Stock Price Ranking by Learning Pairwise Preferences

Financial time series forecasting is a challenging task in machine learning. The noisy and non-stationary nature of financial time series requires efficient support vector machine solvers. In this study, we consider the problem of predicting the future ranks of stocks as a multi-class classification task. We design a pairwise to multi-class classification scheme based on an effective online SVM solver. We treat ranks of daily stock prices as class labels and use pairwise combinations of stocks rather than individual stocks. To assess the proposed approach, we consider an example based on the prices of three stocks from Borsa Istanbul. The experimental results reveal that the algorithm successfully predicts the ranks of stocks from different sectors. We also present a comparative analysis with commonly used classifiers. The results indicate that the proposed approach exhibits better classification performance than other classifiers for all stocks, especially in near future prediction.