Combining the real-time wavelet denoising and long-short-term-memory neural network for predicting stock indexes

Abstract

A stock market index can be a valuable indicator to describe the performance of a stock market in a particular region. Nevertheless, it is very difficult to forecast its future values or trends since the index data often demonstrate a high degree of fluctuations. Intrinsically, the wavelet denoising is a useful method to separate the signals from noise in many practical multi-media applications while the long-short-term-memory neural network (LSTM) is a powerful recurrent neutral network (RNN) architecture of learning and prediction models used in the field of computational intelligence. Nevertheless, few research studies have ever considered their combination for the prediction of stock or index movement. More importantly, some existing proposals trying to combine wavelet denoising with other artificial neural network architectures suffer from two major drawbacks. First, when applying the conventional one-time wavelet transform for denoising the stock data, this approach has made a serious logical flaw to include future stock data in its training phase, thus leading to impressive results in the backtesting yet actually impractical in real-world applications. In addition, the wavelet functions and decomposition levels are typically fixed in those studies for which they will not be able to produce optimal results in terms of the prediction accuracies. Hence, we propose in this paper a novel model to combine real-time wavelet denoising functions with the LSTM to predict the East Asian stock indexes in which the wavelet denoising adopts a sliding window mechanism to exclude the future data while its system configuration is flexibly optimized based on some predefined criteria. The empirical results reveal that the performance of our proposed prediction model shows significant improvements when compared to those of the original LSTM model without utilizing the wavelet denoising function. Furthermore, there are many interesting and possible directions including the integration with other deep learning networks for the future investigation of this work.