Forecasting the unemployment rate over districts with the use of distinct methods

Interdependencies among neighboring regions appear to be important in forming the shape of local labor markets. Nevertheless, only a few studies exist which have applied spatial models to forecast over small spatial units such as cities, districts or counties. The majority of predictions are developed with quarterly or yearly time series for a country or at regional level. The paper presents the above phenomena and deals with the problem of simultaneous forecasting of the unemployment rate over 35 poviats (districts and cities) in one of the Polish provinces. Two extremely different models with spatial dependencies were developed and estimated in this paper: the Spatial Vector Autoregressions (SpVAR) and the Spatial Artificial Neural Network (SpANN). The 13-month out-of-sample forecast is based on high frequency, raw, monthly panel data extracted from 31 local labor offices. The procedure worked out here allows comparing the forecasting performance of spatial models with their non-spatial and seasonal equivalents. The inclusion of a spatial component into the models significantly improves the accuracy of forecasts; however, the overall performance of SpVAR is 30% better than SpANN.