How to Find and Capitalize on Asset-Pricing Anomalies

Asset-pricing anomalies are trading signals that predict future abnormal returns. A November 2023 study “Anomaly Time,” published in the October 2024 issue of The Journal of Finance, demonstrated that anomalies based on annual financial data have provided abnormal returns that were largest in the first month after the release of financial statements and then decayed quickly thereafter. The authors—Boone Bowles, Adam Reed, Matthew Ringgenberg, and Jacob R. Thornock—write: “In informationally efficient capital markets, investors compete to profit from asset-pricing anomalies before mispricing is arbitraged away. This competition creates a race to acquire and process information quickly, and this race begins as soon as information about an anomaly signal becomes public.” Thus, “anomaly returns are at least partly due to delayed information processing by investors.”

The authors also found that “the academic convention of forming portfolios in June causes researchers to significantly underestimate return predictability, which makes some anomalies appear insignificant even though they do reliably predict returns in the days and weeks immediately after information is first released”—”anomaly returns are much stronger using 10-K filing dates, instead of June rebalancing.” And finally, they found that “in recent years, the returns to anomaly strategies are increasingly earned in the first few weeks after information releases, consistent with technological improvements leading to lower processing costs”—the markets are becoming increasingly efficient. From McLean and Pontiff’s list of anomalies, they used 28 anomalies that rely on information released in either earnings announcements or 10-K filings and found:

• The daily abnormal return to the average anomaly portfolio was 9.84%, annualized, over the first month following the release of information.
• Over the first four months after an information release, the average daily abnormal return was just 4.69%, annualized, which implies the returns significantly dissipated after the first month.
• In the period after these first four months, the average daily abnormal return fell to 1.99%, annualized.
• Price discovery happened twice as fast after the implementation in 1996 of EDGAR, relative to before—a reduction in information processing costs leads to faster arbitrage and shorter periods of return predictability and the market becomes more efficient.

Can Anomalies Be Predicted?

In their August 2024 follow-up study, “Predicting Anomalies,” the authors examined whether anomaly returns exhibit predictable patterns in the weeks and months before the trading signals public release. They tested several anomaly prediction models, including:

• Sophisticated machine learning models.
• Autoregressive models, which are statistical models used to describe a time series where the value of a variable at a given time point is modeled as a linear function of its own previous values.
• Simple martingale model, which is a sequence of random variables where the conditional expected value of the next value in the sequence is equal to the present value, regardless of all prior values.

They found that a simple martingale model worked best.

For example, for the asset growth anomaly, they showed that asset growth at third quarter was a highly reliable predictor for asset growth at fourth quarter—making it possible to trade the asset growth anomaly early. The authors examined the same 28 anomalies that were used in their earlier paper. They relied on Compustat Snapshot, a database that shows when accounting information about a firm was made publicly available. Their sample period begins in January 1990 and runs through 2019. The following is a summary of their key findings:

• In general, anomaly signals exhibit positive serial correlation—stock returns follow predictable patterns prior to the publication of anomaly trading signals. For example, for the average anomaly, they found that most stocks in the anomaly portfolio in the third quarter remained in the portfolio at the end of the fourth quarter—66% of the annual portfolio’s stocks were holdovers from the portfolio at third quarter. Thus, by simply assuming persistence in accounting numbers from third to fourth quarter, a trader would make the correct portfolio assignment two thirds of the time, making it possible to trade before financial statements are released.
• For the average anomaly, constructing anomaly portfolios three months before the annual information release using the martingale model resulted in additional returns of 2.99% on an annualized basis. (That is, these returns are in addition to the returns earned from trading on the anomaly after the information becomes public.) Using the machine learning model instead generated additional annualized returns of 3.65%.
• In recent periods, this predictability was concentrated in signals that are harder to forecast, and returns were increasingly earned several quarters before signals were released.

Interestingly, the authors found that the persistence of anomalies performance varied across anomalies; some accounting numbers were highly persistent, while others were not. For example, for both the asset turnover and the profit margin anomalies, 85% of the stocks in the annual portfolio were holdovers from the third quarter. Thus, their prediction models performed strongly with accuracy of more than 91%. On the other hand, only 25% of the stocks in the earnings surprise anomaly portfolio at fourth quarter were also in the portfolio at third quarter. As a result, the prediction models performed poorly for this anomaly, with an accuracy rate of just 28%. (Similarly poor results were found for the revenue surprise anomaly.) This should not be surprising because by definition these are by construction surprises (not predictable).

Most importantly, the authors found: “Though the martingale and machine learning models were consistently accurate over our entire sample period, the returns from trading based on these models have been arbitraged away more quickly in recent years. In the early years of our sample, trading three months early using the martingale and machine learning models earned over 200 basis points. In the recent period, the martingale model lost 13 basis points while the machine learning model earned only 7. These findings suggest traders are using predictive models similar to ours to arbitrage away return predictability in the quarter before information is released.”

Another interesting finding: “In recent years correct predictions are less profitable while ‘wrong’ predictions are even more costly. Indeed, stocks that are in the ‘missed’ portfolio earn a much larger share of returns from trading prior to information releases. This finding suggests that investors have arbitraged away the predictable component of anomalies, but not the difficult to predict component. Interestingly, we also find that although returns to trading three months early have been arbitraged away in recent years, there are still predictable returns to trading six months early using a two quarter ahead martingale model. In other words, predicting anomalies can still be profitable, but you need to act earlier in event time.”

These findings are consistent with those of Robin Greenwood and Marco Sammon, authors of the September 2024 study “The Disappearing Index Effect,” which found that index addition and deletion premiums are moving earlier and earlier in event time as a result of arbitrage competition.

Their findings led Bowles et al. to conclude: “Our findings provide strong evidence that anomalies are really in the data, and they are related to information about anomaly signals.” They added: “Our results show that anomalies are more anomalous than previously realized.”

Investor Takeaways

These studies show how markets become more efficient over time, as information becomes faster and cheaper to access. As findings are published, anomaly returns tend to shrink, and when the anomalies are behavioral-based, they are often eliminated once trading costs are considered. In other words, the markets are becoming increasingly efficient as large databases and faster computers enable arbitragers to uncover anomalies and learn how to exploit them.