Sparse Ensemble Matters: Evidence From Unemployment Rate Forecasting

Abstract

Sparse ensemble forecasting has become an increasingly competitive technique for forecasting research and practice in recent years. This paper examines the role of sparse ensemble in unemployment rates forecasting using expert forecasters. First, we show how the effectiveness of sparse ensembles is influenced by the complexity and accuracy of the base models. Second, we extend sparse regularization techniques to settings with unknown bias and variance employing Monte Carlo simulations. Third, we highlight the critical role of the regularization coefficient $\lambda$, which serves as a key shrinkage factor and necessitates a balance between model sparsity and forecasting accuracy. Experimental results on unemployment rate data demonstrate the superiority of sparse ensemble learning over equal-weight strategies. This framework provides novel insights into predictive modeling within the fields of economics and labor markets.