A1169
Title: Pairwise Markov chains for volatility forecasting
Authors: Elie Azeraf - Azeraf Financial Consulting (France) [presenting]
Abstract: The pairwise Markov Chain (PMC) is a probabilistic graphical model that extends the classical hidden Markov model (HMM). Despite its flexibility, the PMC has rarely been employed for continuous value prediction, mainly due to challenges in modeling observations within generative frameworks. The aim is to propose a new prediction algorithm for the PMC that overcomes these limitations. The approach (i) resolves the feature modeling problem, fully exploiting the PMCs expressive power, and (ii) provides a general mechanism to extend any predictive model with hidden states that evolve over time, thereby introducing non-stationarity in a principled manner. This methodology is applied to financial volatility forecasting, comparing it with standard benchmarks such as GARCH(1,1) and feedforward neural networks across multiple asset pairs. The empirical results highlight that, under regime changes commonly observed in volatility, the PMC-based extension consistently improves predictive accuracy, demonstrating its practical and theoretical value.
