Machine Learning May not Outperform ARMAX in Forecasting Stock Returns

Stock returns predictability has been a long-standing topic in the literature on financial economics. Our study investigates whether stock returns predictability can be improved by applying machine learning methods to the predictive model of Chen et al. (1986), by examining the monthly return of Dow Jones Industrial Average from 1959M3 to 2024M12. With the aid of average window forecasts (AveW) of Pesaran and Pick (2011) as a framework of model selection, evidence from Diebold and Mariano test (1995, 2015) indicates that the best model of machine learning cannot outperform that of the ARMAXs; moreover, the forecast combination shows that the average of all ARMAXs outperform that of all machine learning methods. As it is well known that unaccounted serial correlation enlarges the variance estimates, hence it implies that ARMAXs may account for serial correlation better than machine learning, since machine learning depends on deterministic AR terms.