A Multi-step Ahead Dyadic Particle Filter for Price Prediction

A very challenging task in financial forecasting is the accurate prediction at multiple future time steps based on historical prices. In this paper, a multi-step ahead dyadic particle filter is proposed for stock price prediction. A dyad corresponds to a latent vector modeling each stock and a latent vector modeling the group of companies in the same category. The stock latent vectors evolve through time according to a geometric Brownian drift parameter, which enables stock latent vectors to capture the erratic price evolution appearing in future time steps. The multi-step ahead framework of the dyadic particle filter and the drift parameter lead to more accurate predictions compared to state-of-the-art techniques.