Machine Learning Classification Methods and Portfolio Allocation: An Alternative Probabilistic Perspective

Abstract

We frame the asset pricing problem as a machine learning classification problem. The predictions on 3.34 million observations lead to zero-investment portfolios yielding significant out-of-sample economic gains. Through directly measured accuracies, binomial tests suggest that the classifiers can extract forward-looking contents from historical information. The classifiers exploit the differences in return state transition uncertainties. As reflected by the pre-realization variances of multi-class predicted probabilities, the classifiers are more confident subjectively when predicting high-trading-friction stocks. Consistently, only trading frictions contribute to out-of-sample predictability throughout 26,302 distinct stocks' lifetimes. The adjustment of the classifiers' obsession over certain return states increases the performance.

Keywords: Artificial neural network, big data, binomial test, classification, dropout additive regression tree, gradient boosting machine, information theory, machine learning, portfolio allocation, out-of-sample prediction, random forest, return state transition

JEL Classification: C14, C38, C55, G11, G14