Long term regional forecasting with spatial equation systems

Long-term predictions with a system of dynamic panel models can have tricky properties since the time dimension in regional (cross) sectional models is usually short. This paper describes the possible approaches to make long-term-ahead forecast based on a dynamic panel set, where the dependent variable is a cross-sectional vector of growth rates. Since the variance of the forecasts will depend on number of updating steps, we compare the forecasts behavior of a aggregated and a disaggregated updating procedure. The cross section of the panel data can be modeled by a spatial AR (SAR) or Durbin model, including heteroscedasticity. Since the forecasts are non-linear functions of the model parameters we show what MCMC based approach will produce the best results. We demonstrate the approach by a example where we have to predict 20 years ahead of regional growth in 99 Austrian regions in a space-time dependent system of equations.