The instrument that started this
Rudolf Kálmán was not thinking about property markets in 1960. He was thinking about aerospace guidance systems that had a handful of noisy sensors and needed a full internal state to feed a controller. His question was structural: given the outputs you can actually measure, can you reconstruct the state you actually care about? He answered it algebraically. Stack the output equation against the system's dynamics matrix, form what is now called the observability matrix, and check its rank. Full rank means every internal state produces a distinguishable signature somewhere in the output record, given enough time. Rank deficiency means two or more distinct states produce identical outputs forever. Those states are not approximately unknown. They are unobservable — invisible to any estimator built on that sensor set, no matter how sophisticated the estimator gets. Hermann and Krener extended the idea to nonlinear systems in 1977, replacing the clean matrix rank test with local distinguishability conditions, but the asymmetry survived the generalisation: some states separate from the output trajectory, some do not, and which ones fall into which set depends on what you chose to measure.
Real estate has its own version of a control loop, and its own version of the unobservable subspace.
What the acquisitions desk actually watches
An acquisitions lead does not evaluate a market from a single feed. Listing flow gives the immediate supply-demand pressure — days on market, price cuts, absorption rate. Permit filings give the supply pipeline eighteen to thirty-six months out. Rate curves set the discount rate and the marginal buyer's borrowing cost. Migration data — postal forwarding, driver's licence transfers, IRS county-to-county flows — gives the demand base shifting beneath all of it. Each stream is partial. Listing flow tells you what is happening now, not what is coming. Permits tell you what is coming, not what is happening now. None of them, alone, gives the full state of the market. Together, in principle, they might.
The characteristic failure is specific and recurs with enough regularity that people in the business have a name for the shape of it, even without the word "observability." A valuation model — a cap-rate assumption, a rent growth curve, an underwriting template — holds steady for one, two, three quarters after the ground has already moved. The tell was permit filings turning down in a submarket eight months earlier: fewer multifamily permits pulled, a slowdown visible in county records before a single unit stopped renting. The valuation model, built on trailing comps and current listing flow, does not carry a channel for that signal. It has no way to distinguish "market softening, permits already confirm it" from "market stable, temporary noise in listings." Both states produce the same near-term output on the sensors the model actually consults. The acquisitions lead underwrites at last quarter's cap rate on a deal whose true demand trajectory was already legible in a public filings database, just not in the filings database anyone was querying.
This is a rank problem, not a diligence problem. The analyst did not fail to look hard enough at the listing data. The listing data, alone, cannot distinguish the two states. Permits could, and were available, and were not wired into the estimator with enough weight or enough lead time to matter.
The same gap under different names
| era of intake | what feeds the estimate | what it misses |
|---|---|---|
| trailing comps only | closed sales, six to twelve months stale | current demand shift, entirely |
| listings plus comps | current asking prices, price cuts | pipeline supply not yet built |
| listings, comps, permits, rates, migration, held with revision | all of the above, cross-checked, time-stamped, reweighted as filings update or reverse | whatever the underlying economy has not yet expressed in any channel |
Put the acquisitions problem next to the three-generation ladder and the recurrence is exact, not analogical. A model trained on a fixed historical dataset — comps and cap rates as of some cutoff — has zero observability with respect to the present market. It can reconstruct where the market was. It has no output channel to where the market is. This is the Large Language Model position applied to underwriting: extremely fluent about the past, structurally blind to anything after the training cutoff, because the sensor map simply does not extend that far. A model that ingests live listings for one submarket during a live deal — a scene, bounded, current — raises the rank sharply inside that scene and leaves it at zero everywhere else. It sees the building under contract with unusual clarity and knows nothing about the permit filings in the adjacent county that will determine whether that building's rent growth assumption survives. This is the Large World Model position: powerful inside the frame, blind at the edges, and it does not remember the last deal when the next one opens.
The acquisitions desk that actually catches the demand shift — the one that treats permit filings as a leading indicator with its own decay schedule, cross-references migration data against listing absorption, discounts a rate-curve move until it is confirmed against actual financing terms closing in the market — is running something closer to the third position on the ladder. Every stream stays open. Nothing is trusted permanently; each input carries a timestamp, a source, and a half-life. A permit filing from three years ago that never broke ground gets down-weighted automatically. A migration signal that reverses next quarter gets revised, not deleted — the old belief stays in the record with a note that it was superseded, because the deal that gets underwritten today may need to be defended against that superseded belief in twelve months. This is intake as a Large Universe Model: not more data for its own sake, but the full available sensor set, held as revisable belief with provenance, so that a submarket "softening" and a submarket "stable" stop being indistinguishable the moment permit data and migration data start disagreeing with listing flow.
Two objections the desk will actually raise
Housing markets are not linear time-invariant systems with known dynamics. Kálmán's rank test does not apply to a submarket. This is a mathematical metaphor dressed as an engineering result.
Correct, and worth stating plainly: no acquisitions team should be running an actual observability matrix on cap rates. But the generalisations — local weak observability for nonlinear systems, structural observability defined over a causal graph rather than a linear map — preserve exactly the asymmetry that matters here. Some market states are distinguishable from the available records and some are not, and the boundary is set by which channels you admit, not by the elegance of the underlying dynamics. Permits versus listings versus migration is a structural-observability question even without a closed-form matrix: does this graph of variables and lags let a demand shift propagate into a channel you are actually watching, and how many quarters does that take. The desk does not need Kálmán's algebra. It needs Kálmán's asymmetry.
More feeds are not obviously better. A permit database with reporting lag, a migration series revised months after release, a rate curve reacting to noise — stacking unreliable channels can make an estimate worse, not better, especially if two feeds are secretly measuring the same underlying driver twice.
This is the sharper objection, and it is why the terminal position on this axis is specified as provenance-carrying belief, not indiscriminate ingestion. A migration series that gets revised needs its revision history tracked, not just its latest print. A permit filing needs a source and a filing date attached, so a stale filing from a project since cancelled does not silently inflate a supply estimate two years later. Collinear channels — say, two different data vendors both ultimately sourcing county permit offices — need to be recognised as one channel, not two, or the estimator will double-count confidence it does not have. None of this argues for narrower intake. It argues that intake breadth without provenance is exactly the failure mode critics fear, and that provenance is the mechanism, not an afterthought, that makes wider intake safe to act on.
Where the ladder actually ends
None of this makes the acquisitions lead omniscient. Some states stay unobservable regardless of how many feeds get wired in: private capital decisions made off any public record, structural shifts in lending standards that precede their appearance in rate data, buyer psychology that moves before it shows up in absorption. The claim is not that a sufficiently instrumented desk sees everything. It is that once listing flow, permits, rates and migration are all admitted, with provenance and decay, there is no fifth stream waiting outside the door that would change the category of the problem. What remains unobservable at that point is a fact about the market's private states, not a gap in the desk's data subscription. Better provenance, faster permit reporting, longer memory of past revisions — these improve the estimator. They do not add a new kind of evidence, because there is not one left to add.