Bayesian Value-of-Information Update¶
Sequential inference model — instantiates Sequential Stopping Boundary Design
Recomputes the posterior and the expected value of the next observation after every signal, so continuation is judged against what one more look would actually change.
A Bayesian Value-of-Information Update is the belief-and-value engine that sits underneath a sequential stop decision: it carries an explicit probability distribution over the unknown, revises it with each new signal, and — its defining move — after every revision it computes the expected value of the next observation, meaning how much that observation would improve the eventual decision, not how much it would shrink uncertainty in the abstract. Two things separate it from the rest of the family. First, it prices a look by its power to change the choice you would otherwise make — an observation that would leave the decision unchanged is worthless no matter how much it narrows a confidence interval. Second, it does not itself pull the trigger: it produces a continuation-value number and hands it to a boundary, which is where the actual halt lives.
Example¶
An exploration company has struck a possible oil field and is drilling appraisal wells one at a time to estimate how much is really down there. Its geologists start from a prior — a distribution over recoverable reserves built from the seismic survey and analog fields nearby, centered around an unremarkable 80 million barrels but with a long tail either way. Each appraisal well is a signal: a dry-ish result pulls the posterior down, a strong core sample pulls it up and tightens it.
After the third well, the update does the thing that matters. It asks: given today's posterior, would a fourth well change what we do next — commit to a full development, farm out a stake, or walk away? It simulates the plausible results of well four, weights each by how likely the current posterior says it is, and finds that most outcomes leave the decision exactly where it stands: the field is already, in expectation, above the development threshold, and only a catastrophic fourth result would flip that — a result the posterior now rates unlikely. The expected value of drilling one more well, in decision terms, has fallen to roughly the cost of a rounding error, while the well itself costs about $12M and a drilling season. The update reports that number. It does not say "stop" — that verdict belongs to the boundary that reads it — but it has made the case that another look no longer buys a better decision.
How it works¶
- Start from a stated prior. An explicit distribution over the unknown quantity, built from base rates and analogs, so that later updates are disciplined rather than impressionistic.
- Update on each signal. Apply Bayes' rule to fold the newest observation into the posterior, using a likelihood model that encodes how reliable that kind of signal is.
- Value the next look, decision-first. Simulate the next observation's possible results under the current posterior and score each by whether it would flip the pending choice; the probability-weighted improvement is the continuation value.
- Emit the number, not the verdict. Pass the continuation-value estimate downstream; the update recomputes it every cycle so the signal a boundary reads is always current.
Tuning parameters¶
- Prior informativeness — how much weight the starting distribution carries. A strong prior stabilizes early updates but resists surprising data; a diffuse prior lets evidence dominate but swings wildly on thin samples.
- Likelihood / signal-reliability model — how much a given observation is trusted. Overstating reliability makes the posterior lurch; understating it makes every look feel necessary.
- Lookahead depth — whether continuation value is computed one observation ahead (myopic) or over the remaining sequence. Myopic is cheap and usually fine; it can undervalue a run of cheap looks that only jointly flip the decision.
- Decision-relevance filter — whether only belief changes that would cross the decision threshold count as value. Turning this up is what stops the mechanism from paying for precision no one will act on.
- Recompute cadence — update after every signal, or in batches. Per-signal is most responsive; batching saves computation when signals are cheap and frequent.
When it helps, and when it misleads¶
Its strength is that it dissolves the most common excuse for continuing — "we're still not certain" — by insisting that uncertainty is only worth reducing when the reduction would change an action. That reframing is the value of information: an observation is worth its cost only when the expected improvement in the decision exceeds it.[n1] Used honestly, the update kills the reflexive extra test and reveals when a decision is already effectively settled.
It misleads in two ways. The subtler one is the myopic-lookahead trap: a one-step value calculation can report near-zero for each individual cheap observation while a batch of them would decisively flip the decision, causing the mechanism to recommend stopping too soon. The blunter one is prior/likelihood misspecification — because the whole posterior rides on those two inputs, a quietly optimistic prior or an inflated reliability model manufactures confidence out of assumption, and the tidy number launders it. The guarding discipline is to stress the result against a deliberately skeptical prior and to compute value in decision-relevant terms rather than in raw variance reduction — and, where a cheap-batch flip is plausible, to look more than one observation ahead before trusting a "stop" reading.
How it implements the components¶
prior_distribution_or_base_rate— its starting state is an explicit distribution over the unknown, anchored on base rates and analogs, which every subsequent update revises rather than replaces.continuation_value_estimate— its core output: the decision-first expected value of the next observation, recomputed after each signal so the estimate never goes stale.observation_sequence_window— it models the incoming signals as an ordered stream, folding each into the posterior in arrival order and reflecting how reliability differs across signal types.
It stops short of drawing the halt line itself: the stopping_boundary that turns this value signal into an accept / continue / abandon action belongs to boundary-owning siblings such as Reservation Value Table and Real-Option Exercise Boundary; this update only supplies the continuation value they read.
Related¶
- Instantiates: Sequential Stopping Boundary Design — supplies the value-of-a-further-observation input the boundary compares against continuation cost.
- Sibling mechanisms: Bid Acceptance Cutoff · Real-Option Exercise Boundary · Reservation Value Table · Secretary-Problem Sampling Rule · Sequential Monitoring Stop Rule · Stop-Rule Postmortem
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Recomputes the posterior and the expected value of the next observation after every signal, so continuation is judged against what one more look would actually change, making its operative form a computation or analytic transformation that produces an inference, comparison, or optimized result.
Independent corroboration: The frozen evidence defines Bayesian Value-of-Information Update as 'Recomputes the posterior and the expected value of the next observation after every signal, so continuation is judged against what one more look would actually change', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Operations Research
Origin pattern: Convergent development
Present-day reach: Specialized
Rationale: Bayesian decision analysis computes expected value of sample information by how much another observation can improve the chosen action.
Related originating lineages:
- Economics & Finance — Expected-value reasoning prices the benefit of reduced uncertainty.
- Statistics & Experimental Design — Sequential inference repeatedly updates the posterior after observations.
Review resolution: Operations research is the agreed primary lineage through sequential decision analysis. Economics and Bayesian statistics are independently formative, so the method is best treated as convergent rather than an Encyclopedia synthesis.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
A close cousin under a neighboring archetype is Research Continuation Gate (Marginal Stop Rule), which also values the next unit of inquiry by its information. The difference is structural: that gate weighs a discrete research cycle against marginal cost with no shrinking horizon in play, whereas this update is keyed to an ordered observation stream and a closing window, and its output is a continuation value meant to be read against a horizon-sensitive boundary rather than applied as a gate on its own.
[n1] Value of information — the decision-analytic quantity capturing how much reducing uncertainty is worth before acting; a specifically Bayesian treatment (expected value of sample information) scores a prospective observation by the improvement it is expected to make to the decision, and an observation whose result would not change the chosen action has zero value however precise it is. ↩