Predicting Alpha: A Pedagogical Survey of the State of the Art in Return-Forecasting Signals
Abstract
An alpha signal is any piece of information, available today, that helps predict the part of an asset's future return that is not simply compensation for bearing known risks. Hunting for such signals is the central activity of quantitative investing — and one of the most treacherous, because the same machinery that discovers a true signal will, if used carelessly, manufacture hundreds of false ones. This didactic survey, written for a reader with a quantitative background but no specialist training in empirical asset pricing, explains what the academic literature has learned about predicting alpha. It builds the definition of an alpha signal from first principles (separating a predictor from the risk-adjusted economic edge it may contain), maps where signals have historically come from (the cross-section of firm characteristics, time-series premium predictability, microstructure, and volatility dynamics), surveys the modeling spectrum from linear regression to machine learning, and devotes its core to how to measure whether a candidate signal is real: out-of-sample accuracy and proper scoring, economic significance net of costs, multiplicity-aware significance against data snooping, and distribution-free uncertainty quantification. It closes with the open frontier: non-stationarity, post-publication alpha decay, the high-frequency signal-to-noise ceiling, and the integration of machine learning with rigorous, multiplicity-aware evaluation.
Keywords
JEL codes
Download PDFCite
@techreport{shehadi2026alpha,
author = {Shehadi Candela, Agust\'in},
title = {Predicting Alpha: A Pedagogical Survey of the State of the Art in Return-Forecasting Signals},
institution = {QUAFI Research},
type = {QUAFI Working Paper},
number = {2026-01},
year = {2026}
}Preliminary working paper; circulated for discussion. The views are the author's own. Not investment advice.