Feng, Giglio, and Xiu propose a model-selection method to systematically evaluate whether any new factor contributes to asset pricing beyond a high-dimensional set of existing factors. Unlike standard approaches that assume perfect variable selection, their method accounts for model-selection mistakes that would otherwise create omitted-variable bias. Applying it to a large library of factors, they find most recently-discovered factors are redundant relative to existing ones, while a few have statistically significant explanatory power beyond the hundreds proposed in the past.
"We propose a model selection method to systematically evaluate the contribution to asset pricing of any new factor, above and beyond what a high-dimensional set of existing factors explains. Our methodology accounts for model selection mistakes that produce a bias due to omitted variables, unlike standard approaches that assume perfect variable selection."
The right tool arriving at the right mess: the factor literature had become a data-snooping hall of mirrors, and this paper imports the one econometric idea that fixes it — double-selection makes inference on a target coefficient robust to imperfect control selection, exactly the guarantee single-LASSO lacks. Pairing it with Fama-MacBeth keeps the economics recognizable. The result — most factors redundant, a few real — is both a methodological contribution and a referee's cudgel: "significant alpha versus a three-factor benchmark" no longer clears the bar. Its dependence on the assembled factor library and LASSO tuning is the obvious caveat, and it sits squarely in the lasso/high-dimensional-inference cluster as the finance application of post-selection inference.