01 October 2018 12.00pm - 2.00pm
Machine Learning and Economics
Speaker: Adam Giles, Financial Conduct Authority
Venue: Ashurst, Broadwalk House, 5 Appold Street, London, EC2A 2AG
The Financial Conduct Authority’s Adam Giles gave a fascinating lunchtime Masterclass at the end of September 2018 about the use of machine learning techniques in economics.
Adam distinguished between three types of machine learning - supervised (using data to predict outcomes), unsupervised (used to find structure in data, including text) and reinforcement learning (such as learning to win a game).
The prime focus of Adam’s talk was on supervised learning. While many of the methods used today are not new, it is the increased availability of data and computing power which makes them so effective today. These models focus on out of sample prediction errors rather than the relationships between variables or the asymptotic properties of the model (which is generally the case in econometrics).
Adam provided an overview of some core methods used in machine learning: “penalised linear models”, and “regression trees”. Adam went on to describe how these models are optimised for prediction tasks. In particular a process of splitting the sample data to mimic out of sample performance, and then tuning “regularisation” parameters to find the right balance between fitting the true relationships in the data and not the noise in the sample. Adam also provided an overview of more complex methods that use these methods as building blocks–for example, “random forests” are a collection of regression trees fit on bootstrapped samples of the original data, which substantially improves prediction performance.
Adam’s talk then discussed how many issues in policymaking require some form of prediction. Machine learning is a powerful tool in these circumstances, but suffers from many of the same issues that arise in more standard econometric modelling. For example picking up biases from the raw data, or failing to account for structural changes. However, machine learning methods are more prone to issues with the fairness of the predictions they make. Consider a car insurance company that uses vehicle colour as a proxy for driver aggression to predict risk and price insurance–to the extent that different genders may tend to prefer different car colours, the predictions may unintentionally discriminate between genders.
Finally Adam’s talk covered how machine learning methods, especially “LASSO”, a type of penalised linear model, can be used for more standard econometric problems where the causal relationships between variables and outcomes are of interest, rather than prediction alone.
In conclusion, Adam notes that machine learning techniques in economics are, in some ways, not dissimilar to those in standard econometrics - with similar pitfalls too.
George Buckley, SPE Council