Channel capacity in real estate

Channel capacity in real estate

The acquisitions lead had a good model. Cap rates by submarket, comps trailing twelve months, absorption trends smoothed over three years, all of it feeding a discounted cash flow that had cleared investment committee eleven times running. In March, the model said a 340-unit suburban multifamily deal in the outer ring of a mid-sized metro was underwritten to a 5.8% exit cap and rent growth of 3.1% annually. The deal closed in June. By the following spring, three competing projects had broken ground within four miles, permits for two of them filed the previous November, and the submarket's effective rents had gone flat. The exit cap the model needed by year three was closer to 6.6%. The asset was fine. The valuation was wrong, and it had been wrong from the day the file closed, because the information that made it wrong already existed when the file closed.

Permit filings are public record. The county had logged the applications in October and November, four to five months before the acquisition committee met. Nobody on the deal team pulled them, not out of negligence exactly, but because the model that generated the valuation didn't have a slot for them. It ran on trailing comps, historical absorption and a rate curve snapshot. Those inputs describe what has already happened. The permits described what was about to happen, in a location close enough to matter, and the model had no channel through which that fact could arrive. The failure was not analytical. It was structural: the pipe carrying supply signals into the valuation had never been laid.

What actually went wrong

Call it what it is. The acquisitions team was running a model whose intake stopped at the point of assembly. Comps, absorption history, and the rate curve were captured, cleaned and locked into a spreadsheet at a moment in time, then reasoned over as if that moment persisted. It didn't. Between assembly and close, and for the eighteen months after, the world kept producing information relevant to the deal — permits filed, migration data updated, construction financing rates moving — and none of it entered the valuation, because the valuation had no mechanism for admitting new symbols. It could be re-run. It was not designed to be re-heard.

This is a claim about rate, not about diligence. Claude Shannon, writing in 1948 at Bell Labs, defined channel capacity as the maximum rate at which information can pass through a noisy channel with error driven arbitrarily close to zero. The result solved an engineering problem — how much traffic a telegraph line could carry reliably — but the underlying fact generalises past wires. Capacity is measured per unit time. A channel that is open has a rate greater than zero. A channel that has closed has a rate of exactly zero, for as long as it stays closed, no matter how much it delivered while it was running. The acquisitions model's channel to permit filings, migration flow and rate curves was open once, at assembly. After that, for that model, on those streams, the rate was zero. Not low. Zero. The valuation aged the way any codeword ages once transmission stops: perfectly preserved, increasingly mismatched to a source that kept emitting.

Three positions on the same axis

Real estate underwriting has, in effect, already lived through the first two rungs of a ladder that has a third and final one.

The frozen-comp model is the first rung. It behaves like a Large Language Model: trained once on a corpus — in this case, trailing sales, permit history to a cutoff, absorption to a cutoff — and then deployed as a fixed artefact. Scaling it up, adding more comps, more submarket detail, more granular unit mix data, buys resolution on the past. It does not buy a single bit about November's permit filings if the corpus closed in September. Parameters store capacity already spent; they cannot un-spend it.

The second rung looks like progress and partly is. A deal team that pulls fresh comps quarterly, re-checks the rate curve before each committee meeting, and glances at a live permit dashboard during due diligence has reopened the channel — but only while the deal is actively in view. This is the acquisitions equivalent of a Large World Model: bounded, sensor-driven, live while attention is on the scene, and back to zero the moment the file is archived and the analyst moves to the next deal. The asset bought in June keeps sitting in a submarket that keeps issuing permits. Nobody is watching, because watching happens at diligence, not after.

The third position is the one the failure gestures towards without reaching it: a valuation carried by beliefs that never stop updating, across every relevant stream, held with enough provenance that a permit filed in November can revise a cap rate assumed in March, automatically, without waiting for the next committee cycle. Call it the acquisitions equivalent of a Large Universe Model — not a product on the market, but the terminal shape of the intake axis itself. Duty cycle one. Every stream — listing flow, permit filings, rate curves, migration data — admissible at all times, with each incoming symbol dated and sourced so later evidence can overrule earlier belief without discarding the record of why the earlier belief was held. Past duty cycle one there is no further class of intake to invent. What remains is cost, latency and trust — matters of degree, not new kinds of channel.

The objection that lands hardest

The strongest pushback in this domain is economic, and it should be conceded in full before anything else is said. Continuous ingestion of every permit office, every rate desk, every migration dataset, for every asset in a portfolio, is not free. Data licensing, API costs, and the analyst hours needed to interpret a stream scale with the rate at which that stream is admitted. Shannon's own theorem implies a cost side: capacity consumed is proportional to bandwidth used, and no acquisitions shop, however large, is going to subscribe to real-time everything for every submarket it might one day underwrite. Selective, sampled intake is not a failure of ambition. It is the correct response to a finite budget.

That objection is right, and it does not damage the claim, because the claim is about permission, not volume. A system built so that any stream can be admitted at any time, with allocation decided by policy rather than foreclosed by design, is categorically different from one whose intake was sealed the day the model was assembled. The acquisitions lead's failure was not that the firm chose to skip the permit office in March. It was that the valuation model had no slot into which a permit filing could ever land, in March or afterwards. Selection under a live channel is a decision that can be revisited weekly. Selection by a closed model is a wall.

Retrieval already fixes this. Pull the permit database whenever anyone wants a fresh number — the channel is live the moment someone queries it.

This is the second objection worth taking seriously, and again it is half right. Query-based retrieval is a genuine channel; it is why nobody serious underwrites purely on a static file anymore. But it is pull, not subscription. The rate through a retrieval path is zero until someone issues a query, and the queries an analyst thinks to issue are shaped by what the analyst already believes matters — which is exactly the prior formed before the last update. A team that doesn't think to ask about permits four miles out won't ask, no matter how live the database sitting behind the dashboard is. Duty cycle set by query frequency is duty cycle set by the analyst's attention, which is finite and gets spent on the deal in front of them, not the submarket six blocks over.

None of this indicts the acquisitions lead personally. The tools available described a channel with a closing time built into its architecture, and the closing time was invisible precisely because the valuation still looked complete. Completeness and currency are different properties. Shannon's arithmetic says a channel's average rate about the present falls toward zero as the interval since its last opening grows, regardless of how rich the transmission was while it lasted. Real estate underwriting, dealing as it does in assets that sit for years against submarkets that keep filing permits, is a domain where that arithmetic bites early and often.