Industrial Output Growth Forecast: A Machine Learning Approach Based on Cross-Validation

Abstract

Our work aims to evaluate two strategies to forecast industrial output growth, one of the important macroeconomic indicators for a country. Automatic algorithms that require few or no human interventions and are neutral generate good forecast performance and can be helpful for policymakers. We propose an automatic algorithm that is based on cross-validating different models and finding an optimal combination of them as soon as a new information is available. We evaluate two strategies. The first strategy aims to improve the Granger and Bates method by using cross-validation to better estimate the weights that are used in forecast combinations. The second strategy selects the best model to be used for forecasting based on the Cv technique (CvML). We apply our strategies to the industrial output growth rate data of six countries: Brazil, Germany, the United States, France, Japan, and the United Kingdom. Both cited methods were applied on “pseudo” streaming data, with observations feeding the model one by one, being re-estimated after each step, with an automatic selection/combination of models. Our results show that CvML outperforms all other benchmark models in most cases, especially in the long run. Even when CvML is not the best performing model, it has the same statistical performance as the best one.