How Is This Research Useful to Practitioners?
Stock returns that cannot be explained by traditional asset pricing models may exist because of systematic risk compensation, market efficiency, or extensive data mining. Data-mining concerns may arise in the form of spurious patterns that emerge when data are scrutinized.
The authors find that previous researchers have not examined in depth the impact of data mining on the entire (i.e., published and unpublished) set of cross-sectional stock return anomalies considered by researchers, making it difficult to draw proper inferences. These authors differ from previous researchers by focusing on a large universe of fundamentals-based variables to examine the effects of data mining. Prominent anomalies come from financial statement variables—including unreported variables that may have been considered in previous studies—and, in many cases, from new fundamental signals that have received little attention in previous studies. The authors’ results indicate that top-ranked fundamental signals exhibit superior long–short performance that is not due to random chance.
According to the authors, fundamentals-based anomalies are consistent with mispricing explanations. The predictive ability of top fundamental signals is more pronounced among stocks that are small and have low institutional ownership, high idiosyncratic volatility, and low analyst coverage. They find that anomalous returns are significantly higher after high-sentiment periods and that long–short returns of top fundamental signals are higher during recessions than during expansions, debunking the suspicion that superior performance may be compensation for systematic risk. The authors determine that anomalous variables constructed on the basis of interest expense; tax loss carryforward; and selling, general, and administrative expenses are highly correlated with future stock returns. They extend their approach to past returns-based anomalous variables and find broad support for the intermediate-term momentum effect and for a new longer-run (i.e., beyond the past five years) reversal effect but only selective support for the long-run reversal effect.
How Did the Authors Conduct This Research?
In constructing the universe of fundamental signals, the authors use 240 accounting variables and consider 76 financial ratio configurations for each variable, resulting in a universe of over 18,000 fundamental signals. The sample starts in July 1963 and ends in December 2013.
The authors form long–short portfolios for each fundamental signal, using a bootstrap procedure to assess the significance of long–short returns. As a nonparametric method for estimating the distribution of an estimator by resampling the data, bootstrapping is desirable because long–short returns are highly non-normal, exhibiting complex cross-sectional dependencies across fundamental signals, and because the performance evaluation of a large number of fundamental signals involves a multiple-comparison problem.
The authors form long–short portfolios for each fundamental signal, using a bootstrap procedure to assess the significance of long–short returns. As a nonparametric method for estimating the distribution of an estimator by resampling the data, bootstrapping is desirable because long–short returns are highly non-normal, exhibiting complex cross-sectional dependencies across fundamental signals, and because the performance evaluation of a large number of fundamental signals involves a multiple-comparison problem.
They demonstrate the generality of their approach by applying it to historical returns-based anomalies, the results of which are robust to various sampling schemes, weighting structures, and regression models.
Abstractor's Viewpoint
Using fundamental signals from financial statements may give rise to a mismatch between investor time horizons (intraday trading or weekly/monthly portfolio rebalancing) and the publication of results, which can lag by at least three months. The delay between the release of quarterly financial results and the actual financial quarters may also miss major turning points in the industry or economic cycle (e.g., oil-price drops or a cut in central bank policy rates), which may lead to a totally different direction in stock returns from that suggested by financial results.
It would have been interesting to see whether certain fundamental signals from financial results lose or gain importance over the 50-year sample period. Over time, changes in accounting publication standards could lead to more accounting disclosures, which might aid in discovering more accounting variables that correlate with future stock gains.