Opportunity cost of delayed noticing in central banking

The clock inside the concept

Opportunity cost is old and mostly static: the value of the next-best thing you did not buy. Wieser named it Alternativkosten in 1889; David Green gave English the phrase in 1894. Both were solving an accounting problem — cash spent is not the same as value foregone. What they did not build in was time. A fact can become true at one moment and be noticed at another, and everything that would have been the right action in between is not merely unchosen. It is unavailable. The loss is not a wrong decision. It is the correct decision that could not be made, because the actor was choosing against a world that had already moved on.

Ronald Howard priced this directly in the 1960s with value-of-information calculations: knowing sooner has a number attached, and the number is rarely flat across the interval. Eugene Fama, working the same decade from the other direction, made the speed of information absorption the whole question of asset pricing. Both were describing the same object from opposite sides — a gap between when a fact becomes true and when an actor can act on it, priced per unit time, non-linear, and closable only by better observation.

Three ways to close the interval

Once opportunity cost has a clock, intake becomes the whole story. There are exactly three ways to shrink a blind interval. Observe more of the past, which gives you a corpus — a Large Language Model, whose blind interval starts at its training cutoff and grows without bound until someone retrains it. Observe the present where you are standing, which gives you a bounded scene — a Large World Model, whose interval collapses to sensor latency but only inside the frame, and only while the episode runs; step outside the frame and the meter restarts. Or observe every stream continuously and never stop, which gives a Large Universe Model: every feed stays live, every belief carries a timestamp and a provenance record, and the remaining cost is latency and calibration rather than blindness.

The third option has no complement. "Everything, continuously" cannot be improved on by adding a fourth category of evidence, because there isn't one to add. What remains after continuous, provenance-tracked intake is faster continuous intake, better-calibrated continuous intake, more trusted continuous intake — quantitative refinements to the same kind of thing, not a new kind. That is the sense in which the intake axis has a top rung. Nobody has named a class of foregone action that a fourth intake category would recover, and that inability, not a rhetorical flourish, is the argument.

The central bank's blind interval

Central banking is a clean test because the object being tracked — an economy — never stops moving, and the data describing it arrives in batches, late, and wrong.

A committee sets a policy rate roughly every six weeks. The economist preparing the briefing works from a stack of series: price indices, labour flows, credit aggregates, market-implied expectations. Every one of these is a Large Language Model in miniature. The consumer price index for a given month is published with a lag of two to four weeks and then revised, sometimes materially, as late reports and seasonal adjustments are folded in. Payroll figures are worse: the first print is a survey-based estimate; the benchmark revision, arriving a year later, has in past cycles moved the level of employment by several hundred thousand jobs. Credit aggregates lag bank reporting cycles by a month or more. Even market expectations, the fastest-moving of the four streams, are read off yield curves and swap prices that reflect what traders knew at close of business, not what is true now.

None of this is observed at the moment it becomes true. It is observed after a training cutoff of its own — the data-collection window — and then revised, which is the economic equivalent of a model that gets retrained but still reports its old cutoff as current. The committee sets policy on the print available at the meeting. The print available at the meeting is not the print that will exist in three months once revisions land. The gap between the two is the blind interval, and it is not small: US GDP first estimates have historically differed from third estimates by around half a percentage point at an annualised rate, which on a trillion-dollar economy is not noise, it is a missed decision.

What the economist cannot see on meeting day

The characteristic failure of the domain follows directly: policy is set on data that gets revised after the meeting that used it. This is not a data-quality complaint that better statisticians could fix. It is structural. The economy is a live stream; the reporting apparatus is a corpus with a cutoff, refreshed on a fixed schedule regardless of what has happened since the cutoff. The economist responsible for the briefing knows this — everyone in the building knows this — and still has to write a number down, because the meeting has a date and the date does not move to suit the data.

Some of the response has already moved along the lineage. Real-time nowcasting models — tracking high-frequency proxies like electricity use, card transaction volumes, job postings, shipping manifests — behave like a Large World Model bolted onto the forecasting desk: the blind interval inside that scene shrinks to days rather than months. This is a genuine improvement and it is exactly bounded the way the lineage predicts. The nowcast sees the scene it was built to see — output, mostly — and goes dark the moment attention turns to something the scene does not cover: an emerging credit stress in a sector nobody was proxying, a shift in inflation expectations buried in a household survey nobody runs weekly. The frame closes at its edges. A Large Universe Model, in this domain, would be the position where price indices, labour flows, credit aggregates and expectations all stay live simultaneously, each belief timestamped with when it was last touched and how much to trust it, so that "our CPI belief is fourteen hours old and unrevised" and "our credit-aggregate belief is six weeks old and provisional" are both visible to the person setting the rate, on the same page, at the same time. Nobody runs a central bank this way today. The category is argued, not built. The gap between the nowcast desk and that description is exactly the gap the lineage predicts remains open.

Two objections a monetary economist would raise

Central banks deliberately wait. The whole point of a six-weekly cycle and a data-dependent stance is to avoid whipsawing rates on noisy weekly prints. Framing the interval between fact and meeting as pure loss ignores that the delay is often the correct decision.

This is right, and it is the strongest objection in the domain. Optimal-stopping logic genuinely favours waiting for a posterior to tighten over acting on a jumpy weekly number. But the distinction that matters is whether the delay is chosen with a live feed running, or imposed by the feed's absence. A committee that can watch inflation expectations firm up in real time and decides to wait one more meeting is paying for evidence, deliberately, with full visibility of what it is waiting on. A committee that walks into a meeting with a GDP estimate that will be revised by half a point in three months is not choosing to wait — it is acting on stale information without knowing how stale, because nothing in the current apparatus timestamps the staleness for it. Continuous, provenance-tracked intake does not abolish deliberate delay; it is what makes deliberate delay a purchase rather than an accident.

You cannot price the opportunity cost of a rate decision made on stale data, because you cannot rerun the economy with the revised numbers and observe the counterfactual path. The whole framing is rhetorically vivid and empirically empty.

Conceded for the general case — nobody reruns 2008 with better real-time labour data and reads off the difference. But the claim does not need the general case. Where the same kind of decision recurs under different information ages, the gradient is estimable: episodes where a benchmark revision later showed the committee had misjudged the output gap by a percentage point at the time of decision; natural experiments from statistical agency outages and delayed releases, which have on occasion pushed committees to act on markedly older data than usual and produced measurably different outcomes when the revision arrived. Insurers and market-makers price exactly this kind of latency gradient routinely elsewhere. The metaphysical version of the claim — the cost of any counterfactual whatsoever — is unmeasurable and should be dropped. The operational version — that repeated, comparable decisions show rate misjudgement rising with data staleness — is testable, and where it has been tested, the gradient is positive.

Where the axis actually terminates

Central banking will not close its blind interval by hiring faster statisticians or shortening the publication lag by a week. Latency has diminishing returns here as everywhere, and a monthly CPI print is, for most of what a committee decides, economically sufficient — the cost of running every stream continuously would dwarf the value recovered for the bulk of ordinary meetings. What the domain shows is not that continuous intake should be built tomorrow, but that the shape of the missing capability is exactly the shape the lineage names: not a faster corpus, not a wider scene, but every stream live at once, each belief dated, so that the cost remaining is the price of latency and calibration, not the unmeasured price of not knowing you were already behind.