Data requests & next steps

The regime-flip forecaster (months of supply · crossover dates) runs today on Green Street data alone. This page states its two assumptions, the free validation path, and the exact paid-data requests — in priority order.

01What we built and what it assumes

Flip dates come from Green Street's scenario occupancy paths (quarterly to 2031Q1) crossing each metro's own long-run equilibrium occupancy. Months-of-supply converts the occupancy overhang into units via a backed-out inventory figure. Two stated assumptions: (a) inventory back-out — metro stock is inferred as 12× the GS T12-monthly-average permits divided by permits-as-%-of-inventory (magnitudes check out: Atlanta ~580k units, Austin ~360k — but this deserves independent confirmation); (b) equilibrium anchor — each metro's own full-history mean occupancy (±25bp band). Both are documented in flip_forecast.py and flagged on every surface that shows the numbers. Horizon caveat: "no flip by 2031Q1" means exactly that — the forecast ends there; it never means "never."

02Free validation — DONE (2026-07-11), both assumptions hold

Inventory back-out vs Census ACS 5+ unit stock: PLAUSIBLE. Cross-metro correlation 0.967 (log-log), median ratio 0.94 — the back-out tracks near-total metro apartment stock. Five metros run ~10–20% hot on unit counts (Salt Lake City, Richmond, Columbus, Charlotte, Indianapolis) — a pure scale effect that leaves months-of-supply and flip dates essentially unchanged. Equilibrium occupancy vs Census HVS rental vacancy: PLAUSIBLE. Correlation 0.61 (p<1e-4) across 35 metros; levels consistent with institutional-vs-all-rental coverage; Boston is the one metro flagged for a second look. Full method, tables and raw pulls: docs/census-validation.md and data/ref/census_validation/. Far-dated flips can now be treated as evidence-checked rather than directional.

03Paid-data requests — exact asks, priority order

Request 1 — Green Street Snowflake access (to our GS rep; costs nothing beyond the existing subscription). Ask verbatim: "We want programmatic access to the data we already license — specifically (a) full Sales Comps transaction history without the ~1,050-row export ceiling, (b) Rent Comps history (the web tool exposes only a current 1,000-row snapshot), (c) the full Development Database with delivery schedules, and (d) quarterly forecast releases as data, so we can archive vintages automatically. Please enable Snowflake data sharing or equivalent API access for our account." This single grant replaces browser-download choreography, deepens the sales-comp history to 2015+, unlocks rent time series, and hardens the vintage archive. Add one clarification question to the same email: "Your web tool's asking-rent YTD aggregation runs ~4.7pp above the forecast file's effective-rent growth for the same metros (Q2 2025) — please confirm the definitional difference between the two series." (Found during trade-out validation, 2026-07-11.)

Request 2 — one apartment-market data vendor (RealPage, CoStar, or Yardi Matrix; real budget). What to buy, stated as requirements, vendor-neutral: metro AND submarket occupancy/vacancy history (validates equilibrium better than Census); unit-level pipeline with expected delivery quarters for our 37 metros (replaces GS's smoothed supply path with lumpy reality — directly sharpens flip dates and the delivery-window heatmap); effective rent time series at submarket grain (replaces the Rent Comps gap). Evaluation question for the trial: "can we export full history programmatically, not through a capped web UI?" — we have been burned once.

Request 3 — later, only if the expense thesis matters: FEMA National Risk Index (free) plus a property-insurance premium series (state filings or a broker relationship) to graduate the climate overlay into a scored expense-drag pillar.

04Sequence

1. Send Request 1 today (email to the GS rep — zero cost, unblocks four things). 2. Run the free Census validation this week. 3. Trial one vendor from Request 2 against the requirements list. 4. Each new source lands in data/raw/, gets registered with lineage like f01–f11, and the flip forecaster re-runs with tighter assumptions — the QA gate guarantees nothing silently changes.

Maintained by the data pipeline · Maintained 2026-07-11 · Holland Market Research — powered by TenureOS