Dead time and transport delay in real estate

The best case against this argument

Here is the objection at full strength, the one an acquisitions lead with fifteen years of cycles behind them would make without blinking.

A good underwriting model does not need fresher data. It needs a better model of how today's data will age. Cap rate compression, absorption, migration — these move in patterns experienced practitioners have seen three times before. Build that pattern into the pro forma and you can hold a valuation through a shift you haven't measured yet, because you predicted it. Chasing faster inputs is what inexperienced shops do instead of learning the cycle.

This is not a strawman. It is the working philosophy of most institutional acquisitions desks, and it has a respectable pedigree in control engineering, where it is called a model-based predictor. The claim deserves to be taken seriously before it is answered.

Where the delay actually sits

Real estate underwriting is a feedback loop like any other: a signal comes in, a decision goes out, the market responds, the response is measured, the decision is revised. The signal side of that loop is fed by several streams running at very different speeds. Listing flow updates weekly. Rate curves move daily, sometimes intraday. Migration data — IRS county-to-county filings, USPS change-of-address volumes — lags six to twelve months behind the moves it describes. Permit filings sit in between: a building department records an application the day it is submitted, but the underlying signal it carries, developer conviction about future demand in a submarket, can be six to eighteen months ahead of anything a rent comp or cap rate survey will show.

That last fact is the whole problem. A valuation model built on trailing comps and cap rate surveys is not measuring current demand. It is measuring demand as it existed when the leases that produced those comps were signed, typically nine to fifteen months earlier depending on lease-up and reporting lag. Permit filings for competing product in the same submarket are visible now. The model's dead time — the interval between a real change in the world and the first observable movement in the model's own inputs — is set not by the slowest stream in the abstract but by the slowest stream the model actually uses. If the model ignores permits and leans on comps, its dead time is the comp lag, regardless of how sophisticated the cap rate assumptions are.

This is the mechanism, and it is worth stating precisely because the failure it produces looks like something else. An acquisitions lead underwrites a value-add multifamily deal in a submarket where permit filings for 1,200 new units cleared the planning department fourteen months before closing. The comps used in underwriting still show the old rent growth trend, because the new supply has not delivered yet and has not shown up in any leasing data. The model holds. The deal closes on a valuation that assumes rent growth the permits had already made structurally unlikely. Eighteen months later, absorption of the new units compresses rents in the submarket, the exit cap assumption no longer clears, and the position is underwater on a refinance. Nobody in the process was wrong about any individual number. The model was simply built on a stream with a longer transport delay than the one that mattered, and nobody had flagged the mismatch because the permit data was sitting in a different system with no timestamp attached to the valuation that used it.

The predictor defence, and where it holds

Back to the objection. A cycle-experienced acquisitions lead does, in effect, run a Smith predictor: an internal model of how comps will move given known leading indicators, used to act ahead of the lagging measurement rather than waiting for it. This genuinely works, and it is not a lesser form of rigour — it is the correct response to delay when you cannot shorten the measurement itself. Seasoned underwriters who mentally weight permit pipeline against absorption capacity are doing real predictive control, and it outperforms naive comp-following in every cycle on record.

But the defence has the same brittleness in real estate that it has in a chemical plant. A Smith predictor recovers lost bandwidth only to the extent that the internal model is accurate and the disturbance matches what the model was fitted to. Underwriters calibrated on the 2010s cycle, where new supply absorbed on a fairly predictable eighteen-to-twenty-four-month timeline, carried that internal model into 2021–2023, when construction financing costs and labour shortages stretched delivery timelines unevenly across markets. The mental model mismatched the actual delay structure, and the mismatch was invisible until absorption numbers came in wrong for two years running. This is the general failure mode of model-based prediction under delay: a small error in the assumed transport time does not degrade performance gracefully, it can destabilise the whole judgement, because the correction is now out of phase with the disturbance rather than merely late.

Continuous intake does not replace the acquisitions lead's judgement. It bounds the error the judgement is exposed to, by keeping the permit stream, the rate curve, and the migration data flowing into the same underwriting file with dates attached, so that when the internal model and the fresh stream disagree, that disagreement is visible before close rather than at refinance.

The economics objection, and where it survives

There is a second objection worth taking as seriously as the first, because it is also correct as far as it goes: most of underwriting does not need fast intake. Zoning code, easement law, structural condition, the mechanics of a 1031 exchange — these are stationary on the timescale of a single deal. Rebuilding a full live-data pipeline to re-underwrite a stabilised triple-net lease acquisition with a single national tenant is a waste of infrastructure. The tenant's credit does not move month to month. Paying for continuous permit-and-migration monitoring on every asset in a portfolio, most of which are not exposed to submarket supply risk, is bad economics.

This objection defeats the broad version of the claim and should be conceded fully. It does not defeat the narrow one. Value in acquisitions concentrates in exactly the deals where the loop is non-stationary: submarkets with active construction pipelines, migration-driven secondary markets, anything financed on floating-rate debt where the rate curve is the binding input. In those loops the cost of delay is not linear. A comp-lag error of a few percentage points in cap rate assumption on a stabilised asset is a rounding error in returns; the same lag on a development-adjacent acquisition, where the delay hides an entire new supply wave, is the difference between the deal working and the deal failing. The right architecture is not to accelerate everything. It is to know, stream by stream, what the dead time is, and to spend the monitoring budget on the streams whose delay is large relative to the deal's holding-period risk.

The narrower claim

Delay is not a data-freshness complaint to be solved by refreshing comps more often. It is a structural bound: the achievable responsiveness of an underwriting loop is capped by the slowest stream feeding the decision, and no amount of modelling skill inside that bound raises the cap, only manages living within it. A Large Language Model trained on a corpus with a fixed cutoff inherits the largest dead time of anything in its training data by construction; it cannot see a permit filed after its cutoff regardless of how well it reasons about the ones it saw. A Large World Model built around a live scene — a single submarket snapshot, refreshed for the duration of an analysis — closes the loop tightly while the scene holds, then discards it, so nothing persists to be checked against next quarter's filing. The position that survives scrutiny is the one where every stream — permits, rates, migration, listings — keeps running as a dated, decaying belief rather than a one-time import, so the acquisitions lead's judgement is exercised against known staleness rather than assumed currency. That is not a faster version of underwriting-as-usual. It is underwriting with its delay term made visible, which is the only place delay can actually be managed.