Prediction of Unemployment Rates with Time Series and Machine Learning Techniques

In this paper, are explored and analyzed time series and machine learning models for prediction of unemployment in several countries (Med, Baltic, Balkan, Nordic, Benelux) for different forecasting horizons. FARIMA is a suitable model when long memory exists in a time series and has been applied successfully for predicting unemployment. To overcome the potential issue of heteroskedasticity, we explore whether FARIMA models with GARCH errors achieve more accurate results. To further improve forecasting accuracy, we consider models with non-normal errors. The above models however cannot take into account the non-linearity of the data and due to this fact, we employ three machine learning techniques to forecast unemployment rates, i.e. fully connected feed forward neural networks, support vector regression and multivariate adaptive regression splines. ARIMA and Holt-Winters are considered as benchmark models. Finally, the effects of different forecasting horizons and different geographic locations in terms of forecasting accuracy of the models are explored.