Machine Learning: Superior to Traditional Statistical Models in Forecasting Macroeconomic Time-Series?

Abstract

This thesis investigates machine learning's potential to forecast the Norwegian GDP, unemployment rate, and in ation on monthly or quarterly, and annual terms. We compare machine learning techniques such as penalised regressions and random forest to traditional statistical methods such as the na ve model, autoregressive and vector autoregressive models. This motivates the following thesis question, Is value added by machine learning compared to traditional statistical models in time{series forecasting of macroe- conomic variables? The results show that the machine learning models are relatively better than the traditional statistical models when forecasting except for in ation. Using many exogenous variables to explain in ation is more confusing than value{adding, therefore, the models depending only on in ation itself provide the best forecasts.