Forecasting Unemployment via Machine Learning: The use of Average Windows Forecasts

Abstract

The aim of this study is to find a better approach to forecast unemployment rate, we apply the average windows forecasts (AveW) of Pesaran and Pick (2011) to both machine learning and standard ARMA-based time series approaches; moreover, to account for model uncertainties, we average all model-specific AveWs and show that the resulting forecasting combination further improves forecasting. Specifically, our performance comparison shows that: Firstly, for case without covariates, machine learning methods significantly outperform standard time series models, which is also confirmed by the Diebold-Mariano test. Secondly, for case with covariates, both ARMA-based and machine learning methods perform equally well. Finally, surprisingly, the well-known deep learning method, RNN-LSTM, does not have good performance as expected. Overall, AveW is a useful framework for time series forecasting via machine learning.

Keywords: Forecasting time series, forecasting averaging, machine learning, training by rolling k-fold cross validation

JEL Classification: C53