Uncovering a factor-based expected return conditioning structure with Regression Trees jointly for many stocks

TL;DR

This paper explores how regression trees can uncover non-linear relationships in factor-based models for stock returns, revealing the dominant role of market excess return and analyzing tree structure properties.

Contribution

It introduces a regression tree approach to analyze non-linear dependencies in factor models for stock returns, highlighting the informational content of factors and tree structure insights.

Findings

Market excess return is the most informative factor.

Depth=1 trees often balance due to return distribution properties.

High skew does not necessarily cause tree imbalance.

Abstract

Given the success and almost universal acceptance of the simple linear regression three-factor model, it is interesting to analyze the informational content of the three factors in explaining stock returns when the analysis is allowed to consider non-linear dependencies between factors and stock returns. In order to better understand factor-based conditioning information with respect to expected stock returns within a regression tree setting, the analysis of stock returns is demonstrated using daily stock return data for 5 major US corporations. The first finding is that in all cases (solo and joint) the most informative factor is always the market excess return factor. Further, three major issues are discussed: a) the balance of a depth=1 tree as it relates to properties of the stock return distribution, b) the mechanism behind depth=1 tree balance in a joint regression tree and c) the dominant stock in a joint regression tree. It is shown that high skew values alone cannot explain the imbalance of the resulting tree split as stocks with pronounced skew may produce balanced tree splits.