Noise impact on time-series forecasting using an intelligent pattern matching technique

Intelligent time-series forecasting is important in several applied domains. Artificially intelligent methods for forecasting are being consistently sought. The effect of noise on time-series prediction is important to quantify for accurate forecasting with these systems. Conventionally, noise is considered obstructive to accurate forecasting. In this paper, we analyse the noise impact on time-series forecasting using a pattern recognition technique for one-step ahead forecasting called the "Pattern Modelling and Recognition System". We evaluate the system performance on noise-filtered and noise-injected time series from four different sources: three benchmark series taken from the Santa Fe competition and the US financial index, S&P series. The results are discussed when comparing the proposed method against the established Exponential smoothing method and Neural networks and some important conclusions drawn on their basis. (C) 1999 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.