Forecasting BIST100 Index with Neural Network Ensembles

Abstract

This paper aims to provide a neural network-based approach to forecast the direction of movement of BIST 100 stock price index and investigates the difficulties of such an implementation. It is observed that a neural network implementation is highly sensitive to selection of features and optimization parameters such as learning rate. A methodology to overcome the difficulties of neural network implementations to financial time series is proposed in the paper. Several feature selection methods are employed to obtain a subset of the features that can be used in the training of any classification algorithm. The difficulties and benefits of using an ensemble of neural networks instead of a single neural network are also studied. Results have shown that the use of neural network ensembles yields promising results. Keywords: Neural Networks, Ensemble, Bagging, Forecast.