Time series forecasting with local linear forests
Date and time
March 17, 2022
09:00 - 10:00
I propose to combine the local linear forest (LLF) from Friedberg et al. (2020), which is a tree-based machine learning method, with the Shapley Additive explanation (SHAP) method from Lundberg and Lee (2017) to forecast time series. I designed a series of Monte Carlo experiments with different GDP specifications to compare the predictions generated by the local linear forest with predictions from the traditional implementation of the random forest. In simulated iid data, the LLF predictions are superior than RF predictions in the presence of only linear signals. As an empirical application, I nowcast the Dutch GDP, comparing forecasts of the LLF with random forests and eight other benchmarks. RF nowcasts and forecasts for 1 and 2 quarters ahead have lower error than competing models when averaging all sample periods. LLF predictions had lower forecast error in the period of the European debt crisis.