Kolmapäev, 8. mai 2019 kell 11.00-12.00 J. Liivi 4, ruum 103
Luca Alfieri (Tartu Ülikool)
Forecasting industrial production with boosting in a Big “Fat” Dataset
Recent literature is increasingly interested to the possibilities offered by machine learning models to economic forecast in presence of a large number of covariates. The paper aims to forecast with different Machine Learning techniques, especially boosting, the US Industrial Production Index of the FRED-MD dataset. The FRED-MD dataset is continuously updated and it is composed by more than 130 monthly U.S. macroeconomic time series from January 1959. The different models are coming from two different groups of Machine Learning methods: “the LASSO family” and the Ensembles. The work compares the forecast errors of the estimated models with two classical benchmark models such as autoregressive model (AR) and random walk (RW). The rolling fixed and expanding window pseudo out-of-sample forecasts show that that all the estimated models beat the AR and RW and the LASSO family models performed better than ensemble methods. In the long-horizon, instead, not all the models beat the RW. In the LASSO family group, Ridge cannot beat the RW. Instead, in the ensemble group random forest cannot beat the RW but Boosting beats the RW in expanding window estimations and performs better than random forest in the long-horizon in both fixed and expanding window pseudo out-of-sample forecasts.
Keywords: Macroeconomic Large datasets, Forecasting, Machine Learning
JEL Classification: C22, C53, C55.