Predicting GDP Growth using the U.S. Transportation Sector: A Machine Learning Approach

Abstract

The post-COVID world has made very clear the importance of supply chains and the transportation of goods on overall economic growth. In this paper, a model forecasting real US GDP growth using the performance of the transportation sector is developed to capitalize on this relationship. A machine learning approach is taken where models are trained on the quarterly percent changes in transportation stock prices to predict GDP growth rates from the year 2007 to 2019. Various data preprocessing steps, including feature scaling and selection, were used before feeding the data into nine machine learning models: from traditional models such as ridge regression, SVM, and decision trees, to more complex models such as ensemble random forests and artificial neural networks. Every model was hyperparameter tuned using five-fold cross-validation. Model predictions were evaluated using mean absolute error and root mean squared error. A baseline model that consistently predicts the average GDP growth rate was used as a performance benchmark and scored a MAE of 1.95 on the test set.

For models trained on the whole sector dataset, the results show that the deep neural network model outperformed the baseline model, with a MAE of 1.73. Additional models were trained on industry-based subsets of the data for which the study found the airline, trucking, and delivery industries alone perform on par with the baseline model and thus, do not provide additional information for models to extract. The railroad industry models and the models for the other smaller industries (marine, rental, and transportation services) were found to perform better than the baseline model, signifying these industries provide additional information models can extract. The study concludes that the financial performance of the transportation sector does have the ability to act as a predictor of GDP growth, with certain industries having more predictive power than others, but further research is needed to create more predictive models using more extensive datasets.