Improving the Performance of Multivariate Forecasting Models through Feature Engineering: A South African Unemployment Rate Forecasting Case Study

The ability of machine learning models to forecast unemployment rates better than traditional statistical methods has been well established in literature. The ambition of researchers, in this field, over the last decade has been to demonstrate that machine learning models are able to forecast unemployment rates as well as or better than traditional statistical methods. Feature engineering has thus far been applied to a limited extent when forecasting unemployment rates. Especially when compared to feature selection and feature extraction. This research leverages feature engineering to demonstrate that such techniques could improve the performance of the models. The application of feature engineering on multivariate data to forecast the South African unemployment rate increased the R-squared by over 100% on average and decreased the mean absolute scaled error by ~1%. Demonstrating that such techniques are of value when forecasting multivariate data.