Forecasting using twinned principal curves and twinned self-organising maps

Abstract

We extend the principal curves algorithm by creating twinned principal curves which extend through two related data sets simultaneously. The criteria for accepting a pair of data points as neighbours for any other pair of data points is that each of the relevant points must be close in the appropriate space. We illustrate the algorithm's predictive power on artificial data sets before using it to predict on a real financial time series. We compare the error from this twinning with that achieved by a related algorithm which twins self-organising maps.