Improved time series prediction with a new method for selection of model parameters

TLDR: In this article, a new method for model selection in prediction of time series is proposed, which takes into account the sharp changes in a time series and improves the generalization capability of the KPCR model for better prediction of the unseen test data.

Abstract: A new method for model selection in prediction of time series is proposed. Apart from the conventional criterion of minimizing RMS error, the method also minimizes the error on the distribution of singularities, evaluated through the local Holder estimates and its probability density spectrum. Predictions of two simulated and one real time series have been done using kernel principal component regression (KPCR) and model parameters of KPCR have been selected employing the proposed as well as the conventional method. Results obtained demonstrate that the proposed method takes into account the sharp changes in a time series and improves the generalization capability of the KPCR model for better prediction of the unseen test data.