Early Warning Systems for Banking Supervision: Econometrics, AI and SupTech Masterclass

An early warning system in banking supervision is the process by which a supervisory authority turns the data it already holds about the banks it supervises into a timely signal that a specific institution is deteriorating. It combines inputs — prudential returns, on-site examination findings, payment behaviour, market prices where a listing exists — with a model that ranks or scores institutions, and with the supervisory framework that determines what may be done in response. It differs from a bank’s own early warning indicators, which serve the bank’s management and recovery planning: the supervisor is watching a whole population of institutions from outside, with less information but broader comparison, and must be able to justify any action taken on the basis of the signal.

They answer different questions for different people. A bank’s early warning indicator framework watches one balance sheet from inside, with full access to positions and intentions, and exists to trigger the bank’s own management and recovery actions before thresholds are breached. A supervisory early warning system watches an entire population of institutions from outside, with less granular and later information, and exists to allocate scarce supervisory attention and to justify intervention in a specific bank. The consequences differ too: a bank acting on its own indicator is managing itself, while a supervisor acting on a signal is exercising public authority over a private institution, which is why defensibility and proportionality matter far more on the supervisory side.

Machine learning can improve the ranking of institutions by risk, and it handles non-linearity and interaction between variables better than classical approaches. But bank failure is a rare event, and that changes what the methods can honestly deliver. A model that predicts survival for every institution will appear highly accurate in a population where almost every bank survives, which is why headline accuracy is a misleading measure here. Rare events also mean few examples to learn from, so models are prone to fitting the particular crises in the estimation window rather than the general pattern. The practical position taught on this programme is that machine learning is valuable for prioritising supervisory attention and surfacing candidates for closer examination, and is not a substitute for supervisory judgement or for a framework that can justify acting.

SupTech is the use of technology by supervisory authorities to carry out supervision itself — as distinct from RegTech, which is technology used by regulated firms to meet their obligations. In an early warning context it covers the pipeline that makes a signal possible: automated collection and validation of prudential returns, reconciliation and data quality controls, the scheduling that determines how fresh a signal can be, the analytical layer that scores institutions, and the dashboards and alerts through which supervisors receive it. The technology is rarely the hard part. The recurring difficulties are data quality at the point of collection, dashboards designed for the people who built them rather than the supervisors who must read them, and alerts frequent enough that they stop being read at all.

It is built for a mixed cohort from a single authority or from several: bank supervisors and off-site surveillance teams, financial stability and macroprudential departments, supervisory policy and enforcement staff, economists and modellers, statistics and regulatory reporting teams, and the SupTech, IT and business-intelligence staff who build the dashboards. The modelling sessions are pitched so that a supervisor who does not code can follow the reasoning and challenge the assumptions, and the regulatory sessions so that a modeller understands the constraints their output will meet. That is deliberate: an early warning system fails most often at the seams between these departments, and they rarely sit in the same room.

Both formats are complete: they cover the same nine modules and reach the same conclusion, including the Supervisory Signal Room simulation that closes the programme. The difference is tempo. The one-week format runs intensively across five full days and suits teams who need the capability quickly and can commit without interruption. The two-week format covers the same ground at a more deliberate pace, with longer working sessions, more case discussion, and time for peer exchange between supervisors from different authorities — which many cohorts value as much as the technical content. Both are delivered in English and French, in classroom and live online formats, and in-house for a single authority. The choice is one of pace and availability rather than depth of coverage.