Improving Stock Market Prediction Through Linear Combiners of Predictive Models

Abstract

Accurate and efficient stock market forecasting model design has been attracted attentions of researchers continuously. This leads to development of various statistical and machine learning-based models in the context. Accuracy of a method is immensely problem and domain specific. Hence, identifying a best method, in general is controversial. Combining outputs of different forecasting models to enhance overall accuracies and minimizing the risk of model selection has been suggested extensively in literature. This work presents a linear combiner of five predictive models such as ARIMA, RBFNN, MLP, SVM, and FLANN towards improving the stock market predictive accuracy. Three statistical methods such as trimmed mean, simple average, and the median, and an error-based method are used for appropriate selection of combining weights. The individual forecasts and the linear combiner are employed separately to predict the next day’s closing price of five real stock markets. Extensive simulation work demonstrates the feasibility and superiority of the linear combiner vis-à-vis others.