Performance evaluation of statistical methods for forecasting of unemployment rate

The aim of this article is the accuracy evaluation of suitable models used for prediction of the unemployment rate development in Czech Republic under conditions of economic depression. Models were based on exponential smoothing and training of artificial neural networks. The most suitable models, as it was proved two months ago (see [1]), the exponential eventually damped model with additive seasonality and multilayer perceptron forecasted March’s and April’s unemployment rate as 7.58–7.89 % and 7.63–8.33 % for exponential smoothing respective as 7.3–7.45 % and 6.62–8.22 % for multilayer perceptron. Performance of a models measured by Theil’s U were 0.002–0.022 for exponential smoothing respective 0.016–0.041 for multilayer perceptrons. Recalculation of exponential smoothing model and retraining of artificial neural networks on fresh values of the unemployment rate show that 1) smoothing parameters were little modified; 2) same type of ANNs were suitable for solving this problem – comparison of nets parameters is considered useless. This recalculation/retraining brought new more relevant forecasting regarding to the present and still dynamic economic situation.