What is a supervisory early warning system? Anatomy and the test it has to pass

Every authority holds more data on its banks than it can read. What turns that holding into a signal, the four parts every supervisory early warning system has, and why most fail on timing or on defensibility rather than on modelling.

In short

A supervisory early warning system is the process by which an authority turns the data it already holds about the banks it supervises into a timely, defensible signal that a specific institution is deteriorating.
It has four parts — inputs, a method that ranks or scores, thresholds that convert a score into a supervisory state, and a response ladder. The fourth is the one most often missing, and without it the first three change nothing.
Systems fail on two axes, and neither is modelling: a signal that arrives too late to act on, and a signal too weak to justify acting on.
Data quality at the point of collection sets the ceiling for everything built above it. Authorities that buy a model before fixing the returns pipeline usually discover the model was never the constraint.
The test is one question: name a supervisory decision in the last year that the system changed. A system that cannot answer it is a reporting exercise.

On this page

What a supervisory early warning system is
It is a process, not a model
The four parts
What the data can and cannot tell you
Where it goes wrong
What good looks like
What to do next

What a supervisory early warning system is#

A supervisory early warning system is the process by which an authority turns the data it already holds about the banks it supervises into a timely, defensible signal that a specific institution is deteriorating.

Three words in that sentence carry the weight. **Timely**, because a signal that arrives after the options have closed is a record rather than a warning. **Defensible**, because a supervisor who acts is exercising public authority over a private institution and must be able to say why. **Specific**, because knowing the system is fragile does not tell you which bank to examine on Monday.

It is a process, not a model#

The common error is to treat an early warning system as a model with a process attached. It is the reverse: a supervisory process with a model inside it. The model ranks; the process decides what the ranking means and who does what about it.

This distinction is not academic. It explains why authorities that procure a well-built model are often disappointed, and why authorities with modest methods sometimes supervise very effectively.

The four parts#

A complete system has four parts. The proportions vary with the authority; the parts do not.

**Inputs** — prudential returns, on-site examination findings, supervisory judgement recorded in a usable form, payment-system behaviour, and market prices where the institution is listed.
**A method** that combines those inputs into a ranking, a score or a set of flags, whether statistical, judgemental or both.
**Thresholds** that convert a score into a supervisory state, so that a number becomes a category with a meaning attached to it.
**A response ladder** stating what changes at each state — examination intensity, information requests, meetings at a named level, and the formal measures available.

The fourth part is the one most often missing, and its absence is decisive. Without it the system produces a ranking that is read, noted and filed, and the institution at the top of the list is supervised exactly as it was the month before.

A ranking that changes nobody’s week is a ranking, not a warning.

What the data can and cannot tell you#

Each input family turns at a different point in a bank’s decline, and knowing which turns first is more useful than adding more of them.

Prudential returns are comprehensive and comparable, and they are also periodic, revisable and prepared by the institution being assessed.
On-site findings are the richest evidence an authority holds and the least frequent, so they age between examinations.
Payment-system behaviour is close to real time and narrow — it sees liquidity strain well and capital erosion not at all.
Market prices, where a listing exists, move earliest and are noisiest, and they say as much about sentiment as about the institution.

Data quality at the point of collection sets the ceiling for everything above it. An authority whose returns arrive late, are revised heavily, or are compiled inconsistently across institutions has a data problem that no method will solve, and the sophistication of the model chosen makes no difference to it.

Where it goes wrong#

Systems fail along two axes, and neither of them is modelling.

**Too late to act on.** The signal is accurate and arrives when the remaining options are the expensive ones. Often the underlying indicator turned earlier and the reporting cycle absorbed the difference.
**Too weak to act on.** The signal is early and the authority cannot build a case from it, so nothing happens until something more concrete appears — by which time the signal was not the constraint.
**The alert nobody reads.** Thresholds set so that a large share of institutions are flagged in any given month, after which flags stop carrying information.
**The system that has never fired.** Read as reassurance, it is more often evidence that thresholds sit where the population already is.
**The orphaned score.** The method is owned by an analytics unit, the response ladder by a supervision department, and no one owns the join.

What good looks like#

A working system is recognisable from outside by four properties, none of which concerns the method.

Each supervisory state has a stated consequence, and someone can point to a month in which that consequence occurred.
The lag between the underlying condition and the signal is known and stated, rather than assumed to be zero.
Alerts are infrequent enough to be read and are recorded when dismissed, so a pattern of dismissal is visible.
One named function owns the whole chain from collection to supervisory response, and can be asked to explain any link in it.

What to do next#

Take the last twelve months and ask a single question of the system: name a supervisory decision it changed. An examination brought forward, an information request issued, a meeting called, an intensity classification altered.

If nothing comes to mind, the gap is almost never in the method. It is in the fourth part — the response ladder — and that is both the cheapest part to build and the part no vendor supplies.

Frequently asked

What is a supervisory early warning system?

A supervisory early warning system is the process by which a supervisory authority turns the data it already holds about the banks it supervises into a timely, defensible signal that a specific institution is deteriorating. It has four parts: inputs such as prudential returns, on-site examination findings, payment behaviour and market prices; a method that combines them into a ranking or score; thresholds converting that score into a supervisory state; and a response ladder stating what changes at each state. It is a supervisory process with a model inside it rather than a model with a process attached, which is why authorities with modest methods sometimes supervise very effectively and authorities with sophisticated ones are sometimes disappointed.

What data goes into a supervisory early warning system?

Four families, each turning at a different point in a bank’s decline. Prudential returns are comprehensive and comparable across institutions, but periodic, revisable and prepared by the institution being assessed. On-site examination findings are the richest evidence an authority holds and the least frequent, so they age between examinations. Payment-system behaviour is close to real time but narrow — it sees liquidity strain well and capital erosion not at all. Market prices, where a listing exists, move earliest and are noisiest, saying as much about sentiment as about the institution. Knowing which family turns first matters more than adding further inputs, and data quality at the point of collection sets the ceiling for everything built above it.

Why do supervisory early warning systems fail?

Along two axes, and neither is modelling. A signal can be too late to act on — accurate, but arriving when only the expensive options remain, often because the underlying indicator turned earlier and the reporting cycle absorbed the difference. Or it can be too weak to act on — early, but not something an authority can build a case from, so nothing happens until something more concrete appears. Three further patterns recur: thresholds set so loosely that flags stop carrying information, a system that has never fired being read as reassurance rather than as evidence the thresholds sit where the population already is, and an orphaned score owned by an analytics unit while the response ladder belongs to a supervision department, with nobody owning the join.

How do you tell whether a supervisory early warning system is working?

Name a supervisory decision in the last twelve months that the system changed — an examination brought forward, an information request issued, a meeting called at a named level, an intensity classification altered. If nothing comes to mind, the gap is almost never in the method; it is in the response ladder, the part that states what actually changes at each supervisory state. That is the cheapest part of the system to build and the part no vendor supplies. Three supporting checks: whether the lag between condition and signal is known rather than assumed to be zero, whether alerts are rare enough to be read, and whether dismissals are recorded so a pattern of dismissal becomes visible.

The programme behind this article

Work through this material with the practitioners who wrote it.