Learning to Trade with Incremental Support Vector Regression Experts

Support vector regression (SVR) is an established non-linear regression technique that has been successfully applied to a variety of predictive problems arising in computational finance, such as forecasting asset returns and volatilities. In real-time applications with streaming data two major issues that need particular care are the inefficiency of batch-mode learning, and the arduous task of training the learning machine in presence of non-stationary behavior. We tackle these issues in the context of algorithmic trading, where sequential decisions need to be made quickly as new data points arrive, and where the data generating process may change continuously with time. We propose a master algorithm that evolves a pool of on-line SVR experts and learns to trade by dynamically weighting the experts’ opinions. We report on risk-adjusted returns generated by the hybrid algorithm for two large exchange-traded funds, the iShare S&P 500 and Don Jones EuroStoxx 50.