SARIMA vs Deep Learning Hybrid Models for Italian Unemployment Forecasting

In this study we develop and compare forecasting models for the monthly unemployment rate in Italy—aggregate and gender-specific—over the period 2004–2025. We employ classical SARIMA models as well as hybrid deep learning approaches that couple SARIMA with LSTM and GRU networks to capture nonlinear structures in the data-generating process. The aim is to assess whether neural network models significantly outperform the traditional Box-Jenkins approach. Forecast accuracy is evaluated using the standard error metrics in both in-sample and out-of-sample settings. In-sample results show slight improvements from hybrid models for women and the global rate whereas the SARIMA approach provides the best performance for men. Out-of-sample forecasts over a 12-month-ahead horizon indicate that SARIMA achieves the lowest forecast errors for men and global unemployment rates.