The Efficient Market Hypothesis posits that the stock market is efficient at incorporating publicly available information. If that is true, then it should be next to impossible to select stocks that will produce returns superior to the overall market’s returns, after making allowance for the cost of risk. But decades of academic research have turned up numerous “anomalies.” These anomalies appear to show that there have been systematic ways to choose stocks that will “beat the market,” and that success is attributable not just to a few individuals with idiosyncratic skills.
In some instances, it has been possible to show that the trading costs of taking advantage of these anomalies would wipe out any superior returns that they might produce. But that does not explain the persistence of most of them. One of the first anomalies to be discovered, and perhaps the best known, is the historical tendency of the stocks of small-capitalization companies to produce superior returns to the overall market (though certainly not at all times).
As the label “anomaly” suggests, the hypothesis of market efficiency is intuitively reasonable as a representation of the normal state of affairs, from which these phenomena are seen as departures. If many participants in the market know a strategy for beating the market, then they will use it until their trading drives up prices and the strategy ceases to provide value. The Efficient Markets Hypothesis posits that advantages are traded away almost instantaneously.
R. David McLean, of the University of Alberta, and his colleague Jeffrey Pontiff, of Boston College, have been conducting a large-scale study to see if the advantages of stock-market anomalies are traded away once they are published in academic papers and therefore become widely known, and if so, how quickly. McLean and Pontiff have not yet published their research, but McLean discussed it at a luncheon meeting of the Boston Security Analysts Society on April 15.
Three explanations for anomalies
McLean explained that there have been three responses to the documented existence of such anomalies.
One has been to say that they merely reflect “data mining” (also sometimes called “data snooping”). McLean referred to a 1998 paper by Eugene Fama, one of the fathers of the Efficient Markets Hypothesis, that made just this point. The argument is that if you test thousands of stock selection strategies, a few of them are bound to work through sheer luck. And if luck is the source of the winning strategies, hardly any of them can be expected to work during a time period outside the original test.
Another response, also identified with Fama, is that the success of the strategies is real, but the realized excess returns, called “alphas,” aren’t actually alphas. Rather, they are compensation for unrecognized systematic risks that are not included in risk models of the stock market but ought to be. This is the old problem of the “joint hypothesis” of market efficiency and any risk model: It is not possible to test the hypothesis of market efficiency independently of a model that provides estimates of the risks being incorporated into prices. (This response lies behind the Fama-French model that is an alternative to the Capital Asset Pricing Model.)
This is a conundrum. McLean said that he does not, however, have much sympathy for the argument, which is more a supposition that systematic risk should be able to explain away any regularity in the success of the stock-selection strategies. But he pointed out that one consequence of the argument is that, contrary to the first response, it implies that these strategies should be successful outside the time period of the original tests.
The third response is that these stock-selection strategies genuinely produce superior results. A likely consequence of this argument is that, because the strategies work, their effectiveness should decline, as they become widely known, to the point where they cease to work after taking into account trading costs.
The three responses to the evidence for successful stock-selection strategies imply three different possibly measureable results.
How the research was conducted
McLean’s and Pontiff’s empirical investigation has three primary parts: identifying and replicating existing research on market anomalies, determining whether each anomaly continued to exist between the end of the period studied and the publication of the research and assessing whether, and to what degree, the anomaly continued to exist after the research was published. (The study does not consider market timing, only stock selection.)
McLean and Pontiff, after reviewing the academic literature, initially identified 95 published strategies that ostensibly predicted stock returns. (As their research has continued, they have accumulated more than 100, including some from unpublished papers.) In the full spirit of experimental investigation, they tried to replicate these studies and found that they could replicate the findings of only 78 of them within the spans of time of the original studies. They pulled these strategies mainly from three leading finance journals — McLean didn’t say which ones — with a few strategies drawn from other publications. The oldest paper that they drew upon was Blume and Husic’s report on the low-price anomaly, published in in the Journal of Finance in 1972. For replication of the studies, they used Compustat for accounting data, the Center for Research in Security Prices (CRSP) for price and return data, and the Institutional Brokers’ Estimate System (IBES) for analysts’ earnings estimates.
McLean commented that only in very few cases did it seem as if the strategy was derived from theoretical reasoning. He said that it looked more as if the researchers were just coming up with an investment methodology that could have been suggested by anything, testing it to see if it worked and therefore violated classical financial theory, and then, sometimes, producing a story or rationale to justify why it worked. He added that this seemed to him to reverse the way that science usually proceeds.
For those studies that they could replicate, McLean and Pontiff looked to see if each strategy continued to work during the interval between the completion of the study and its publication in a journal or its posting on the Social Science Research Network (SSRN). McLean said that it typically takes about four to five years from the end of the study period for such work to appear in print, but all the while, the researchers are giving talks about their research and sending working papers to each other before publication (much as McLean and Pontiff have posted on SSRN a working paper representing a slightly earlier version of McLean’s talk). During this period, the strategies are being made known, so anyone could try to apply them in practice, potentially trading away their advantage. Even before the academics began their research, it’s entirely possible, McLean said, that the strategies were known to and being used by practitioners, but such knowledge was probably limited to just a few.
For each strategy, McLean’s and Pontiff’s test was very simple. They constructed a long-short portfolio of stocks that were undervalued or overvalued according to the strategy’s measure at the start of each month, and they called the subsequent one-month (market-neutral) return to the portfolio its “alpha.” They did not adjust for trading costs or risk. McLean explained that there was no obvious way of correcting for risk, because depending upon how one did so, it could explain away all excess returns. That is, they could fall into the trap of the joint hypothesis.
The results of the study
During the span of the original tests (329 months on average), the average return for all the replicable strategies was about 79 basis points a month (before transaction costs). Between the end of the study of a strategy and its publication (44 months on average), the “alpha” of the long-short portfolio, as constructed by McLean and Pontiff, declined on average by 13% to 68.5 basis points a month. The difference, however, was not statistically significant. Post-publication (through 2012, so the post-publication number of months was different for nearly all the strategies in the study), there was a further average decline of 31%, to 48 basis points a month. The total average decline was 44%. Put very roughly, the advantage was halved from the first time period to the third.
But these are average declines. McLean and Pontiff found that strategies that they deemed relatively easy to trade upon, mainly because they involved large-capitalization and highly liquid stocks or, like price-momentum strategies, they used straightforward, readily available information, typically had a greater decline in effectiveness. Strategies that they deemed relatively difficult to trade upon typically had a lesser decline. Also, the larger the initial alpha, the greater the decline in effectiveness, and the greater the statistical significance of the initial test result, on average, the greater the decline in effectiveness.
There was one surprising outlier among the strategies: earnings surprise. The version of this strategy that McLean and Pontiff tested, which has been known and applied since the 1980s, is to divide the difference between a company’s most recent quarterly earnings and its earnings four quarters earlier by the standard deviation of earnings during the previous eight quarters. During the sample period, this strategy produced an average alpha of 2.5 basis points per month (before transactions costs), and this figure did not decay during either the post-testing or the post-publication period. Why this strategy should persist so strongly, McLean said, puzzled him. He suggested that perhaps it is difficult to apply in practice, although it is, indeed, used.
One consequence of these results for investment practitioners, McLean warned, is that those doing backtests of a published strategy (perhaps on a different universe of stocks) shouldn’t confine their tests to the period used by the academic researchers, because the results will likely be upwardly biased.
McLean and Pontiff continued their study to see how quickly the effectiveness of the strategies declined after publication. They found that during the first year after publication, the average alpha declined by only 10 basis points. During the second year, the decline was, on average, 15 basis points. During the third year, the decline was more than 20 basis points, and during the fourth year, 50 basis points. This is not, McLean pointed out, like the responses we see in typical “event studies,” which generally find that when news is released, the market responds almost instantaneously with next to no lingering after effect. Rather, it takes years, on average, for the market to trade away the advantages of a published anomaly.
McLean and Pontiff asked themselves whether decay in effectiveness was attributable to the publication of the strategies or whether it merely reflected an overall tendency, during the last few decades, for the market to become more efficient (implying that the strategies would have become less effective anyway). They were able to show the former, that the effect of publication caused the strategies’ effectiveness to decline.
Which strategies worked the best?
McLean and Pontiff grouped the strategies into four clusters:
- Event strategies (reflecting responses to particular kinds of company announcements or economic events, like the announcement of a share repurchase)
- Fundamental strategies (based on accounting data)
- Market strategies (based on company size, volume or price, like price momentum)
- Valuation (ratios, like price to book or price to sales).
By repeating their original analysis cluster by cluster, they found that the market strategies, during the period of research, had the highest alphas and, post publication, the biggest declines in alpha. Similarly, and in general, the easiest strategies to execute, which happened also to be the most profitable, had the biggest declines in effectiveness after publication.
Strategy performance has its own momentum. McLean and Pontiff found that if a strategy worked during one month, it had an increased likelihood of working the following month. Unsurprisingly, they found that some strategies worked better with some kinds of stocks (for example, those of small-capitalization companies) than with other kinds.
Furthermore, they found that strategies that entailed stocks with the biggest dollar-trading volume, which are the ones that large institutions can more easily trade (because of the lesser impact on price of their trading), declined more in effectiveness. Strategies with large idiosyncratic volatilities, that is, volatilities that are more difficult to hedge, tended to decline less. In sum, if a strategy is costly to execute, it declines less in effectiveness after publication, presumably because investors use the strategy less, even when they know about it.
McLean and Pontiff looked at behavior other than returns before and after the publication of a strategy. In particular, they investigated whether stock variance, trading volume, dollar volume and short interest of stocks on the short side of their strategies increased after publication. For example, McLean said, when investors learn that a price momentum strategy is effective, would they increase trading in stocks that had positive or negative momentum, thereby increasing trading volume, volatility and the short interest in stocks with negative momentum?
Sure enough, McLean and Pontiff found that all four of these characteristics (variance, trading volume, dollar volume and short interest of stocks on the short side of a strategy) increased after publication of a strategy. Short selling was especially interesting. They found that before publication, there was, on average, more short interest on the short side than on the long side of a strategy, perhaps reflecting some knowledge and trading upon the strategy. But after publication, the difference between short interest on the short side and short interest on the long side increased, on average, by a factor of three, a clear indication that there was much more trading on the strategy.
The last component of McLean and Pontiff’s investigation concerned the relationships among the strategies. They found that there was, on average, a very small but statistically significant relationship among the strategies before publication: They had an average correlation of 5%. When they regressed the returns to each strategy with the returns to other strategies before publication, the beta was about 0.70. After publication, however, there was a divergence. Published strategies became completely uncorrelated with unpublished strategies, and the beta fell to 0.013, but they became more highly correlated with other published strategies, and the beta increased to 0.56. This, McLean said, suggests that there are considerations external to the strategies themselves, perhaps related to fund flows, that cause investors to use strategies together after they begin to execute them on a large scale. Even so, some strategies are negatively correlated, which is good for investors, because that increases diversification of portfolios.
Can investors profit from academic research?
McLean concluded with a brief overview of how this work might be used in practice. He said that his colleague, Pontiff, has been successfully trading upon some of the strategies over the last four years, and the two of them may write another paper on this.
McLean said that, ultimately, the investor wants a single equation that produces a forecast return for each stock for a desired holding period by combining and weighting all the strategies that have been shown to work and adjusting for trading costs. The equation can be re-estimated as new strategies come along and as existing ones lose their effectiveness. He was not offering a recipe for success, or even recommending specific strategies. Although he said that different strategies work better for different kinds of stocks, he did not discuss how to incorporate this complication into the equation.
Institutional investors and advisory firms have used this kind of methodology for a long time, as I’ve seen from the outset of my career in investment management. For example, Chicago Investment Analytics long produced forecast alphas based upon the output of a large number of tested and successful investment strategies. The approach is not very different from the simple “multi-factor models” used by quantitatively oriented institutional investment managers like Batterymarch and Trinity Investment Management since the late 1970s.
The real value and consequence of McLean’s and Pontiff’s study is more intellectual than practical. They have shown that academic research into stock-selection methods is taken up by investors and increases market efficiency where that efficiency did not exist before.
Adam Jared Apt, CFA, is a financial advisor and the owner of Peabody River Asset Management, based in Cambridge, MA.
Read more articles by Adam Jared Apt