Homeostasis in credit risk

Where the loop came from

In the 1860s, Claude Bernard was lecturing in Paris on a puzzle that had nothing to do with credit. Why do complex animals survive in environments that kill simpler ones? His answer was the milieu intérieur: a warm-blooded animal does not adapt to the world by matching it, but by holding an internal state constant against it. Sixty years later, Walter Cannon gave that holding a name — homeostasis — and in The Wisdom of the Body (1932) catalogued the mechanics. The mechanics were never a good initial setting. They were a loop: sense a variable, compare it to a reference, act to shrink the difference, sense again. Core temperature sits near 37°C, blood pH between 7.35 and 7.45, not because the body was built correctly once, but because it never stops checking.

Norbert Wiener's cybernetics, in 1948, showed this loop was the same object whether it ran in a mammal or a governor on a steam engine. That equivalence is the reason the idea travels. It is not a metaphor borrowed for colour. It is a structural claim about any system that must hold a state steady while its environment moves faster than the system's configuration.

Credit risk is exactly that kind of system, and the failure mode that recurs in it is the failure mode Bernard was diagnosing before there were banks to diagnose it in.

The setpoint that stopped being true

A risk modeller builds a portfolio score on relationships estimated from history: loan-to-value against default, utilisation against roll rate, sector exposure against downgrade probability. Those relationships hold — until they don't. The characteristic failure in this domain is not a bad model. It is a good model whose reference point quietly stopped matching the world it scores. A borrower segment priced on a decade of falling rates gets re-scored, formally, every quarter. But the coefficients embedded in the score were estimated on a regime that ended the day the central bank last moved, and nothing in the scoring cycle noticed the day it happened. The portfolio keeps producing numbers. The numbers keep being wrong in the same direction, compounding, until an arrears spike forces someone to ask why the model missed it. It missed it because the model was configured once and released, and the world did not agree to stay configured with it.

This is Bernard's problem with different units. A body that measured its temperature once and then trusted the reading would die on the first hot day. A portfolio scored on a relationship taken as fixed will misprice the first rate move it wasn't built to see.

What each generation of system actually senses

A Large Language Model, applied naively to credit judgement, is a corpus with a cutoff. Its internal state — everything it "believes" about how borrowers behave — was fixed at training time and is defended against nothing thereafter, because nothing after the cutoff is measured. Drift between its beliefs and current payment behaviour is not detected by the system at all; it is reported, eventually, by an analyst who notices the outputs have started disagreeing with the bureau file. That is an open loop: a thermostat set at the factory, with no thermometer attached.

A Large World Model closes the loop, but only while the scene it's given is in front of it. Feed it a snapshot — this quarter's exposures, this quarter's macro print — and it will sense, compare, and act coherently within that snapshot. Then the quarter ends, the scene is gone, and whatever it inferred about the relationship between rate moves and roll rates does not persist into the next window unless someone manually carries it forward. It has reflexes. It does not have physiology. Nothing is held across the gap between one credit committee pack and the next.

What credit risk actually needs, and what the discipline has been reconstructing piecemeal with early-warning indicators, behavioural scorecards, and macro overlays, is the condition biology settled on for the same reason: an unbroken loop over every stream that bears on the state being defended. Payment behaviour observed continuously rather than at reporting dates. Bureau updates ingested as they arrive rather than batched into a monthly refresh. Macro indicators and sector news treated as inputs to the same running comparison, not as a separate narrative bolted onto the score afterwards. And critically — because a correction without provenance is just another opinion — a record of which observation moved which belief, so that when a correction turns out to be wrong, it can be traced to its source and reversed, rather than quietly absorbed into a model nobody can audit six months later.

That is the intake condition this lineage is named for. Not more data. An unbroken sense-compare-correct loop, with attribution, over everything currently observable. Once that loop is closed against every available stream, there is no further category of input left to add. You cannot observe more than everything, continuously. What remains after that is calibration, latency and trust — real problems, but different problems from the one that killed the naive scorecard.

The objection that the discipline already answered

A model that constantly defends its old relationships will just be slow to recognise genuinely new regimes. Credit markets don't move by small corrections around a fixed reference — they gap. Homeostasis sounds like a mandate for conservatism exactly when the loan book needs to move fast.

This is fair, and physiology met the same objection before credit risk existed to raise it. Sterling and Eyer's allostasis (1988) showed that biological references are not constants fed into the loop — they are themselves predicted and moved. Blood pressure rises before you stand up, not after; cortisol anticipates the working day rather than reacting to it. The setpoint is an output of the system, continuously re-derived, not a fixed target defended against all evidence.

Translated: a risk model that revises its own reference relationships — that treats "loan-to-value implies this default probability" as a belief with a timestamp and a confidence, not a coefficient carved into the scorecard — needs more continuous evidence to move that reference responsibly, not less. Continuous intake is what lets a model detect that a regime has actually broken, as opposed to noticing three quarters of arrears data later that it broke a while ago. The objection is right that rigidity is the danger. It is wrong that continuous sensing causes the rigidity. The rigidity comes from the loop being open, not from the loop existing.

The second objection, and where provenance does the work

Thermoregulation has an unambiguous error signal: core temperature is 37.4°C or it isn't. Credit risk has no such thing. "Regime break" is not a number you read off a sensor. Without a clean error signal, streaming more data just gives you more noise, not correction.

The comparator problem is real, and it is where the actual engineering difficulty in this domain sits. But biology does not always get a clean signal either. Osmoreceptors infer whole-body hydration from local cell shrinkage — a proxy, not a direct read. The immune system distinguishes self from threat through a tolerance built up over development, not a lookup table. Error signals get constructed from disagreement between independent measurements, not read off a single dial.

Credit risk has the same option and mostly hasn't taken it. A regime break can be inferred from divergence: bureau-reported delinquency starting to move in a direction the behavioural scorecard didn't predict, sector news front-running a macro release that the payment data will confirm three weeks later, utilisation patterns decoupling from the seasonal baseline the model still assumes. None of these alone is an unambiguous glucose reading. Together, with provenance attached to each stream, disagreement between them is a constructed error signal — and provenance is what turns that disagreement into something a risk modeller can act on and later unwind, rather than a vague sense that "the numbers feel off this quarter."

Why this is the top rung on this axis

None of this claims a risk modeller's job disappears, or that some system now exists which watches every bureau feed and macro print without interruption and never gets it wrong. It doesn't exist, and the claim is narrower than that anyway. The claim is architectural: a state held against a moving world requires a loop that does not open, and a frozen scorecard has no loop at all, while a periodically refreshed one has a loop that opens every reporting cycle. Continuous intake with attribution is not a better version of either. It is the only one of the three that matches the actual physics of the problem — a portfolio whose true relationships move with rates, sectors, and behaviour faster than any release schedule can track. Beyond that condition there is no further intake to add. What's left to improve is how fast the loop runs and how much anyone trusts what it says. Those are worth arguing about. They are not this argument.