Macroeconomic Predictions Using Payments Data and Machine Learning

Monitoring and predicting the economy’s short-term dynamics are vital in economic decision making. However, major economic indicators are released with a substantial delay, and policy-makers must therefore rely on sophisticated models to accurately estimate them. 

Consumers are increasingly adopting electronic payment methods—a trend that accelerated dramatically during the COVID-19 pandemic. The vast amounts of high-frequency data generated by electronic payments are available almost in real time. And thanks to recent advances in artificial intelligence and machine learning, we have sophisticated econometric tools for analyzing non-traditional data and nonlinear relationships. In this paper we aim to show that payments data and machine learning models are useful in predicting short-term macroeconomic dynamics, such as nowcasting gross domestic product and retail and wholesale trade in Canada. We also address the challenges of interpretation and overfitting of machine learning models in order to improve their performance and our understanding of their predictions. 

We find that payments data that capture a variety of economic transactions are useful for estimating the state of the economy in real time. Moreover, the econometric tools provided by machine learning can capture large and nonlinear effects from a crisis. We also find that a few payment streams in Canada’s retail payment systems become significantly more important during periods of crisis, which substantially improves our model performance during those periods.