Abstract

Machine learning algorithms dedicated to financial time series forecastinghave gained a lot of interest over the last few years. One difficulty lies inthe choice between several algorithms, as their estimation accuracy may beunstable through time. In this paper, we propose to apply an onlineaggregation-based forecasting model combining several machine learningtechniques to build a portfolio which dynamically adapts itself to marketconditions. We apply this aggregation technique to the construction of along-short-portfolio of individual stocks ranked on their financialcharacteristics and we demonstrate how aggregation outperforms singlealgorithms both in terms of performances and of stability.

 

Quick Read (beta)