How FWDRE measures Manhattan retail

FWDRE maintains live status verdicts for Manhattan storefronts, fused from tens of thousands of retail-location records — the live, per-neighborhood counts on our neighborhood pages refresh nightly. No single feed is trusted on its own — every verdict is fused from three refresh axes that fail independently, so a blind spot in one cannot silently become a blind spot in the answer.

When new evidence lands on any axis, the affected verdicts recompute: roughly 90% of status verdicts have been re-resolved within the last 48 hours. as of Jul 2026

Most data vendors force every storefront into open or closed. We don't. Every FWDRE verdict lives in one of three tiers, and the third tier is the one that keeps the other two honest.

• A closure verdict must be earned. "Closed" requires two or more independent sources, one of which must be a direct observation of the storefront. Paper evidence alone can never kill a business in our data.
• Absence of evidence is not evidence of death. A lapsed license is a question, not a verdict — plenty of real, operating businesses leave no fresh paper trail.
• Every signal carries its real evidence date. The inspection date, the license expiry, the observation timestamp — never the date we happened to ingest it. Old facts are never dressed up as fresh ones.

Evidence ages. A health inspection from last week says more about a restaurant than one from last year, and different kinds of evidence go stale at different speeds. So every signal in the fusion carries a weight that decays on a half-life set by its category — fast for momentary signals, slow for annually-renewed ones.

• Fresh on any axis reads open. One current, dated proof of operation is enough.
• A hard negative outranks a stale positive. When every axis has gone stale and a direct negative observation arrives, the negative wins — even against a lingering "open" from a lagging commercial feed.
• Nothing fresh, nothing decisive — honestly-unknown. The model is built to withhold credit, not to manufacture certainty.

weight(signal) = 2 ^ ( - age_days / half_life[category] )

verdict:
  open     — any axis fresh: dated, current proof of operation
  closed   — 2+ independent negatives, incl. a direct observation
  unknown  — everything stale, nothing decisive

The formula class is public and the fitted parameters in force — category half-lives and thresholds — are printed in every report. Nothing in the verdict path is a black box.

We print the interval, not just the point — an honest range beats a flattering number. The audit re-runs on an ongoing cadence (25 storefronts July 1 → 100 storefronts July 6; the hard “closed” tier ran 20-for-20 in the latest pass), and the current figure with its interval is printed in every report. It is also backed operationally: any status call a client disputes is re-verified against live sources within 48 hours, free.

The same honesty applies to the closure rate itself. Closures are structurally undercounted — a storefront can go dark months before any paper trail catches up. A single point estimate would flatter us, so FWDRE publishes closure as a band between a hard floor and a soft ceiling:

Every input, its refresh cadence, and the exact role it is allowed to play. Sources earn specific jobs — none of them gets to answer questions outside its competence.

Identity data includes Overture Maps Foundation data and Foursquare Open Source Places, used under their open licenses with attribution. Google is a rented, lagging signal used as an input to fusion only — it is never stored as the foundation and never redisplayed. LL157 registry filings are self-reported and carry neither rent nor square footage; we use them strictly for vacancy, lease-expiry, and turnover evidence. Distances in FWDRE reports are shown in feet and miles (US commercial-real-estate convention) but computed in meters from coordinates; method and source lines print the metric figure once alongside, so every distance stays reproducible.

• No 100% accuracy. Nobody measuring the physical city has it, including us. What we offer instead is a published error rate, its confidence interval, and the direction it errs.
• Listings are not availability. A listing is marketing; availability is a fact about a lease. We never infer one from the other.
• Candidate "leads" are evidence flags, not verified listings. When we surface a space as a likely opportunity, we show the evidence and its dates — we do not represent it as on the market.
• Coverage is Manhattan today. Brooklyn is in progress. Outer-borough registry data exists in our pipeline, but it has not yet met the verification bar this page describes — so we don't sell it as if it had.

If you believe a figure is wrong, tell us. Disputes are logged, the underlying evidence is re-pulled, and the outcome — a correction or a published errata note — is attached to the affected report.