. 3-56.
Huang, D., Jiang, F., Tu, J. and Zhou, G. (2015), “Investor sentiment aligned: a powerful predictor of stock returns”, Review of Financial Studies, Vol. 28 No. 3, pp. 791-837.
Kelly, B. and Pruitt, S. (2013), “Market expectations in the cross-section of present values”, The Journal of Finance, Vol. 68 No. 5, pp. 1721-1756.
Metrick, A. (1999), “Performance evaluation with transactions data: the stock selection of investment newsletters”, The Journal of Finance, Vol. 54 No. 5, pp. 1743-1775.
Mikhail, M.B., Walther, B.R. and Willis, R.H. (2004), “Do security analysts exhibit persistent differences in stock picking ability?”, Journal of Financial Economics, Vol. 74 No. 1, pp. 67-91.
Sibley, S., Xing, Y. and Zhang, X. (2016), “The information content of the sentiment index”, Journal of Banking & Finance, Vol. 62 No. 1, pp. 164-179.
Tetlock, P.C. (2007), “Giving content to investor sentiment: the role of media in the stock market”, The Journal of Finance, Vol. 62 No. 3, pp. 1139-1168.