Comparing Unsupervised Layers in Neural Networks for Financial Time Series Prediction

Abstract

In this study, we propose and compare neural network models that use unsupervised layers for the prediction of financial time series. We compare the novel FL-RBM and FL-SMIA-RMB models that integrate a Restricted Boltzmann Machine (RBM) and the self-organizing layer of the Selforganized Multi-Layer Network using the Immune Algorithm (SMIA) with the FL-SMIA network and a standard MLP. We aim to investigate the performance of unsupervised learning in comparison to purely supervised and other mixed models. The FL-RBM model combines the products of raw input features (the Functional Link, FL), with the Restricted Boltzmann Machine RBM as a self-organizing first hidden layer, while the FL-SMIA model uses the Immune Algorithm on the first layer. The FLSMIA- RBM model, combines both self-organizing layers with a back-propagation network. The results show that the FL-SMIA model outperforms the FL-RBM, the FL-SMIA-RBM and the MLP as measured by Annualized Return (AR) in one-day-ahead prediction on exchange rates time series. In terms of volatility, the FL-SMIA and MLP perform similarly.