Return Prediction and Stock Selection from Unidentified Historical Data

Abstract

The experimental approach is applied to explore the value of unidentified historical information in stock-return prediction. Return sequences were randomly drawn cross section and time from historical S&P500 data. Subjects were requested to predict returns or select stocks from 12 preceding realizations. The hypothesis that predictions are randomly assigned to historical sequences is rejected in permutation tests and prediction-errors decrease with expertise. The best-stock portfolios by experimental predictions significantly outperform worst-stock portfolios in joint examination of mean-return and volatility. Actual predictions are significantly more effective than various statistical rules in separating the “best” stock from the “worst” in random 6-stock menus.

Keywords: return forecasting, predictability, expertise, prediction regime

JEL Classification: D8, G1, C9