Methods and apparatus for use in predicting non-stationary time-series data

摘要

A computer-implemented data prediction method carries out a prediction cycle in which a neural network outputs a predicted value of non-stationary time-series data at a next time step, where values of the data are known for and up to the current time step. In each cycle the inputs for a neural network are selected dynamically by performing an input selection process. This process comprises determining, for a set of candidate indicators, the causality between values of the data and values of each indicator using a causality measure, such as Granger causality, and selecting a subset of indicators which yield the highest causal correlation with the data, to be the inputs for the neural network at that time-step. The primary application disclosed is predicting next day direction of a stock price.