How WindMayor computes everything you see.

Every number on every WindMayor surface comes from a public, free, cited source — or it's labeled [ILLUSTRATIVE]. This page is the paper trail. If you can't reproduce a value from the source we cite, that's a bug. Tell us.

The DSCR formula (v4)

DSCR (Debt Service Coverage Ratio) is the single most important number on the site. It answers: does the rent on this property self-fund a typical loan against it? A DSCR ≥ 1.20 is the conventional minimum that lenders underwrite to. The same exact formula is used across military bases, colleges, and cities.

Property tax is the bigger lever

Property tax varies far more across states than insurance does, so replacing the flat 1.10% is what actually moves DSCR. The effective rate runs from 0.28% in Alabama and Hawaii to 1.85% in New Jersey — a 6.6× spread (Tax Foundation, "Property Taxes by State and County," property taxes paid as a % of owner-occupied housing value). Texas (1.49%) is a clear example: moving from the flat 1.10% to the real Texas rate cuts a typical Texas DSCR by roughly 0.17. Insurance uses the NAIC 2022 Homeowners Insurance Report (released 2025), Table 4 — the HO-3 state-average premium divided by the exposure-weighted amount of insurance, then a labeled ×1.20 uplift to a DP-3 dwelling-fire / landlord proxy, because these are rentals, not owner-occupied policies. The ×1.20 is an explicit assumption: NAIC does not publish a clean DP-3 average. The national HO-3 average is 0.40% of insured value, but the state spread is real — the DP-3 rate runs from 0.26% in Utah and Oregon to 0.98% in Louisiana and 0.92% in Oklahoma. Every record carries prop_tax_pct_used and insurance_pct_used so the inputs are auditable per row.

The variable maintenance model (v4) — an estimate, labeled

v3 carried a flat 0.50% maintenance reserve on every property in the country. That is plainly wrong — it costs more to maintain a 90-year-old house in a high-cost-of-construction metro than a new build in a cheap one. v4 replaces the constant with a per-ZIP estimate written to zipdata/data/maintenance_by_zip.csv for all 31,646 ZIPs:

Across all ZIPs the model produces a median maintenance reserve of 0.97% — close to the 1% rule-of-thumb base, with a real spread from the 0.6% floor to the 1.8% cap. The BLS construction-cost index is populated for 31,468 of 31,646 ZIPs; where it is missing the multiplier defaults to 1.00 and the row says so in its source string.

Property-type pricing — price what the investor would actually buy

Zillow ZHVI is now pulled by property type — single-family (SFR), condo, 2-bedroom, 3-bedroom — not just the all-homes blend. DSCR can then be priced off the unit an investor would realistically buy in that ZIP rather than a market-wide average. Colleges carry dscr_condo (the UI primary), dscr_sfr, and dscr_2br side by side; each row records the property_basis it used and falls back down a ladder (type → sfrcondo blend → all-homes) when a type series is missing.

Apples-to-apples DSCR (v4) — houses pencil better than condos

A DSCR computed from condo rent over single-family value (or vice versa) is apples-to-oranges. v4 ships a strictly matched-type DSCR in zipdata/data/dscr_by_type.csv: dscr_sfr uses single-family rent ÷ single-family value, and dscr_condo uses condo rent ÷ condo value — never a cross.

The matched-type comparison settles a question the all-homes blend hid: across the 2,044 ZIPs with a single-family DSCR and the 1,029 with a condo DSCR, the median dscr_sfr is 0.74 against a median dscr_condo of 0.65. Houses pencil better than condos — condo values have run ahead of condo rents, and an HOA-free house carries a leaner cost stack. Neither type clears 1.20 at today's rate; this is a relative ranking, not a green light.

What we are still NOT modeling

• HOA fees (varies wildly; would need per-property data)
• CapEx / rehab cost — the v4 per-ZIP maintenance reserve is a structural estimate (construction cost × building age), not a measured CapEx plan tied to a specific roof or HVAC
• Closing costs, lender fees, points
• Within-state tax variation — the rate is a state effective average; individual counties and millage districts vary around it
• Refinance or rate reset risk

This DSCR is a screen, not an underwrite. It tells you which markets are worth your attention. The actual deal needs property-specific inputs.

Data sources, by surface

What does NULL mean?

When you see a blank or — in the UI, that field has no real source value for that record. We do not estimate, smooth, interpolate, or fill from a neighbor. Examples:

• A small military installation with no Wikipedia page → troop_count = NULL
• A college whose ZIP is not in Zillow's ZHVI universe (truly rural) → dscr = NULL (a future ticket adds city / county fallback)
• An extreme outlier (e.g. p2r > 60, or |dscr| > 5) → NULL for display; raw value kept in *_raw field

What does [ILLUSTRATIVE] mean?

Listings on TaxFight currently show illustrative example deals — clearly tagged [Illustrative] — because the marketplace is pre-launch. Once live inventory ships, the tag goes away on real listings. Estimates inside an illustrative listing (rent estimate, DSCR estimate) are bracketed (est) for the same reason.

The rate spine

Every DSCR calculation across all surfaces uses the same FRED MORTGAGE30US rate — pulled live where possible, otherwise the latest observation from the cached rate series. The +0.70% spread above the Treasury is documented OfferMarket pricing for DSCR investor loans, used because investor loans price higher than owner-occupied. The single rate spine means no surface can quietly drift from another.

The rate model (v3)

v3 splits the rate question into two honest pieces — a short-horizon model that we tested and only ship where it has real walk-forward skill, and a 5-year forward cone that is explicitly a range of outcomes, not a forecast. We do not pretend to predict multi-year rates.

Short-horizon model — weekly, real edge at 1 month, no edge at 3

The model is OLS at weekly frequency — the cadence of the Freddie Mac PMMS survey. The textbook Treasury-leg + spread-leg decomposition was tested and discarded: forecasting the 10-year Treasury added noise. The signal that actually exists is the weekly lag-catch-up of the PMMS survey rate to Treasury moves that have already happened. Three features: 10-year Treasury 1-week momentum, 10-year 4-week momentum, and the mortgage–Treasury spread's deviation from its 52-week mean. Walk-forward — at each week the model is fit only on pairs whose outcome was observed strictly before that week, no lookahead.

The 5-year forward cone — a range, not a prediction

The forward cone draws a central path and a fan of bands out to 2031. It is built from 2,878 weekly MORTGAGE30US observations (1971–2026). Read it as a range of plausible outcomes — multi-year rates are not forecastable and the cone does not claim to be a forecast.

• Central path. Weeks 0–13 use the short-horizon model (the only horizons with tested skill). Weeks 13–260 are a geometric mean-reversion glide toward a long-run anchor of 4.87% — the trailing-decade mean of MORTGAGE30US, with a ~2-year half-life. No week-to-week wiggle is predicted past 3 months.
• The anchor deliberately excludes the 1980s. 4.87% is the mean of the trailing 520 weekly observations, not the full 1971–2026 history — that history embeds the 1980s double-digit-rate regime, which is not a sensible reversion target for a projection starting in a post-2022 environment.
• The bands are empirical, not assumed. The p10 / p25 / p75 / p90 bands are the actual historical percentiles of rate(t+h)−rate(t) moves over each horizon, centered on their own median. No assumed volatility. The cone fans out only because rates have in fact moved more over longer windows.

That empirical dispersion is wide and worth seeing plainly. The 10th-to-90th percentile spread of historical rate moves is 0.57 pts at 1 month, 2.47 pts at 1 year, and 6.16 pts at 5 years. The 5-year band runs from roughly 3.0% (p10) to 9.2% (p90) by 2031, around a central path near 5.3%. Anyone underwriting a 5-year hold should size against the band, not the central line.

Display outlier handling

Some IPEDS / Zillow records produce wild values — a college reporting a 29,000% enrollment YoY (probably a coding error), a ZIP with $5M ZHVI and $4k rent (real but breaks the screen). We clamp these for display only:

• enrollment_yoy_pct outside [−50, +50]% → NULL in UI
• home_value > $3M or rent > $8k/mo → NULL; flag hv_quality / rent_quality = extreme
• price_to_rent > 60 or < 5 → NULL; flag p2r_quality = extreme / thin_sample
• The raw value is preserved in *_raw so nothing is lost

Condo-basis DSCR — and where it misleads

College and city DSCR is computed on Zillow's condo-tier ZHVI (City_zhvi_uc_condo_tier_…) rather than the all-homes blend. The reason: the investable unit near a campus or in a dense city core is usually a condo or small multifamily — not the detached single-family house that dominates the all-homes index. Pricing DSCR off a $600k median house when the actual buy is a $300k condo would understate yield and overstate the loan. Switching the basis flipped the DSCR sign for 1,133 colleges and shifts every city with a condo series.

What the research found

Several WindMayor surfaces have a research arm — out-of-sample regressions that test whether the patterns we screen on actually predict. We publish these including the failures. A negative R² is reported as a negative R². The point of this page is that a reviewer can trust it; that only works if the losses are as visible as the wins.

BAH regression — the model does not extrapolate

We asked: can local economics predict how fast a base's Basic Allowance for Housing grows? OLS on n=177 military housing areas, target bah_5y_cagr, 80/20 train/test split.

The only two features that carried a real signal were zhvi_5y_cagr (β ≈ +0.009, p < 1e−5) and zillow_zori_5y_cagr (β ≈ +0.005, p ≈ 0.005). Plainly: BAH growth tracks local home-value and rent growth, and nothing else added signal. Troop count, years active, latitude, distance to coast, education, median age, unemployment, IRS net migration — every one was statistically indistinguishable from zero. Worse, adding those eight features in v2 made the test R² more negative (−0.27 → −0.51): the extra parameters overfit the 142-row training set. A negative test R² means the model predicts held-out BAH worse than the training mean. We do not ship this as a forecast. The honest takeaway is the null result itself: if you want to anticipate BAH, watch Zillow for that ZIP — the DoD schedule is already doing roughly that.

City rent regression — better, but ZHVI-dominated, and migration inverted

Same method, applied to 250-city rent. OLS, n=244, target zori_5y_cagr. R² train 0.482, R² test +0.276, MAPE test 16.8%. A genuinely positive out-of-sample R² — the model does carry real signal — but it is again dominated by zhvi_5y_cagr (β +0.0071, p ≈ 4e−15): home-value growth predicts rent growth.

AVM v1 — the pipeline is the deliverable; the accuracy is degenerate

The AVM research preview is a hedonic OLS home-value model built entirely on free public data — 7,141 OpenStreetMap building footprints for the Seattle metro, ZIP-level ZHVI as the label.

DSCR rate stress test — who survives a +73bp rate shock

The DSCR stress test reprices every surface at the rate-forecast upper band. Rate now: 6.51% (FRED MORTGAGE30US). Five-year upper band (p90): 7.24% for 2031 — a +73bp shock. Rent and home value are held constant; only the rate moves.

Military bases are the fragile surface: at +73bp, 7 of the 10 bases that clear DSCR 1.0 today fall below it — a 70% dropout, with the survivors clustered at razor-thin margins. Seattle shows 0% dropout for the worst reason: no Seattle ZIP clears 1.0 even today (median DSCR 0.16), so there is nothing left to lose — it is already underwater, not resilient.

One caveat on vintage: the stress page still reprices against the older forecast_ensemble p90 band (7.24% for 2031) and the pre-v3 flat operating-cost constant. The newer empirical 5-year cone above is wider; a refresh of the stress test onto the v3 cone and the per-state DSCR is a tracked follow-up.

The ZIP spine — zip_master

Every nationwide surface (the ZCTA heatmap, Opportunity Sigma, the zipdata research) is built on one file: zipdata/data/zip_master.json, 31,646 ZCTAs. It joins ACS 5-year 2023 demographics, Zillow ZHVI/ZORI, HUD SAFMR FY2026, BLS county unemployment, IRS 2022–2023 migration flows and several new layers shipped in this round:

• FEMA National Risk Index — a composite risk score plus all 18 individual hazard scores (riverine flood, tornado, wildfire, hail, hurricane, earthquake, heat wave, drought and more), county→ZIP. Coverage 98.6% for the composite. v4 corrects the scaling of these layers — see The FEMA hazard fix below.
• Price-per-sqft — real median $/sqft from Realtor.com ZIP market trends, 90.5% coverage. It turns out to be a cleaner yield signal than headline home value (see below).
• Tornado counts — 30-year ZIP tornado counts from the NOAA SPC database 1950–2023.
• Deepened rent — a best_rent field blends ACS median rent, Zillow ZORI and HUD FMR to reach 100% rent coverage; the blend mixes market rent and policy rent and is flagged as such.

The FEMA hazard fix (v4) — a map that was lying

Every FEMA hazard layer on the ZCTA heatmap was wrong, and we shipped it wrong. The fix is worth documenting plainly.

v4 re-pulls the absolute, dollar-denominated NRI Risk Index Values — the genuinely right-skewed measure where most counties sit near zero and only truly hazardous counties are elevated. Those values are cube-root power-transformed (a monotone transform that keeps the right-skew — a log transform would flatten it back toward normal, re-introducing the very bug) and linearly rescaled 0–100 against the 99.5th percentile. The result is a map that tells the truth:

The wildfire median dropping from ~54 to 11.7 is the fix in one number: most ZIPs do not face meaningful wildfire risk, and the corrected layer now shows the West hot and the East calm — a right-skewed distribution with a long warm tail instead of a uniform wash. The composite fema_risk_score and fema_eal_score were corrected the same way.

Climate-adjusted yield (v4) — a screen, not a quote

A high gross yield in a hurricane-exposed ZIP is not the same as a high net yield. The climate-yield layer (zipdata/data/climate_yield.json, 29,734 ZIPs) estimates how much of a ZIP's gross yield is eaten by insurance:

Nationally the median insurance drag is 0.48% of value (about $1,268/yr), and the median ZIP loses 9.9% of its gross yield to insurance. But the spread is the story. Florida ZIPs lose a median 20.1% of gross yield to insurance — followed by Texas (16.4%), Louisiana (15.7%), Rhode Island (15.6%) and California (15.3%). A high-yield Florida ZIP can be a "high-yield-but-uninsured-risk" market: the headline number survives only until the carrier quote arrives.

The demand-anchor score (v4) — an honest null result

The demand-anchor score asks a reasonable question: does a ZIP's proximity to big, durable employers — military bases, colleges and semiconductor fabs — predict its rent growth? For 28,630 ZIPs we computed a distance-decayed sum (exp(−d/30mi), 90-mile cutoff) of nearby anchor "mass" across 6,236 anchors — troop count for bases, enrollment for colleges, market cap for fabs — weighted up for ZIPs near all three anchor types.

The rent forecast (v4) — making DSCR forward-looking

Spot DSCR is a snapshot — today's rent over today's value. The v4 rent forecast (rate/data/rent_forecast.json) projects where metro rent is heading so DSCR can be read 1 and 3 years forward. The model is a pooled AR(1) mean-reversion on year-over-year rent growth: g(t+h) = a + b·g(t), fit by OLS across all metro-month pairs (29,770 observations) and clamped to the empirical p1–p99 of observed growth. Resolution is metro level — ZIP-level monthly rent is too noisy to forecast honestly.

The appreciation forecast (R-MODEL.appreciation-forecast) — the value side

DSCR has two drivers — rent (numerator) and home value (denominator, through the loan). The rent-forecast above makes the rent side forward-looking; the appreciation forecast does the value side. Same model family, same diagnostics, same honesty: pooled AR(1) on YoY home-value growth, clamped to the empirical p1–p99 of observed growth, fit on 72,908 metro-month pairs.

The appreciation-forecast skill is real but smaller than rent's (+4.7%/+9.2% vs rent's +9.1%/+23.8%). Home-value growth is highly autocorrelated — sticky momentum means a naive "growth continues at its current pace" baseline is hard to beat. Beating naive by single-digit MAE points at the 1-year horizon, and by ~9% at 3 years, is the honest ceiling on what a pooled AR(1) on a single autocorrelated series will deliver.

Backtests — does any of this predict, or only describe?

WindMayor sells DSCR and Opportunity Sigma as present-tense screens — can a property cash-flow today, and is its yield statistically unusual relative to peers. Whether either has forward predictive value is a separate question. We backtested both, honestly, and report the result as-is.

DSCR backtest (R-VALIDATE.dscr-backtest) — an honest null

At each backtest date T we computed DSCR for 242–244 cities using only data known at T — point-in-time rent and home-value (from the rent_ts/home_value_ts arrays) and the FRED MORTGAGE30US rate prevailing at T. Then we measured the forward 5-year rent CAGR and value CAGR that actually arrived, and regressed forward growth on DSCR-at-T. Two windows: 2019→2024 (rate ~4.46%) and 2021→2026 (rate ~2.74%).

Sigma backtest (R-VALIDATE.sigma-backtest) — same null shape

Same methodology, same windows, same n. Sigma at T is reconstructed by taking point-in-time gross yield and robust-z-scoring it across the cross-section at that date — identical formula to compute_sigma.py, no lookahead.

Opportunity Sigma — where to look, not what to buy

Opportunity Sigma scores how unusual a ZIP's yield (or DSCR) is relative to its peers. It is a robust z-score: it uses the median and 1.4826 × MAD (median absolute deviation), not the mean and standard deviation. The reason is that ZIP-level yield has a long fat tail — a handful of $40k-home ZIPs post 90%+ gross yields. A mean/std z-score lets those outliers inflate the spread and quietly flatten everything else; the median/MAD version is not dragged by them.

R-SIGMA.normalize-v2. A reviewer caught a ZIP that scored +39.6σ — not a real anomaly but a Zillow ZORI scrape error ($27,865/mo at ZIP 28570) producing an "impossible" yield. v2 fixes the normalization in two ways:

• Artifact gate before z-scoring. Six ZIPs are removed entirely as physical impossibilities — not down-weighted, gated to NULL. The rules: gross yield > 33% (impossible for owned residential), home value < $25,000 (sub-rural noise), or ZORI rent > 6× the HUD FMR for the same ZIP (a scrape error or vacation-rental pricing). The six gated ZIPs and their reasons are written into sigma_data.json under artifact_gate — nothing is silently suppressed.
• Z-score in log space. Yield is a ratio metric; its distribution is right-skewed. v2 z-scores log(yield) rather than yield itself. After both fixes the maximum national yield σ is +4.33 (was +39.6), the robust center is 6.19% with a robust log-spread of 0.377, and 91 ZIPs sit above +3σ nationally. Those 91 are not artifacts — they are the genuine distressed-market fat tail (Detroit, East St. Louis, Jackson MS, the Mahoning Valley). The remaining 7 ZIPs above +4σ are inside that same tail; manual spot-check confirms each is a real low-price / real-rent observation, not a scrape error.

On the national yield base (n=14,956 after gating) the robust center is 6.19%; the log-space robust spread is 0.377. Band thresholds below apply on the log-yield z-score for the national surface; per-surface sigmas (cities, milbase, collegemap, seattle) keep their own robust scales.

What the ZIP data says — and the modest R²

We ran the full 31,646-ZIP spine through a correlation matrix, a yield regression and a k-means clustering. The findings are observational — correlation, not causation; nothing below controls for confounders.

Price-per-sqft beats headline price as a yield killer

Among the cleaner findings: price-per-sqft correlates −0.16 with gross yield. Per-sqft price captures land scarcity better than total home value, and is one of the strongest single negative predictors of yield.

Disaster risk tracks HIGHER home values, not lower

The most counter-intuitive result: the FEMA composite risk score correlates +0.34 with home value across 29,734 ZCTAs. The market is not discounting hazard — it is paying up for the same coastal and metro places that score risky.

One vintage note: this correlation matrix was computed before the v4 FEMA hazard fix described above. The cube-root rescale is monotone, so the sign and broad direction hold, but the exact coefficients should be re-run on the corrected absolute scale — a tracked follow-up.

k-means (k=6) sorts the spine into six archetypes — from High-yield / Cheap ($78k median home, 15.2% gross yield, 1,683 ZIPs) to Highly-educated / Expensive ($666k home, 4.1% yield, 4,890 ZIPs). Cluster boundaries are fuzzy; the names describe centroids, not every member.

New tools (v4)

Two interactive surfaces shipped this round; both read the same audited data the rest of the site uses, no separate pipelines.

The Buy-Box Screener

The Buy-Box Screener filters the full 31,646-ZIP spine against an investor's criteria — yield floor, price ceiling, FEMA risk cap, income, ownership rate and more. It is a pure client-side query over a slim ZIP extract; nothing on it is a recommendation. It is the screen made explicit: set your box, see which ZIPs fall inside it, then do the property-specific diligence.

The AI / Chip Stock Heatmap

The AI / Chip Stock Heatmap tracks 53 public companies across the AI and semiconductor supply chain — returns and metadata from the public Yahoo Finance chart API. It pairs with the fab map, which places the physical fabs and offices on a map. It is a context layer for reading where AI capital is concentrating — not a stock screen, and carries no valuations of private companies, only public-record HQ cities.

Stock-wealth geography (R-MODEL.stock-wealth-geography) — heavily caveated

When a public company's stock rises, the equity wealth lands with its employees — through RSUs, ESPP and options — who live near the company's offices. Stock appreciation is therefore a geographically concentrated injection of housing-demand money. The stock-wealth-geography surface estimates this per office, ZIP and metro across the 2021-01 → 2026-05 window.

Implied cap rate (R-MODEL.cap-rate) — the unlevered view

DSCR is a levered screen — it asks whether rent covers a real mortgage. The cap-rate page is the unlevered companion: how much yield does the ZIP throw off on the property itself, before financing? Cap rate = NOI / home value, scored on 30,002 ZIPs.

National median cap rate: 2.54% (mean 3.31%; p10 0.69%, p90 6.46%). 898 of 30,002 ZIPs show a negative NOI — modeled operating costs exceed effective rent. These are retained as a real signal that modeled carry is upside-down, not nulled. Cap rate is unlevered by definition; pair with DSCR for the levered picture.

What changed in v5 — data fixes, new fields, new models

v5 is the biggest round since v3. DSCR's formula constants are unchanged from v4, but v5 adds:

• Three data-quality correctness fixes (demand-anchor merge, ZIP 28570 rent scrape, collegemap DSCR stale)
• A structural split of the ZIP data file (zip_render vs zip_master) to cut the heatmap critical path from ~49 MB to ~2.6 MB gz
• Six new data fields (crash, pop density, broadband, schools, traffic, crime proxy)
• An artifact gate + log-space rebuild of Opportunity Sigma (R-SIGMA.normalize-v2) — documented in the sigma section above; max national σ collapsed from +39.6 to +4.33
• Two new models — implied cap rate per ZIP (R-MODEL.cap-rate, 30,002 ZIPs) and the appreciation forecast (R-MODEL.appreciation-forecast, pairing with rent-forecast to deliver a true forward DSCR)
• Two honest backtests — DSCR (R-VALIDATE.dscr-backtest) and Sigma (R-VALIDATE.sigma-backtest), both null-result findings reported as such
• The stock-wealth-geography surface and the dscr-positive / portfolio-insurance insight tools
• An independent source-traceability audit (T-DATA.source-spotcheck, 90 / 90 PASS) verifying the “real data, sourced, no fabrication” claim

zip_render / zip_master split (B-DATA.zip-master-bloat + slim_zip_render)

The 78 MB zip_master.json was too large for the heatmap to fetch on every page load. v5 splits it into two files and key-shortens the served one:

• zip_render.json — 9.3 MB (~2.6 MB gzipped over the wire), 31 fields for 31,646 ZIPs, key-shortened (each long field name has a 1–3 char alias; e.g. median_home_value_acs → hv). The full alias table is included as _keymap at the top of the file and decoded by the heatmap at runtime — so the served file is compact, but no fields are renamed in the pipeline. Tracked in git; this is what the site actually reads.
• zip_master.json — 78 MB, all ~120 pipeline fields, full long names. Gitignored; regenerated locally by the pipeline. The canonical intermediate product for data agents; not served to the browser.

Net effect on cold-load (B-PERF.heatmap-load): heatmap critical path went from ~49 MB to ~2.6 MB gz. Shape-first paint draws the ZIP polygons before the data file lands; data layers wash in as zip_render.json arrives.

Data-quality fix: demand_anchor_score was the wrong value (B-DATA-FIX-12)

The demand-anchor merge wrote anchor_score_raw (the raw distance-weighted sum) into the demand_anchor_score slot instead of the headline diversity-weighted value. The correct formula is: headline = anchor_score_raw × (0.4 + 0.6 × type_diversity / 3) — where type_diversity (0–3) rewards a ZIP that draws from all three anchor types rather than a single dominant one. Of 27,535 paired ZIPs, 27,450 carried the wrong value before the fix. The corrected values are now the source of truth in demand_anchor.json; the pipeline no longer lets a "never overwrite" guard silently suppress them.

Data-quality fix: ZIP 28570 corrupt rent (B-DATA-FIX-13)

ZIP 28570 (Morehead City NC) carried a Zillow ZORI best_rent of $27,865/mo — a 24.7× multiple of the HUD FMR ($1,130) and 26× the ACS median rent ($951). The value propagated into zip_render.json and cap_rate.json, producing a computed DSCR of 13.55 for that ZIP. In v5 the corrupt ZORI value is removed; best_rent for 28570 falls back to the ACS value ($951) and is no longer a heatmap outlier.

The fix is generalized at the pipeline level: the rent ladder now guards ZORI against HUD FMR — any ZORI value exceeding 15× FMR is dropped as a scrape error; ZIPs where ZORI runs 6–15× FMR are retained but tagged rent_quality = zori_high_vs_fmr (these are real luxury / vacation-rental markets such as Aspen CO, the Hamptons NY, and Westhampton Beach NY).

Data-quality fix: collegemap DSCR stale (B-COLLEGE.dscr-stale)

Zillow ZHVI and ZORI data for the 6,086 collegemap institutions were refreshed, but the stored dscr and dscr_condo values were not recomputed. A fresh run of the canonical v3 formula against current rent and home-value inputs found 987 SFR rows and 1,746 condo rows whose stored DSCR differed from the recomputed value by more than 0.10. All were corrected; a cross-check confirms 0 stale rows on both bases. The fix also recovered +113 collegemap DSCR rows (5,747 → 5,860) by recomputing off the all-homes price ladder, which covers more rural ZIPs than the condo-only basis the prior pass had narrowed to.

New fields in v5

New insight tools (v5)

Two new research surfaces published using the v5 data:

DSCR ≥ 1.0 Map

A full characterization of the 1,151 U.S. ZIP codes (4.4% of the 26,177 computable) where rental income covers all operating costs plus a 7.21% DSCR-loan debt service at 80% LTV. The finding: 95% of cash-flowing ZIPs have home values under $200k and are concentrated in Appalachia (WV, KY), the Rust Belt (OH, MI, IL, PA) and the Deep South (AL, MS, LA). A second category — the Hamptons, Fort Myers Beach, Kauai — shows DSCR ≥ 1.0 because ZORI captures vacation-rental pricing, a different investment thesis. Median DSCR among positive ZIPs: 1.28× at a $97k median home value with 12.2% gross yield.

Portfolio Insurance-Risk Tool

Paste any list of ZIP codes — a portfolio you own, a market you are underwriting, or a comparison set — and receive an aggregate FEMA risk distribution, total annual insurance drag in dollars, yield impact (gross vs. climate-adjusted), and a per-property table. Covers 29,734 ZIPs, fully client-side. Five preset portfolios: Rust Belt cash-flow ZIPs, coastal metros, Sun Belt, high-wildfire exposure, and hurricane alley.

Audit — real data, sourced, no fabrication 90 / 90 PASS

The most consequential claim WindMayor makes is that the numbers on screen come from real public data, traceable to source, with no fabrication. T-DATA.source-spotcheck (re-run 2026-05-22) is the independent audit of that claim. Methodology: pick a random 5-row sample from each surface (deterministic seed=2026), grab one or two cited numeric fields per row, and trace each back to the original public source — fetching the source file fresh during the audit itself for live APIs.

Across 11 distinct data surfaces and 7 independent public sources every stored value traced to source to the dollar, with the worst observed delta = 0.19% (a cap-rate row) explained entirely by stored 4-decimal rounding of NOI and home-value before division. Where a ZIP has no Zillow coverage, the row honestly cites a fallback (HUD FMR or Zillow City) instead of fabricating — and the fallback value is also verified.

Sources & attribution

• FRED (Federal Reserve Bank of St. Louis): public domain. fred.stlouisfed.org
• Zillow Research (ZHVI, ZORI): free public data, used with attribution. zillow.com/research/data
• IPEDS (NCES, U.S. Department of Education): public domain. nces.ed.gov/ipeds
• DoD BAH (Defense Travel Management Office): public schedule. travel.dod.mil/…/BAH
• Wikipedia infobox: Creative Commons Attribution-ShareAlike
• GlobalSecurity.org: public publishing, attribution provided
• Census ACS (U.S. Census Bureau): public domain
• BLS OEWS / CPI (Bureau of Labor Statistics): public domain
• OpenStreetMap / Overpass API: Open Database License (ODbL) — attribution + share-alike
• IRS SOI migration data (county-to-county): public domain
• HUD Fair Market Rents / SAFMR: public domain
• FEMA National Risk Index: public domain — composite + 18 hazard scores
• NAIC Homeowners Insurance Report (2022 data, released 2025): per-state HO-3 premiums, cross-checked vs Insurance Information Institute
• Tax Foundation — "Property Taxes by State and County": effective property-tax rates
• NOAA Storm Prediction Center: tornado database 1950–2023, public domain
• Realtor.com ZIP market trends: median price-per-sqft, attribution per fetch
• EPA AQS: Annual AQI by county, public domain
• Yahoo Finance chart API: public endpoint; attribution per fetch
• NHTSA FARS (Fatal Accident Reporting System): public domain — fatal crash counts 2021–2023 by county
• FCC National Broadband Map: public domain — broadband availability by location
• NCES EdFacts (U.S. Department of Education): public domain — state assessment proficiency rates SY2019–20
• FHWA HPMS (Highway Performance Monitoring System): public domain — annual average daily traffic counts by road segment
• County Health Rankings (Robert Wood Johnson Foundation / U. of Wisconsin): public domain — FBI UCR violent crime proxy, 2022

Disclaimer

WindMayor is a research and underwriting screen, not investment advice, not legal advice, not tax advice. Numbers are screens for further diligence, not closing-grade values. Past data is not prediction. Real deals require property-specific underwriting with current local comparables.