An On-Line Machine Learning Return Prediction

his paper presents a new prediction methodology for long-short portfolio return in its multiplicative version. Our method relies on the on-line universal portfolio construction. We derive a closed-form predicting formula whose coefficients are solely determined by historical data. We empirically and robustly demonstrate that the predictive error can be as small as 2\%. This methodology provides a promising application of on-line machine learning to portfolio return prediction with a sufficiently large dataset