A Hidden Markov Model as a Dynamic Bayesian Classifier, With an Application to Forecasting Business-Cycle Turning Points

We introduce a method for dynamic classification of vector time series data into different regimes. A hidden Markov regime-switching model is used in classification. Past regimes are determined in advance and characterized by first and second moments of the observation vector. In estimation and model selection, instead of the maximum likelihood principle, we use Brier´s probability score making it possible to perform feature extraction, eg. noise-removing filtering. When calibrated to the forecast horizon, the method provides a simple and computationally efficient way to utilize leading information in forecasting regimes in time series. The method is applied on forecasting turning points of Sweden`s industrial production, where the Stock Market Index and a Business Tendency Survey series together express expectations, providing leading information. The method is also tested on forecasting the business cycle of the US, using GDP and the Department of Commerce Composite Index of Leading Indicators.