Single-stage portfolio optimization with automated machine learning for M6

The goal of the M6 forecasting competition was to shed light on the efficient market hypothesis by evaluating the forecasting abilities of participants and performance of their investment strategies. In this paper, we challenge the ‘estimate-then-optimize’ approach with one that directly optimizes portfolio weights from data. We frame portfolio selection as a constrained penalized regression problem. We present a data-driven approach that automatically performs model selection and hyperparameter tuning to maximize the objective without noisy or potentially misspecified intermediate steps. Finally, we show how the portfolio weights can be optimized using the Method of Moving Asymptotes. Testing on the M6 competition data, our approach achieves a global rate of return of 9.5% and an information ratio of 5.045, which is in stark contrast to the mean IR of the M6 competition teams of −3.421 and the IR of 0.453 for the M6 benchmark.