Adaptive forecasting in dynamic markets: An evaluation of AutoTS within the M6 competition

Abstract

In contemporary forecasting, the challenges of navigating the intricacies of erratic human-induced patterns combine with the challenges of navigating the overwhelming number of methods and models available to manage these data. The M6 Competition, which emphasized repeated, real-time monthly forecasting of stock markets, featured many of these difficulties. Here, AutoTS, an open-source Python package designed specifically for probabilistic time series predictions, is evaluated within the context of this competition. AutoTS includes an extensive repertoire of models, augmented by robust data preprocessing utilities, and employs genetic algorithms to fine-tune model parameters, contingent upon user-delineated evaluation metrics. This study describes the deployment of AutoTS in the M6 Competition, which won the investment decision challenge, and outlines the model selection pipeline and the process of converting forecasts into decisions which produced this result. Although a single definitive model remains elusive, these findings underscore the potential value of methodologies that are dynamic and largely autonomous.