Evaluation and comparison of machine learning and classical econometric AR model on financial time series data

Abstract

This paper examines the effects of time series data behaviour on the predictive performance of classical econometric univariate autoregressive and machine learning autoregressive models. The research aims to understand which forecasting approach would perform better in extreme scenarios. Even though some empirical studies demonstrate the superiority of machine learning methods relative to classical econometric methods, it is still arguable under what conditions one method can be constantly better than the other. And if there are any cases when econometric models are preferable than machine learning. Data is derived from simulation, ensuring the presence of different outlier and error distributions in small and relatively larger samples. The simulation results show that the machine learning approach outperforms econometric models in most of the cases. However, the existence of outliers worsens the performance of machine learning on small datasets. Even with the presence of outliers, as the sample size grows, the result improves so much for machine learning that it dominates the econometric model. The same models were used to forecast with rolling sample approaches on real financial data.