Escalation of Forecasting Accuracy through Linear Combiners of Predictive Models

Precise and proficient modelling and forecasting financial time series has been paying attention of researchers, which leads to the development of various statistical and machine learning based models. Accuracy of a particular method is problem and domain specific, hence identifying best method is controversial. To boost up overall accuracies and minimizing risk of model selection, combination of outputs of different models has been recommended in the literature. This work presents a linear combiner of five predictive models i.e. ARIMA, RBFNN, MLP, SVM, and FLANN for improving prediction accuracy. Four statistical methods i.e. trimmed mean, simple average, median, and an error based method are used for suitable choice of combining weights. The individual forecasts and the linear combiner are used separately to predict closing price of five stock markets and exchange rate of five global markets. Extensive simulation work demonstrates the feasibility and supremacy of the linear combiner.