Neural networks for cross-sectional employment forecasts: a comparison of model specifications for germany

In this paper, we present a review of various computational experiments - and consequent results - concerning Neural Network (NN) models developed for regional employment forecasting. NNs are widely used in several fields because of their flexible specification structure. Their utilization in studying/predicting economic variables, such as employment or migration, is justified by the ability of NNs of learning from data, in other words, of finding functional relationships - by means of data - among the economic variables under analysis. A series of NN experiments is presented in the paper. Using two data sets on German NUTS 3 districts (326 and 113 labour market districts in the former West and East Germany, respectively), the results emerging from the implementation of various NN models - in order to forecast variations in full-time employment - are provided and discussed. In our approach, single forecasts are computed by the models for each distinct district. Different specifications of the NN models are first tested in terms of: (a) explanatory variables; and (b) NN structures. The average statistical results of simulated out-of-sample forecasts on different periods are summarized and commented on. In addition to variable and structure specification, the choice of NN learning parameters and internal functions is also critical to the success of NNs. Comprehensive testing of these parameters is, however, limited in the literature. A sensitivity analysis is therefore carried out and discussed, in order to evaluate different combinations of NN parameters. The paper concludes with methodological and empirical remarks, as well as with suggestions for future research.