Probabilistic Data Analysis in Non-Stationary Processes Forecasting

Abstract

A modified approach to the Bayesian data analysis is proposed and an original methodology for identifying and accounting the possible uncertainties that exist in modelling and predicting non-stationary processes problems is developed. A new methodology for the combined model implementation is considered. It includes the following elements: the Bayesian data processing method, the optimal data processing filter, regression models for the statistical data formally describing and predicting the conditional variance, and the probabilistic Bayesian network model for predicting non-stationary processes. The conducted computational experiments for the financial market indicators forecasting showed that the quality of short-term volatility forecasts is significantly improved due to the optimal filtering procedure and the right choice of model structure.