Regional employment forecasts with spatial interdependencies

The labour-market policy-mix in Germany is increasingly being decided on a regional level. This requires additional knowledge about the regional development which (disaggregated) national forecasts cannot provide. Therefore, we separately forecast employment for the 176 German labour- market districts on a monthly basis. We first compare the prediction accuracy of standard time-series methods: autoregressive integrated moving averages (ARIMA), exponentially weighted moving averages (EWMA) and the structural-components approach (SC) in these small spatial units. Second, we augment the SC model by including autoregressive elements (SCAR) in order to incorporate the influence of former periods of the dependent variable on its current value. Due to the importance of spatial interdependencies in small labour-market units, we further augment the basic SC model by lagged values of neighbouring districts in a spatial dynamic panel (SCSAR). The prediction accuracies of the models are compared using the mean absolute percentage forecast error (MAPFE) for the simulated out-of-sample forecast for 2005. Our results show that the SCSAR is superior to the SCAR and basic SC model. ARIMA and EWMA models perform slightly better than SCSAR in many of the German labour-market districts. This reflects that these two moving-average models can better capture the trend reversal beginning in some regions at the end of 2004. All our models have a high forecast quality with an average MAPFE lower than 2.2 percent.