Machine learning algorithms dedicated to financial time series forecasting have gained a lot of interest over the last few years. One difficulty lies in the choice between several algorithms, as their estimation accuracy may be unstable over time. Aggregation combines a finite set of forecasting models, called experts, without making assumptions about the models and dynamically adapts to market conditions. We apply expert aggregation to the construction of long-short strategies, built from the individual stock return forecasts. The online mixture outperforms individual algorithms in terms of both portfolio performance and stability. Extensions to both expert and aggregation specializations are also proposed and improve the overall mixture on portfolio evaluation metrics.
View on arXiv