Application of Sliding Window Technique for Prediction of Financial Time Series Using Time Delay Neural Networks

This paper proposes application of sliding window technique to time-delay neural network (TDNN) for prediction of financial time series. Neural network is a data-driven approach, in which we have huge data samples but limited information about the model structure. In this paper, we measure performance of the prediction and apply sliding window technique to select the most favorable neural network structure, time-delay taps and the most desirable training data size that result in the best prediction performance. The method was evaluated by using real data of share price of four firms traded in London Stock Exchange. The results show remarkable decrease for the root mean squared error, mean absolute percentage error and the linear regression of TDNN output offset.