Fama–French three versus five, which model is better? A machine learning approach

Document Type

Article

Publication Title

Journal of Forecasting

Abstract

This research proposes new estimations of the Fama–French three- and five-factor models via a machine learning approach. Specifically, it uses a Bayesian optimization-support vector regression (BSVR) approach to obtain predictions of portfolio returns. On data from five industries' portfolio returns in the United States over the period July 1926 to January 2019, the BSVR models perform well. Specifically, our new model, called the Fama–French BSVR three-factor model, outperformed the Fama–French BSVR five-factor model. More precisely, the Fama–French BSVR three-factor estimations attain out-of-sample (testing dataset) correlation coefficients of 94% for portfolio returns for the consumption and manufacturing industries. A correlation of 92% between the predicted and experimental values of portfolio returns was found for the high-tech industry; 91% was found for the mining, construction, transportation, hotels, entertainment, and finance industries. However, for the Fama–French BSVR five-factor model, the correlation coefficients lie between 48% (health industry) and 89% (high-tech industry).

First Page

1461

Last Page

1475

DOI

10.1002/for.2970

Publication Date

2-14-2023

Keywords

asset pricing model, Bayesian optimization, factor model, machine learning, support vector regression

Comments

IR Deposit conditions:

OA version: Accepted version

24 months embargo

Publisher source must be acknowledged with citation

Must link to publisher version with set statement (see policy)

Link to Full Text

Share

COinS