Deep neural-network prediction for study of informational efficiency

Abstract

In this paper, we attempt to verify a hypothesis of informational efficiency of financial markets, known as “random walk” introduced by Fama. Such hypotheses could be considered in relation to financial crises. In our study the hypothesis is tested on data taken from Warsaw Stock Exchange in 2007–2009 years. The hypothesis is tested by predictive modelling based on Machine Learning (ML). We compare conventional ML techniques and the proposed “deep” neural-network structures grown by Group Method of Data Handling (GMDH). In our experiments a GMDH-type neural-network model has outperformed the conventional ML techniques, which is important for achieving the reliable results of predictive modelling and testing the hypothesis. GMDH-type modelling does not require the knowledge of network structure, as a desired network of near-optimal connectivity is learnt from the data. The experimental results compared in terms of prediction error show that the GMDH-type prediction model has a significantly smaller error than the conventional autoregressive and neural-network models.