A novel MCDM ensemble approach of designing an ELM based predictor for stock index price forecasting

Abstract

In support of its faster learning capacity and better generalization, Extreme Learning Machine (ELM) has gained the attention of researchers as a means to solve various real-world prediction problems. However, the performance of ELM is heavily dependent on the activation functions used in it. In this study, design of ELM is addressed as a Multi Criteria Decision Making (MCDM) problem. The selection of the activation function for the ELM based predictor model is done through a novel MCDM ensemble approach. On the basis of 9 prediction metrics, MCDM techniques such as TOPSIS, PROMETHEE-II, and VIKOR were used to assess and rank 15 activation functions on ELM performance. In light of the fact that the ranks determined by each MCDM technique do not coincide, a novel ensemble approach was proposed to calculate the final rank score by considering the occurrences of each model in the primary ranking and its respective rank score. In the end, the most highly ranked activation function is taken into account in the ELM-based predictor model. The proposed model is assessed over three benchmark stock indices such as BSE SENSEX, NIFTY 50 and BSE S&P 500. The empirical analysis clearly shows that the ELM based predictor model designed using ELU activation function performs competitively compared to other reported models.