Analysis of the predictability of TOPIX returns using neural networks

Abstract

This study presents the results of various neural networks for predicting the TOPIX (Tokyo Security Exchange Stock Prices Index) monthly changes with macroeconomic time-series data as inputs. Stock-market prediction is a complex problem given the poor understanding of the underlying mechanisms governing the system's evolution. Studies which have been conducted assume that the observations of stock-market prices are produced by an underlying low-dimensional attractor with characteristics of a nonlinear dynamical system. Following the analysis of the nonlinearity and complexity of the TOPIX data, different neural network models for predicting the TOPIX monthly changes have been constructed. The networks are able to predict better than chance by using single-step prediction. In addition, sensitivity analysis of relevant inputs is conducted for different time-periods to estimate the relationship between different inputs and the output. The results of sensitivity analysis may help to simplify neural network models and infer relationships between the observed time series.