Seeing liquidity stress coming

Recent bank failures have shown how fast liquidity stress can move. Deposits and market funding can leave in days, not weeks. By the time the standard metrics turn red, many of the options for a calm response have already gone.

Banks already watch early warning indicators, and supervisors have expected them to since the Basel Committee's 2008 principles for sound liquidity risk management. The hard part is turning a dashboard of signals into one clear call: is stress building now? My co-author Murat Donduran and I asked whether machine learning could make that call earlier and more reliably, from a bank's own point of view rather than a policymaker's.

What we did

We used daily, publicly available market data from January 2007 to January 2021, 3,537 days covering both the global financial crisis and the COVID-19 shock. The indicators were deliberately simple: the VIX, crude oil and Nasdaq 100 futures, the Euronext 100 index, the TED spread and US commercial paper spreads. For most of them we tracked daily moves, seven-day volatility and intraday range.

To teach the models what stress looks like, we used the St. Louis Fed Financial Stress Index. The calmest 85% or so of days were labelled green, the next 10% amber and the most stressed 5% red, the familiar RAG status a bank's management already works with. We then trained 24 machine learning models to predict the next day's status, with cross-validation to keep them honest.

What we found

On plain accuracy, several models looked good, because most days are calm and predicting "green" is easy. That is the wrong test. A bank needs to be warned on the amber and red days. On that measure, an ensemble model built on the RUSBoost algorithm, designed for rare events, clearly came out on top.

It flagged 95% of amber and red days, against an average of 74% for the best of the other model types. It caught 83% of red days and 84% of amber days, and it never mistook a red day for a green one. The price was some false alarms on calm days, which is the right trade-off for an early warning system.

Why it matters

Earlier warning buys management time: time to assess the situation, test options and plan a response before markets force the issue. A model like this does not replace judgement, and its thresholds need regular back-testing. But it can turn many signals into one daily status that can be automated and fed into management information.

The study used public, US-centric market data only. The real prize is for a bank to combine this approach with its own daily data, such as liquidity positions, deposit outflows and collateral calls, which public research cannot see. Supervisors, with data across many banks, could go further still.