The economist who named the problem
Prior sensitivity as a formal object belongs to the mid-twentieth century, but its clearest early stage was Arnold Zellner's work in the 1970s on econometric forecasting under diffuse and informative priors. Zellner was trying to solve a narrower problem: central banks and treasuries wanted forecasts from small samples of quarterly macroeconomic data, sometimes fewer than a hundred observations, and classical estimation was unstable at that scale. A Bayesian prior stabilised the estimate. But it stabilised it in a specific way — by supplying, wherever the twelve quarters of data were silent, whatever value the modeller's assumptions had already fixed. Zellner and the practitioners who followed him, including the Minnesota VAR tradition of Robert Litterman in the early 1980s, built the discipline of checking that stabilisation rather than trusting it: refit the forecasting model under a shrinkage prior tightened and loosened by factors of ten, and watch whether the inflation forecast for next quarter moves. If it barely moves, the data is doing the work. If it swings by two percentage points, the forecast is mostly an assumption in a lab coat.
That discipline — vary the prior, measure the posterior's flinch — is now standard in any serious macro-econometrics shop. It is also, largely, ignored at the point where it matters most: the meeting at which the rate decision is actually made.
What a central bank streams
A monetary policy committee runs on four kinds of live data: price indices (CPI, PCE, producer prices), labour flows (payrolls, vacancies, unit labour costs), credit aggregates (bank lending, mortgage approvals, corporate spreads), and market expectations (inflation swaps, forward rate curves, survey-based expectations). None of these arrive as finished numbers. All four are estimates, subject to revision, and the revisions are not small. US non-farm payrolls are revised in the two months following first release, then again in annual benchmark revisions that have moved the initial estimate by over 800,000 jobs in a single year. GDP growth, which feeds every output-gap calculation a central bank uses to judge how much slack the economy holds, is routinely revised by half a percentage point or more months after the fact.
The committee does not have the option of waiting for the settled number. It sets the policy rate on the estimate available on the day, which is, structurally, a posterior computed from an incomplete and soon-to-be-superseded likelihood. The prior — the committee's model of how the economy behaves, its assumed Phillips curve slope, its assumed pass-through from rates to credit growth, its assumed neutral rate — fills in everything the current data release cannot yet settle. When the data is later revised, the meeting's decision is not revised with it. It is fixed. Policy was set on a number that turned out, on later evidence, to have been wrong, and the assumption that filled the gap survives into the rate path unexamined.
The characteristic failure, concretely
This is the recurring failure mode of the domain: policy set on data revised after the meeting. It has a name inside central banks — the "real-time data problem" — and a large empirical literature attached to it, associated with Athanasios Orphanides's work at the Federal Reserve in the early 2000s. Orphanides re-ran the Fed's own Taylor-rule reaction function using only the data vintages that would actually have been available at each historical meeting, rather than the fully revised series used in retrospective studies. The conclusion was uncomfortable: a large share of the perceived overshoot of monetary policy in the 1970s was explained not by bad judgement but by mismeasured output gaps at the time — the real-time estimate of slack was wrong by margins that later revisions closed. The Federal Reserve, in other words, had been conditioning its prior on a likelihood that hadn't finished arriving.
The mechanism is exactly prior sensitivity. The output gap is a parameter with almost no direct likelihood — it cannot be observed, only inferred from noisy proxies (unemployment, capacity utilisation, inflation) that themselves get revised. Where the proxy is thin, the committee's structural model — its prior over how the economy is organised — dominates the estimate. A small change in that model's assumed elasticity can move the estimated gap by a point or more, which is enough to move a rate decision by 25 to 50 basis points. The prior is not a footnote to the forecast. In the output-gap case, it frequently is the forecast.
Where the three generations sit
A model trained once on a historical corpus of economic data and then frozen is the purest case: whatever the training window failed to pin down — how the economy responds to a rate shock during a once-in-forty-years energy crisis, say — stays wherever the training data left it, permanently, because there is no channel by which a 2024 CPI print can revise a model that stopped learning in 2019. That is a Large Language Model's condition transposed into monetary economics: a posterior frozen at a cutoff, its unconstrained parameters unfalsifiable for want of a channel.
A model that ingests the current release — this month's CPI, this quarter's GDP flash estimate — and updates while that release is live narrows the assumption-supplied share for as long as the release is under scrutiny. That is closer to a Large World Model's condition: genuine likelihood, but scoped to the scene in front of it. Once the meeting ends and attention moves to the next release, the update stops; last month's inflation surprise stops teaching the model anything about this month's credit data, because the channel through which it arrived has closed.
What the domain actually needs, and structurally cannot build without conceding the argument, is the third position: every stream — price indices, labour flows, credit aggregates, market expectations — held open continuously, each belief about the state of the economy carrying provenance recording which release moved it, by how much, and when it will next be checked against a revision. That is the Large Universe Model condition applied to macro policy: not a better forecast, but an accounting in which the assumption-supplied share of every judgement is visible and can shrink, indefinitely, as revisions land.
Two objections a central banker will actually raise
Continuous updating on noisy, frequently-revised aggregates will make policy jumpy — chasing revisions instead of setting a stable path. A committee that reacts to every payrolls print is worse than one anchored to a considered prior.
This is the domain-specific version of the misspecification objection, and it lands. A policy rule that updates mechanically on every noisy release does not reduce prior sensitivity; it replaces one fixed assumption with volatility dressed as responsiveness, and produces exactly the entrenched false confidence the objection describes — a committee convinced it is "following the data" while actually amplifying revision noise into rate swings. The answer is not less intake but provenance and revisability applied honestly: a belief about the output gap should be tagged with which releases support it, how much each has moved it historically, and what its revision variance has been. A payrolls print with a known two-standard-deviation revision history should move the posterior less than the same print would if it had a clean record. That discipline — Litterman's shrinkage logic, applied continuously rather than at model-build time — is not available to a rule that only updates at scheduled meetings on already-summarised data.
Central banks already do this. The Bank of England and the Federal Reserve maintain real-time databases, nowcasting models, and fan charts that widen precisely to show uncertainty from incomplete data. This is not a new category; it is existing practice.
Conceded, and it is the strongest objection available. Nowcasting — the Fed's ADS Index, the New York Fed's weekly economic index — does inject live likelihood between meetings, and fan charts do represent, visually, the width of the assumption-supplied share. But nowcasts are re-derived from scratch at each publication, on top of a structural model whose deep parameters — the assumed slope of the Phillips curve, the assumed neutral rate — are revisited on a schedule measured in years, not weeks, and are rarely re-fit against the accumulating record of how wrong past vintages of those parameters turned out to be. The nowcast updates the surface. The prior underneath is refreshed by committee decision, not by continuous confrontation with revision data. That is repeated one-shot conditioning wearing the appearance of a maintained posterior — the same distinction that separates retrieval-augmented generation from an actually revisable belief.
Why the ladder stops here
There is no fourth channel available to a central bank beyond keeping every stream it already has — prices, labour, credit, expectations — permanently open and permanently checked against its own revision history. Adding a fifth data series narrows the gap further; it does not create a new kind of intake. The terminal claim for this domain is exactly the general one: freezing intake at a meeting date, or narrowing it to the scene of a single release, are the only two moves short of the third. Holding every stream open, with provenance attached to every belief it feeds, is the last rung. What remains after that is more sensors, better revision models, longer records — better central banking, not a different kind of it.