Reversion Forecasting For Online Portfolio Selection

Focused on the issue that the non-stationarity of time series and the inefficiency of single model forecasting are not fully considered in the existing reversion strategies,two new multi-period portfolio selection strategies are proposed,i.e."Online Autoregressive moving average reversion"(OLAR)strategy and "Online Combination Forecasting Reversion"(CFR)strategy.For the OLAR strategy,firstly the ARIMA model was simplified,and then a closed update of the parameters was design within the framework of online learning,and finally the online passive aggressive algorithm was used to adjust the portfolio.For the CFR strategy,based on the same online settings,FOCFR and SOCFR algorithms based on first-order online learning and second-order online learning were designed respectively,and the online passive aggressive algorithm is also used to adjust the portfolio.More importantly,here show the regret upper bounds of these two CFR algorithms.The sub-linear regret bounds indicate that the algorithms are well generalized.Finally,a large number of experiments were conducted to evaluate the performance of proposed strategies in real market.Empirical results show that OLAR and CFR can effectively overcome the drawbacks of the existing strategies and obtain consistent and optimal performance.