Modelling with Big Data and Machine Learning: Interpretability and Model Uncertainty

The Bank of England and King's College London present a joint conference

Event dates: 4 - 5 November 2019

The confluence of access to large granular data sources (‘Big Data’) and the rapid advance of modelling techniques like those from machine learning (ML) promises new insights into the economy and a larger information set for policymakers. The Bank of England (BoE) and the Data Analytics for Finance and Macro (DAFM) Research Centre at King’s College London held a two-day workshop to discuss recent developments and, crucially, focus on particular aspects of Big Data and ML approaches which are of increased interest to applied researchers. Two such aspects formed the focus of this two-day conference. The first related to a commonly cited weakness of ML methods when applied to economic problems and data, which is lack of interpretability of ML model outputs. This makes the adoption of such models difficult for economists who wish to have a more structural understanding of the underlying economic issues. The second, and related, focus was on the estimation and/or calibration of the uncertainty associated with model outputs. Both these matters have not received as much attention in the mainstream ML literature as economists would like to.

The conference covered a range of topics including:

Large granular structured or unstructured data sources, e.g. administrative data, web data, from the “digital exhaust”, text data.
Machine Learning for prediction and understanding the economy and its interpretation.
Interpretability and uncertainty measurement of non-parametric methods, e.g. ML.
Data methods, e.g. matching, filtering or cleaning techniques.
Theory, e.g. estimation with many covariates or strong non-linearities, model and estimation uncertainty of ML approaches.

Keynote speakers included: