Passive Aggressive Portfolio Selection with Neural Network Ensembles

In this paper, we design a new portfolio selection system named Passive Aggressive Neural Network Ensembles (PANNE). PANNE integrates an Adaboost improved BP Neural Network price pattern recognition strategy with Passive Aggressive Online Learning algorithm. In order to make the intelligent system more robust, we consider an autocorrelation filter that can be applied to initial stock screening. Within the algorithmic system, we propose 8 strategies in total to elaborate the predictive information from various facets. From intensive data experiments, the autocorrelation filtered PANNE strategy exceeds all other strategies in terms of its annual return that is more than 30% over 15 years and Sharpe Ratio 1.32. Other than this, three other strategies yield over 20% annually and all with Sharpe ratio higher than 1.0. In addition, left tail scenario and sensitivities over different variates have been thoroughly analyzed to provide better insights and interpretations regarding the strategies’ performance and persistence. As further discussed, this new comprehensive and robust PANNE portfolio selection system can be incorporated with various prevailing financial anomalies making it a highly attractive and promising tool for industrial application.