Bayesian Updating
TL;DR
Bayesian updating is belief revision with bookkeeping: state what you believed before the evidence arrived, weigh each new item by how strongly it discriminates between the hypotheses — not by how vivid it is — and move belief by that much, in either direction. As a graph, the prior is a recorded artifact with a source, each evidence item carries its diagnosticity, and the calibration gate polices the two classic corruptions: double-counting the same fact twice because it was phrased two ways, and belief that only ever moves one direction. The trace that falls out is the thing regulated monitoring always lacks: why this risk score, decomposed into which evidence moved it and by how much.
The suspicion that only ever went up
The enhanced-review file has been open for four months, and the risk rating has climbed every month. A wire to a new counterparty: up. A quiet month: up — dormancy before activity is a known typology. The counterparty turned out to be a school district: up, somehow — layering through legitimate entities. Reading the file end to end, the analyst’s successor notices what nobody inside the case could see: there is no observation in the world that would have moved this rating down.
She starts over with two numbers. First: of accounts that look like this one at onboarding, how many turn out to be illicit? The base rate is small, and it is the starting point, written down before any of the file’s color. Then, item by item: does this fact appear meaningfully more often in illicit accounts than in clean ones? The wire to a school district does not. The quiet month barely. Two facts genuinely discriminate, and the file’s conclusion now rests on those two — visible, weighable, arguable — instead of on four months of accumulated adjectives.
An unguided model narrates evidence into whichever story is already running — vividness reads as weight. The Bayesian graph makes every item pay for its influence in diagnosticity, and shows the belief moving both directions or explains why it didn’t.
How the graph works, step by step
Retrieve the prior, with its source. Before any case detail is admitted, the graph records the starting belief and where it comes from — a portfolio base rate, a segment statistic, a documented judgment. This is the strategy’s anchor against the oldest failure in risk work: reasoning as if the interesting case were the typical one.
Admit evidence one item at a time. Each fact enters individually, deduplicated at the door: the same underlying observation phrased three ways in three reports is one item, not three. Provenance rides along, because an item’s weight may be revisited when its source is.
Evaluate diagnosticity. The discipline node. For each item: how much more likely is this observation if the hypothesis is true than if it is false? Facts that are common in both worlds — however alarming they sound — carry little weight and are recorded carrying little weight. This is where vividness stops being a substitute for evidence.
Update, loop, and gate. Belief moves in proportion to each item’s weight, in whichever direction the item points, and the running trajectory is part of the trace. The calibration gate then audits the whole run: every item counted once, movements proportional to stated weights, and — the scene’s test — if every update pointed the same way, the gate demands the recorded reason why no exculpatory evidence was found or credited.
The gates, operationally
| Gate | What it checks | Fails when |
|---|---|---|
| Prior gate | The starting belief is stated with a source before case evidence is admitted | The prior is back-filled to fit the conclusion, or the base rate never enters at all |
| Diagnosticity gate (per item) | Each item’s weight reflects how well it discriminates between hypotheses, and each underlying fact is counted once | Vivid-but-common facts move belief, or one observation is credited under three phrasings |
| Calibration gate | Belief moved proportionally and bidirectionally across the run, or the one-way trajectory is explicitly justified | The rating ratchets — every item confirms — and nobody can name what would have moved it down |
Where it fits — three use cases
1. Ongoing monitoring and periodic review
KYC refresh and enhanced due diligence are belief-revision problems by definition: a rating exists, new facts arrive, the rating should move — both directions. A Bayesian trace turns “why is this customer high-risk?” from a narrative into a decomposition: this prior, these items, these weights. It also surfaces de-risking’s mirror failure: ratings that never come down.
2. Alert and case risk scoring
Layered detection produces overlapping signals — the same underlying behavior tripping three rules. Item-at-a-time admission with deduplication is precisely the fix for scores inflated by triple-counted evidence, and the per-item weights give model validation something concrete to challenge.
3. Early-warning credit review
A covenant wobble, a late financial statement, an industry downgrade — each discriminates between “temporary noise” and “deterioration” with different force. Priced updates produce watch-list decisions with visible arithmetic, and the trajectory record shows exactly when the file should have moved — the question every post-mortem asks.
When to reach for it
Reach for Bayesian updating when belief must track evidence over time — monitoring, scoring, watch-lists — and especially where the failure mode is the ratchet: suspicion that only rises, ratings that never fall. Its honest limits are the two inputs: priors and weights are judgments, and the graph does not pretend otherwise — it records them, which is what makes them challengeable. That is the point: not certainty, but belief whose every movement has a receipt.
- Prior as artifact
- The starting belief, sourced and timestamped before evidence arrives — the anchor that keeps the interesting case from being treated as the typical one.
- Diagnosticity
- An item’s real evidential weight: how much more likely it is under one hypothesis than the other. Vividness is not weight.
- The ratchet
- Belief that only moves one direction — the signature of confirmation absorbed into process, and the specific corruption the calibration gate exists to catch.