Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
You can tell that people are getting serious when their research has a yellow cover like one of those expensive mathematics textbooks:
And the team at the European Banking Authority is indeed getting serious — it’s just published a distinctly better than workmanlike piece on the possibility of using random forests, neural networks and similar techniques to potentially make a bit of progress towards one of the great dreams of central bank economists. That is to say, the possibility to automate the dreary and unprestigious job of bank supervision, using machines to monitor the data rather than going through supervisory returns yourself.
The frustrating thing about this exercise is that it always sort of works. As with indicators of financial stress (another favourite research project), there is some actual pattern and structure in the underlying data, and so the right use of statistical methodology will find it.
The EBA researchers actually do quite a bit better than previous efforts. “Breaches of supervisory concern levels on a few key ratios” are their better source of training data points than “actual failures”, meaning they can train a variety of models and use an ensemble approach.
Unfortunately, though, models like this are subject to the same fatal flaw as indicators of financial stress, which is that they’re always a finely polished rear-view mirror. When financial crises and bank failures actually happen, they tend to happen for reasons that can’t be predicted, because they’re not in the data set, because they’ve never happened before. Or at least, they’ve never happened before in exactly this way, so nobody was collecting the right data on them in the right way.
Even more annoyingly, we know it’s going to happen this way again. Financial supervisors keep telling us, with growing anxiety, that one consequence of the last round of banking regulations is that a load of business has moved into the “non-bank financial institutions” sector, where it isn’t regulated and no data is collected. While one might say “what exactly did you think would happen”, they have a point.
But wait! The data might be able to help with that too! In the EU, at least, every repo and derivatives transaction has to be reported and is stored in a massive data repository. Some more researchers at the ECB have found a way to link up this trade-by-trade data with the small amount of data they do have on hedge funds, and create something that might kind-of-sort-of serve as a measure of hedge fund leverage.
It’s possible to see a picture of the future here. Most financial things that you can do will generate at least some sort of data footprint. If that footprint is in some way legible to the regulators, then they might be able to arrange it into the sort of data set that can be observed to see if patterns of excessive leverage are growing up. And as AI and machine learning systems get better, they are going to get better at rearranging atoms of transactional data into meaningful structures — that’s what they’re for.
So the dream of electronic bank supervision is alive. Unfortunately, making the data footprint available to the supervisors is also what’s known as “regulatory burden”, and everyone is currently trying to minimise it. So human beings will have to do the job for a bit longer.
