Skip to main content
  • This website sets cookies on your device. To find out more about how we use cookies please refer to our Privacy and Cookie Policy. By continuing to use the site, we’ll assume that you are content for us to set these on your device.
  • Close
Home > Research > Staff Working Paper No. 674: Machine learning at central banks - Chiranjit Chakraborty and Andreas Joseph
 

Staff Working Paper No. 674: Machine learning at central banks - Chiranjit Chakraborty and Andreas Joseph

01 September 2017

​Staff Working Paper No. 674: Machine learning at central banks
Chiranjit Chakraborty and Andreas Joseph

We introduce machine learning in the context of central banking and policy analyses. Our aim is to give an overview broad enough to allow the reader to place machine learning within the wider range of statistical modelling and computational analyses, and provide an idea of its scope and limitations. We review the underlying technical sources and the nascent literature applying machine learning to economic and policy problems. We present popular modelling approaches, such as artificial neural networks, tree-based models, support vector machines, recommender systems and different clustering techniques. Important concepts like the bias-variance trade-off, optimal model complexity, regularisation and cross-validation are discussed to enrich the econometrics toolbox in their own right. We present three case studies relevant to central bank policy, financial regulation and economic modelling more widely. First, we model the detection of alerts on the balance sheets of financial institutions in the context of banking supervision. Second, we perform a projection exercise for UK CPI inflation on a medium-term horizon of two years. Here, we introduce a simple training-testing framework for time series analyses. Third, we investigate the funding patterns of technology start-ups with the aim to detect potentially disruptive innovators in financial technology. Machine learning models generally outperform traditional modelling approaches in prediction tasks, while open research questions remain with regard to their causal inference properties.

The code and data used in “Case 2: UK CPI inflation forecasting” (Section 5.2) are available on GitHub.

Share