A comparison between machine learning and classic algorithms for GDP forecast
Tutor / directorGabarró Vallés, Joaquin
Document typeMaster thesis
Rights accessOpen Access
In the recent years there has been an explosive increase in the number of research papers using machine learning methods for forecasting. In this work, I will focus on comparing the estimation of GDP using classical and machine learning methods. In particular, I am interested in analyzing if the index S\&P 500 can help forecast the GDP, therefore I will take as a basis a work where a VAR model is used to analyze the relationship between GDP and S\&P 500. From there I use three recently developed implementations of the gradient boosting decision tree to model GDP and S\&P 500. The recent implementations are XGBoost, LightGBM and CatBoost and are very famous because they are widely used in winner solutions in Kaggle competitions. The metric I use to do the comparisson is the Mean Squared Error and by using the three machine learning algorithms, I find that they provide better results than the Vector Autoregressive. I also perform grid search with several parameters, taking into account regularization, to avoid overfitting and obtain the lowest mean squared error. To select the best model I consider the evaluation results along with the overfitting measure and finally I take a look to the prediction charts of all the algorithms. In my opinion XGBoost offers the best predictions in this exercise.