Improving Robustness of a SVR Based Algorithm Trading Model with Carefully Crafted Features and a Diversified Portfolio

Algorithm trading is to use computer programs to automate trading. Our goal is to improve the robustness of a SVR based algorithm trading model for short term trades. Robustness of a trading model means that the profit curve of the model has low volatility without sudden sizable ups and downs. (1) Firstly, we carefully craft some features for prediction of stock price in short term future, based on commonly used indicators to capture typical price movement patterns, overbought/oversold, and divergence situations. The features are normalized before being used to train the SVR based trading model to generalize the model to all stocks. (2) Secondly, we design a portfolio diversifying method based on the trading model. Correlations between stocks in a portfolio can compromise profitability of an algorithm trading model. Prices of stocks with strong correlations move in the same direction. If the trading model predicts price movement in the wrong direction, stop loss will be triggered and these stocks will cause loss in a mutual accelerated manner. We propose a method to improve diversity of the portfolio. Stocks are clustered into different groups using a similarity measure based on historical performances of the trading model on the stocks, and the portfolio is built by selecting stocks from different groups. (3) Experimental results show that our trading model can earn an excess return compared to risk free fixed savings. And the volatility of the profit curve of the trading model is reduced, which means that robustness of the trading model is improved.