A dynamic dyadic particle filter for price prediction

Abstract

Financial forecasting deals with dynamically evolving phenomena through time. Accurate price prediction based on historical data is a challenging task in this field, because a large degree of uncertainty governs price evolution. Stock and flight price prediction are two cases studied here. A dynamic dyadic particle filter is proposed that is based on sequential importance resampling. To ensure the generation of efficient particles, a time-evolving structure is constructed by employing pairs of latent state vectors called dyads. The variance of the observation noise is treated as a random variable obeying a heavy-tailed distribution. For stock price prediction, the dyad consists of a latent state vector modeling each stock and a latent state vector modeling the group of companies of the same category. It is demonstrated by experiments that the dynamic evolution of latent state vectors leads to more efficient prediction compared to state-of-the-art techniques. For flight price prediction, the dyad consists of two latent state vectors, corresponding to route and arrival, respectively. Promising flight price predictions are disclosed given the wide price range in each route.