Online Self-reorganizing Neuro-fuzzy Reasoning in Interval-Forecasting for Financial Time-Series

Abstract

“The only thing constant is change.”—Ray Kroc (Founder of McDonald’s). Self-organizing neuro-fuzzy machines are maturing in their online learning process for time-invariant conditions. To, however, maximize the operative value of these self-organizing approaches for online-reasoning, such self-sustaining mechanisms must embed capabilities that aid the reorganizing of knowledge structures in real-time dynamic environments. Also, neuro-fuzzy machines are well-regarded as approximate reasoning tools because of their strong tolerance to imprecision and handling of uncertainty. Recently, Tan and Quek (2010) discussed an online self-reorganizing neuro-fuzzy approach called SeroFAM for financial time-series forecasting. The approach is based on the BCM theory of neurological learning via metaplasticity principles (Bienenstock et al., 1982), which addresses the stability limitations imposed by the monotonic behavior in Hebbian theory for online learning (Rochester et al., 1956). In this paper, we examine an adapted version called iSeroFAM for interval-forecasting of financial time-series that follows a computational efficient approach adapted from Lalla et al. (2008) and Carlsson and Fuller (2001). An experimental proof-of-concept is presented for interval-forecasting of 80 years of Dow Jones Industrial Average Index, and the preliminary findings are encouraging.