Predicting individual stock returns using optimized neural networks

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Business | Master's thesis

Date

2019

Major/Subject

Mcode

Degree programme

Finance

Language

en

Pages

47 + 10

Series

Abstract

I investigate individual monthly U.S. stock return predictability through a comparative study on neural networks and ordinary least squares benchmarks, using a predictor set of 102 lagged firm characteristics and the market return from 1980 to 2018. I find monthly out-of-sample (OOS) R2 of 0.80% for the best neural network, confirming similar findings of marginal predictability from existing literature applying machine learning to empirical finance. OOS R2 increases to 7.12% for the best neural network, when considering average market return predictability using market return predictions constructed bottom-up from equal-weighting individual stock predictions. I also find significant monthly four-factor alphas of 1.55% and annualized Sharpe ratios of 2.62 on long-short top-bottom decile portfolios sorted on predicted returns – not taking into account trading costs. Investigating variable importances within neural networks reveals that networks using Rectifier as their activation function focus on momentum and liquidity variables, similar to existing findings, but networks using Maxout focus on firm fundamentals and risk measures instead – a new observation for the anomalies literature. Lastly, my findings confirm that return anomalies are stronger in small stocks, and prediction performance is generally stronger during market turbulence.

Description

Thesis advisor

Lof, Matthijs

Keywords

koneoppiminen, neuroverkot, osaketuotot, ennustaminen

Other note

Citation