Artificial Neural Networks or Regression Modelling: Does it Matter?

Abstract

This study explores the ability of artificial neural networks (ANNs) to digest anomalies from factor models and investigates whether ANN models play a similar role in dissecting anomaly returns between developed and developing markets. The sample includes stocks from both the United States (U.S.) and Chinese markets, covering the period from 1995 to 2021. Neural network and traditional regression models are constructed using in-sample data, and their predictive performance is evaluated on out-of-sample data. Six well-known asset-pricing factors are selected as input variables, and the long-short spreads of nine anomaly strategies are the single output variable. The results show that the proposed ANN models outperform traditional regression models in holdout performance, regardless of the number of input factors or market type. Furthermore, ANN models in both the U.S. and Chinese markets demonstrate the same incremental accuracy performance. This provides evidence in support of using ANNs for financial modelling.