Application of online multitask learning based on least squares support vector regression in the financial market

Abstract

As is known, the financial market prediction and high investing value is receiving more increasing attentions nowadays. But affected by many complex factors, it is difficult to perform the financial market forecast accurately. Among the solving methods, the time-series prediction has caused the focus for its great predictive effect in many fields. However, most of the existing works focus on single-time-series analysis and cannot obtain good learning results because it trains tasks independently and ignores the cross-correlation among multiple time series. Motivated by the multitask learning, a novel online multitask learning based on the least squares support vector regression (OMTL-LS-SVR) algorithm is proposed for multi-step-ahead financial time-series prediction. OMTL-LS-SVR regards multiple related time series as different learning tasks, which are trained in parallel to obtain the prediction model and shorten the training time. Under this scheme, the knowledge from one certain task can benefit others, allowing it to exploit the relatedness among multiple subtasks. The OMTL-LS-SVR is applied to perform the time-series tendency prediction in four branches of China’s financial market, and the experimental results demonstrate the effectiveness of the proposed multitask learning algorithm.