Predicting the relative performance among financial assets: A comparative analysis of different approaches

We perform a comparative analysis of a wide array of approaches for the problem of forecasting the relative performance among different tradable assets in the framework of the M6 competition. To produce the forecasts, we employ various models spanning probabilistic, classification, and time-series methods, each approaching the problem from a different perspective. We demonstrate that in the case of financial forecasting, simple machine learning approaches have better performance compared to more complex deep-learning models. Furthermore, approaching the problem as a classification task appears to be beneficial. We also confirm findings from existing literature that using simple ensemble techniques can improve performance, and that forecasting performance is better for exchange-traded funds and assets that have lower idiosyncratic volatility. Finally, we benchmark our results against the performance of teams that participated in the M6 competition.