The Desk

How this research came together, what it can do for you, where the data still has holes, and the access that completes it. One page — printable.

01How the data came together — our thinking

We started from a rule, not a dataset: every number on this desk must be traceable to a registered source, and every claim must survive an independent check before it is trusted. Eleven raw Green Street files (forecasts, grades, sales comps, SFR, credit) plus Holland's own house-view deck and third-party context were registered with lineage, rebuilt into one panel, and locked behind nine QA gates — including a clean-room rebuild that must reproduce the panel byte-for-byte. Nothing is filled silently, market names are matched exactly or not at all, and every figure is stamped with its source and as-of date. Where we had to assume (metro inventory, equilibrium occupancy), we validated against Census and published the verdicts. Where sources disagree — Holland's real lease trade-outs vs. Green Street, our challenger model vs. the house forecast — the disagreement is kept visible and treated as the most valuable research lead we have. Anecdotes (market chatter) are collected but quarantined: they never touch the research signal.

02Capabilities — what you can ask this desk for

Market ranking & entry screens

37-metro research signal rank on five lenses with IC-backtested weights (undersupply pillar IC +0.221, positive in 81% of periods) — plus top-10 develop and top-10 purchase-below-replacement screens.

Market studies

A fixed-spine study (sections A–H) for every covered metro — fundamentals, demand bridge, valuation, flip clock — as web page, PDF, and machine-readable JSON.

Cycle timing

Regime-flip forecaster: months of supply and the quarter each market crosses between oversupplied and constrained, out to 2031Q1, Census-validated. Delivery-window heatmap and "out of the ground by" deadlines per metro.

Buy vs. build

Replacement-cost quadrant per metro backed by Green Street sales-comp evidence — where acquiring below replacement beats developing, and where it flips.

Portfolio & vintages

Holland holdings overlaid on every signal, and vintage-capture analysis: what each acquisition year harvested versus what the market offered.

Scenario engine

All five Green Street scenarios are first-class on every surface — one selector re-renders the whole desk, so a downside case is one click, not a new report.

Independent forecast challenge

A walk-forward challenger model tests Green Street’s rent path every quarter. Current verdict, reported honestly: a simple AR(4) still beats the neural nets — divergences are flagged as research leads, not hidden.

Validation & self-audit

Three-way trade-out check (HPG’s real leases vs. Green Street vs. the desk), Census validation of every forecaster assumption, and a standing Evidence & Disagreements panel that shows what held and what didn’t.

Ask the desk

An on-device AI analyst grounded in the same data feed the site renders — it cites its sources, carries every caveat, says "not in the desk’s data" rather than guessing, and licensed data never leaves the building.

Bespoke questions

Give the desk a question — a new screen, a backtest, a market deep-dive. It lands through the same pipeline, so it arrives with lineage, QA, and provenance, not as a one-off spreadsheet.

03The checklist — what the story still needs

Everything below is the gap between the desk today and the full story it is built to tell. Checked items are shipped and verified; open items name the exact data or decision that unblocks them. Exact vendor-request language lives on the data requests page.

Shipped & verified

Forecast ledger frozen — 242 machine-derived calls for 2026Q1 (window calls, flip medians, trigger conditions, timing values, 5 revision calls); byte-identical regeneration, grading armed.
Flip-forecaster assumptions Census-validated — inventory back-out (corr 0.967) and equilibrium anchors (corr 0.61), verdicts published.
Three-way trade-out validation — HPG's real leases vs Green Street vs the desk agree on direction and rank order.
House intel extracted with integrity checks — Card Counter 17/17 cross-check, supply wave reconciled to 428,100 units; quarantined, never blended with the research signal.

Open — decisions on our side

Trigger adoption pass — done 2026-07-12: all 15 rules adopted; the trigger board is standing house policy, recorded in the ledger and re-proposed for review at each release.
Send the Green Street Snowflake request — one email, zero cost; removes the comp export ceilings, unlocks the full development database and automatic vintage archiving, and carries the asking-vs-effective series question. Unblocked by: sending it.
2026Q2 release day — one command sequence rebuilds the desk, archives the vintage, and grades all 242 frozen calls: the calibration story, the revision watchlist scores, and the revision monitor all begin that day. Unblocked by: the next Green Street release (time).
Swap the DRAFT house deck — the Concord/Newmark view is Q2 2025 DRAFT vintage; replace with the final when Holland issues it. Unblocked by: Holland.

Open — data we need to complete the story

Second forecast vintage — the single biggest unlock: turns scenario-implied probabilities into calibrated ones and every frozen call into a graded one. Arrives: 2026Q2.
Unit-level delivery pipeline (vendor: RealPage / CoStar / Yardi Matrix) — Green Street's supply path is smoothed; real delivery schedules make flip dates lumpy and true, and sharpen the timing-value index.
Submarket occupancy & rent history (same vendor) — validates equilibrium anchors properly and extends the desk below metro grain, where the house decks already operate.
Rent-comp history (Snowflake) — today's trade-out validation is one quarter, five metros; history makes it a standing check.
Insurance premium series + FEMA NRI — graduates climate from context overlay to a scored expense pillar, if the expense thesis matters.
Deal-level cost & yield assumptions (HPG internal) — the timing-value index speaks in basis points on purpose; deal assumptions are what convert it to dollars honestly.
Colorado Springs, Tucson, Boise — on the 40-metro board, not in Green Street's coverage; needs an alternative source or a stated decision to accept the gap. (Credit data stays national-grain by design — that one is not a gap.)

04The PyTorch path — data that unlocks deep forecasting

The desk already runs PyTorch on-premise (NVIDIA GPU): our challenger harness trains neural forecasters every quarter and judges them walk-forward against simple baselines. Today's honest verdict: an AR(4) still beats the LSTM — because the training panel is small (37 metros × ~100 quarters × one full cycle), not because deep learning can't forecast rents. The list below is the specific data, per source, that changes the verdict. The harness re-judges automatically as each source lands; nothing ships unless it beats the baseline out-of-sample.

Free & public — pulled 2026-07-12, awaiting ingestion (f16+)

Census BPS monthly permits by CBSA — pulled: 432/432 monthly files, 1988→2023-12, with sha256 manifest. Metro grain was discontinued after 2023; 2024+ continues via County files aggregated to CBSA (recorded in the manifest). Unlocks: supply-pipeline sequence models; 3× the observations of our quarterly panel.
BLS metro employment, monthly — pulled: all 38 series (37 metros; metdiv codes for LA/OC/Oakland/SF/Miami, Raleigh+Durham as two CBSAs), 1990→, every series returned data. Unlocks: demand nowcasting features across three recessions.
FHFA house-price index by metro — pulled: quarterly from 1975, all metros — the longest metro series available anywhere. Unlocks: cycle-regime classifiers trained on more than one crash.
Zillow ZORI — pulled: monthly metro rents 2015→. Unlocks: higher-frequency rent targets + an external validation set.
Ingestion into the training panel — done: f16–f19 registered with lineage, 45,872-row panel (37/37 markets per source), overlap validation vs Green Street: permits 0.984 · employment 0.893 · ZORI 0.925 (Minneapolis flagged) · FHFA 0.302 (related asset).
Challenger v2 retrained (2026-07-12) — verdict, reported honestly: AR(4) still wins; the free public features degraded the learned models out-of-sample. Consistent with this page's ranking — the real unlocks remain vintage labels (2026Q2) and vendor submarket cross-section. Full report: docs/challenger-v2-report.md.

Via the Snowflake request (zero cost — already on the checklist)

Green Street forecast vintages as data — every past quarterly forecast, not just the current one. The crown jewel: realized forecast ERRORS become training labels, which is what calibration models and the revision watchlist need to learn from rather than wait for.
Full sales-comps history (2015→) — unlocks pricing and liquidity models instead of a single 12-month snapshot.
Rent-comps history — unlocks property-grain rent training series.
Development database with status transitions — proposed → approved → built / stalled, with dates. Unlocks a delivery-probability model: the learned version of the Card Counter's financed-vs-shadow split.

Vendor budget (RealPage / CoStar / Yardi Matrix)

Submarket occupancy & rent panels (1995→) — ~500 submarkets instead of 37 metros. The single biggest unlock for deep models: ~40× the cross-section is what lets sequence models finally out-learn the AR(4).
Unit-level pipeline with outcome histories — unlocks lease-up-curve models: predicted absorption pace per delivery, which is the learned upgrade to the timing-value index.

Internal (HPG)

Portfolio lease-level trade-out history — the deck's p03 table, but as a time series. Unlocks: ground-truth micro-labels no vendor sells; models tuned to HPG's actual assets.
Historical underwrites vs realized outcomes — unlocks the dollars layer: a model of where our own pro-formas run hot or cold, which is the only honest way to put dollar figures on forecasts.

Standing rule for all of it: new data lands in data/raw/ with lineage like f01–f15, the challenger harness retrains and reports walk-forward results against the AR(4) baseline, and the model only replaces the baseline when it wins out-of-sample — the same honesty that produced today's verdict. House intel stays quarantined from training; forecasts stay frozen in the ledger and graded.

Holland Market Research — powered by TenureOS · Green Street 2026Q1 release · every figure stamped source + as-of · 9/9 QA gates passing