A Bayesian Approach to Adapting Forecasts to Structural Changes in a Simple State-Space Model

Most forecasting models often fail to produce appropriate forecasts because they are built on the assumption that data is being generated from only one stochastic process. However, in many real world problems, the time series data are generated from one stochastic process initially and then abruptly undergo certain structural changes. In this paper, we assume that the basic underlying process is the simple state-space model with random level and deterministic drift, but is interrupted by three types of exogenous shocks; level shift, drift change, and outlier. A Bayesian procedure to detect, estimate, and adapt to the structural changes is developed and compared to simple, double, and adaptive exponential smoothing using simulated data and the U.S. leading composite index.