The authors’ research confirms the generally accepted phenomenon of alpha decay in quantitative equity strategies. The methodology can be summarized in three parts: univariate regressions looking at the predictability of characteristics in isolation, multiple regressions consisting of the entire universe of factors, and long–short hedge strategies based on sorting stocks into deciles and then trading the extremes.

For the entire 1980–2014 sample, 12 factors survive as independent predictors: asset growth; growth in industry-adjusted sales; growth in inventory; growth in book equity; growth in capital expenditure; growth in long-term net operating assets; growth in property, plant, and equipment (PP&E) plus inventory; number of consecutive quarters with earnings higher than the same quarter a year ago; growth in sales less growth in inventory; standardized unexpected quarterly earnings; percentage change in shares outstanding; and earnings announcement return.

Most strikingly, the authors identify a breakpoint from 2003 onward in which the returns to long–short hedge strategies that exclude micro-capitalization stocks are insignificantly different from zero. They attribute this finding to increased market efficiency because of reduced limits to arbitrage.

Starting with an initial set of 330 characteristics drawn from the anomalies/predictability literature, the authors reduce the set to 102 characteristics that can be implemented using the CRSP, Compustat, and I/B/E/S databases. Multicollinearity among the predictors is mitigated by further reducing the set to 94 broadly independent factors.

The authors control for the effects of micro caps, which tend to upwardly bias return statistics, by running three types of regressions. The first includes the entire universe but mitigates the micro-cap bias by placing more weight on large caps using value-weighted least-squares regressions. The second directly excludes micro caps, and the third looks at the entire universe using ordinary least-squares regressions.

By estimating the statistical significance (p-values) of regression coefficients that are adjusted upward to take into account the chance of falsely discovering anomalies in a large set of characteristics, the authors adjust for the effects of data-snooping bias in their regression results.

They also test the economic significance of their results by constructing long–short hedged strategies in which companies’ monthly returns are predicted and sorted into deciles, with the top decile forming a long portfolio and the bottom a short portfolio. Three strategies are considered: one in which all NYSE companies are included and the baskets are value weighted by market cap, an equal-weighted scheme, and a scheme that excludes micro caps.

The authors help to explain the current explosion in interest at quantitative trading firms for alternative datasets and techniques from data science, such as machine learning and artificial intelligence. The days of constructing profitable long–short equity strategies from a handful of company fundamentals appear to be long gone. It is interesting to note that the authors obtain the best practical results by including all 94 characteristics rather than first reducing the set down to the most statistically significant. This result supports a significant role for ensemble learning methods and clustering schemes, as well as increased attention on unstructured data and proprietary datasets.