A new approach of integrating piecewise linear representation and weighted support vector machine for forecasting stock turning points

Financial data forecasting is one of the most important areas in financial markets. In the stock market, if the stock falls or rises to a point and then rises or falls for a long time, these points are turning points (TPs). Everyone wants to buy or sell stocks at the TP to maximize profits. This paper integrates the piecewise linear representation (PLR) and the weighted support vector machine (WSVM) to forecast stock TPs and proposes several methods to enhance the performance of the PLR-WSVM model. Firstly, a fitness function is proposed to select the threshold of the PLR automatically. Secondly, an oversampling method suitable for the problem of forecasting stock TPs is proposed. The random undersampling combined with the oversampling is used to balance the number of samples. Thirdly, the relative strength index (RSI) is integrated to determine whether the predicted TP is a buying point or selling point. Twenty stocks are used to test the proposed model. The experimental results show that the proposed model significantly outperforms other models. The coefficient of variation of the revenues obtained by the proposed model is the lowest, indicating the proposed model is the most stable. (C) 2019 Elsevier B.V. All rights reserved.