# PMA Scoring Methodology This document defines each dimension of the Public Market Analysis (PMA) score, its weight, the formula used to normalise it to 0–100, and the risk flag thresholds that trigger user-visible warnings. --- ## Overview The PMA score summarises affordable housing site viability in five dimensions: | Dimension | Weight | Signal | |---|---|---| | **Demand** | 30% | Affordability pressure, renter household share | | **Capture Risk** | 25% | Existing + proposed units vs. qualified renters | | **Rent Pressure** | 15% | Market rent vs. affordable rent threshold | | **Land / Supply** | 15% | Vacancy rate bands | | **Workforce** | 15% | Placeholder (future: LODES workforce data) | A higher score means **stronger market support** for an affordable housing project. --- ## PMA Tract Apportionment PMA aggregation selects Colorado census tracts whose TIGER bbox intersects the selected buffer, then apportions count fields by the tract polygon's share inside the circular buffer. The analytic share is computed from `data/market/pma_tract_display_geometry.geojson`, a lightweight derivative of `data/market/tract_boundaries_co.geojson`: tract rings are projected to local miles around the site, clipped against a 72-segment circle, and divided by the tract's projected polygon area. This replaces the older bbox rectangle approximation. If the display geometry artifact is unavailable in an offline session, the PMA engine falls back to the bbox approximation and discloses a data warning instead of blocking the page. ### 2026-07-17 Migration Impact The PR-B migration changed only the apportionment weights. It did not change which tracts are selected, score weights, thresholds, source vintages, or LIHTC supply logic. Representative 3-mile buffer checks run locally against `market-analysis.html` showed the following PMA score movement: | Site fixture | Market type | Old bbox score | New polygon score | Delta | Household delta | |---|---:|---:|---:|---:|---:| | Denver Civic Center | Urban core | 58 | 57 | -1 | -12,725 | | Lakewood / Belmar | Suburban | 65 | 62 | -3 | -12,621 | | Silt | Rural | 71 | 69 | -2 | -244 | | Breckenridge | Resort | 31 | 30 | -1 | -17 | The direction is mostly downward because rectangle intersections previously over-counted edge tracts. Resort/rural cases can move individual vacancy and workforce sub-scores when very large tracts are clipped more accurately. --- ## Dimension Definitions ### 1. Demand (30%) Measures unmet rental housing need driven by cost burden and renter household prevalence. ``` cbScore = min(100, (cost_burden_rate / 0.55) × 100) renterScore = min(100, (renter_share / 0.60) × 100) demandScore = cbScore × 0.60 + renterScore × 0.40 ``` Where: - `cost_burden_rate` = share of renters paying ≥30% of income on rent - `renter_share` = renter_hh / total_hh ### 2. Capture Risk (25%) Measures market saturation by comparing existing LIHTC supply to *LIHTC-eligible* renter households. **Lower capture ratio → lower risk → higher score** (head-room signal). ``` capture = (existingLihtcUnits + proposedUnits) / qualifiedRenters captureScore = max(0, min(100, (1 − capture / 0.50) × 100)) ``` **Where `qualifiedRenters` comes from (D, 2026-05):** The denominator is the count of renter HHs ≤80% AMI — the population federal LIHTC income limits actually cap. CHFA market studies underwrite to this pool, not to the full ACS renter base. 1. **Preferred**: `qualifiedRenters` = sum of CHAS renter HHs in tiers `lte30 + 31to50 + 51to80`, apportioned across the buffer's tracts by their F80 polygon-clip shares. CHAS data from `data/hna/chas_affordability_gap.json` (HUD CHAS 2018-2022 vintage). 2. **Fallback**: when CHAS is unavailable, falls back to ACS total `renter_hh` from `acs_tract_metrics_co.json`. This over-counts the LIHTC demand pool and under-states capture risk — surfaced as `pma_data_coverage.capture_risk: "partial"` in the result so renderers can disclose the proxy. 3. **Last resort**: 1 (literal one) when both are missing. Coverage = `"fallback"`. The `captureDenominator` field on the result object exposes the value used + the source (`chas_lihtc_eligible` / `acs_total_renter_hh` / `fallback_1`) + the per-tier breakdown for renderer disclosure. ### 3. Rent Pressure (15%) Measures the gap between market rent and the affordable rent threshold. If market rents substantially exceed affordable rents, demand is unmet. ``` affordableRent = (AMI × 0.60 × 0.30) / 12 # 60% AMI, 30% rule, monthly rentRatio = median_gross_rent / affordableRent rentScore = max(0, min(100, (rentRatio − 0.70) / (1.50 − 0.70) × 100)) ``` Where `AMI` = Colorado statewide Area Median Income. See [HUD Income Limits](https://www.huduser.gov/portal/datasets/il.html) for the current AMI value and vintage year. ### 4. Land / Supply (15%) Measures rental-market supply tightness via **STR-adjusted rental vacancy** (#1163/#1171). Very low rental vacancy signals unmet demand; 10%+ signals lease-up risk. **The input is rental vacancy, not total vacancy.** Total ACS vacancy (`B25004_001E`) counts seasonal and second homes as "vacant," which scored every Colorado resort county 0 on this dimension (Summit County: 63% median tract "vacancy") — reading the state's most workforce-housing-starved rental markets as oversupplied. Rental vacancy follows the Census Housing Vacancy Survey convention, then applies a seasonal-share discount as a proxy for the share of "for rent" units that are short-term/vacation product unavailable to the long-term rental market: ``` seasonal_share = Σ vacant_seasonal / Σ vacant (B25004_006E) (B25004_001E) # 0 when Σ vacant = 0 adj_vacant_for_rent = Σ vacant_for_rent × (1 − seasonal_share) str_adjusted_rental_vacancy_rate = adj_vacant_for_rent / (Σ renter_hh + adj_vacant_for_rent + Σ rented_not_occupied) landScore = max(0, min(100, (1 − str_adjusted_rental_vacancy_rate / 0.10) × 100)) ``` The seasonal discount is a proxy, not a direct STR inventory count: Colorado has no statewide STR registry, and local license counts are patchy and use a different universe than ACS "for rent at survey time." License counts are used only in a warn-only calibration benchmark (`npm run audit:str-discount-calibration`) and are never scoring inputs. **Both code paths use this same preferred input and the same 0.10 ceiling** (`PMAMarketScoring.RENTAL_VACANCY_CEILING`) — the historical 0.10-vs-0.12 divergence (#1149) remains resolved: - `scoreMarketTightness()` (`js/market-analysis-scoring.js`) — the default path. Buffer-level adjusted rental vacancy is derived from summed, buffer- share-apportioned counts (not averaged tract rates) in `aggregateAcs()`. - `scoreLandSupplyWithBridge()` → `scoreLandSupply()` (`js/market-analysis/site-selection-score.js`) — the Bridge-gated path. Its only remaining difference is the 60/40 land-cost blend; enabling Bridge/MLS access no longer changes the vacancy normalization. The 0.10 ceiling is underwriting-grounded: 10%+ vacancy is where lease-up risk and absorption pace materially threaten LIHTC underwriting. **Fallbacks**: - When tract data predates the `vacant_seasonal` field, scoring uses raw `rental_vacancy_rate` (`rental_vacancy` basis) at the same 0.10 ceiling. - When tract data predates all rental-vacancy fields (or a buffer has no rental universe at all), scoring falls back to the historical behavior — total vacancy at a 0.12 ceiling, suppressed input defaulting to 0.05 (neutral ≈ 58) — and the dimension note discloses `legacy_total_vacancy` as the basis. **Known limitation — residual STR contamination in resort cores.** ACS still does not identify short-term rentals directly. The seasonal-share discount mitigates the worst resort-core inversion, but it can under- or over-discount where seasonal housing and STR licenses diverge. Read adjusted resort-market scores alongside local STR-license and long-term-listing data. For example, a year-round-marketed Breckenridge STR can be counted as ACS "for rent" without also being counted as "seasonal," so residual STR contamination can remain after the seasonal-share discount. Buffers meeting **both** of the following are automatically flagged "STR-DISTORTED" in the dimension note (`PMAMarketScoring.isStrDistorted`, disclosure only): - `str_adjusted_rental_vacancy_rate ≥ 0.08` when present, otherwise raw `rental_vacancy_rate ≥ 0.08` for stale data (score ≤ 20 — materially depressed), and - total `vacancy_rate ≥ 0.25` (seasonal-dominated market, the STR tell). Residual flags mean the proxy-adjusted rate is still high in a seasonal- dominated market; they do not change the score. ### 5. Workforce (15%) Weighted composite score from up to 5 Colorado-specific data sources: | Sub-source | Weight | Data File | Module | |---|---|---|---| | LODES job accessibility | 25% | `data/market/lodes_co.json` | `window.LodesCommute` | | ACS income/education proxy | 25% | ACS tract metrics | (inline) | | CDLE vacancy rates | 20% | `data/market/cdle_job_postings_co.json` | `window.CdleJobs` | | CDE school quality | 15% | `data/market/cde_schools_co.json` | `window.CdeSchools` | | CDOT traffic connectivity | 15% | `data/market/cdot_traffic_co.json` | `window.CdotTraffic` | ``` workforceScore = lodesScore × 0.25 + acsWfScore × 0.25 + cdleScore × 0.20 + cdeScore × 0.15 + cdotScore × 0.15 ``` Each sub-source falls back to a neutral value (40–55) when its data module is unavailable. Coverage level is reported as `full`, `partial`, or `fallback`. --- ## Overall Score ``` score = demandScore × 0.30 + captureScore × 0.25 + rentScore × 0.15 + landScore × 0.15 + workforceScore × 0.15 ``` ### Score tiers | Range | Tier | Interpretation | |---|---|---| | 80–100 | **Strong** | Strong market support; high viability | | 60–79 | **Moderate** | Reasonable support; review risk flags | | 40–59 | **Marginal** | Limited market support; further study needed | | 0–39 | **Weak** | Weak signal; site may face absorption challenges | --- ## Risk Flag Thresholds | Flag | Threshold | Severity | |---|---|---| | High capture risk | capture rate ≥ 25% | ⚠ Warning | | High cost-burden pressure | cost_burden_rate ≥ 45% | ✕ High | | Elevated rent pressure | rent ratio ≥ 1.10 | ⚠ Warning | When no flag is triggered, a "No critical risk flags" ✓ OK flag is displayed. --- ## Capture-rate Simulator The capture-rate simulator (CHFA-style) computes: ``` captureRate = proposedUnits / renter_hh ``` Risk levels: - **Low**: capture < 15% - **Moderate**: 15% ≤ capture < 25% - **High**: capture ≥ 25% The AMI mix inputs (`30%`, `40%`, `50%`, `60%`, `80%`) allow users to test different unit mixes. When the sum of AMI-mix units is > 0, it overrides the "Total proposed units" field. --- ## Competitive Supply Share (absorption risk) **Not a capture rate.** Capture rate (above) divides units by a *demand pool* (income-qualified renter households). Competitive Supply Share divides supply by supply — the proposed project's share of the competitive inventory it will lease up against. The two move in opposite directions for the same market signal (more existing LIHTC supply raises capture rate but lowers supply share), so they must never share a label. Implemented by `calculateAbsorptionRisk()` in `js/pma-competitive-set.js`: ``` competitiveSupplyShare = proposedUnits / (totalCompetitiveUnits + proposedUnits) ``` Thresholds (`SATURATION_LIMIT = 0.10`): - **Low**: share < 5% - **Moderate**: 5% ≤ share < 10% - **High**: share ≥ 10% Rendered in the "Absorption Risk — Competitive Supply Share" card on the Market Analysis page. The result object's internal field name remains `captureRate` for backwards compatibility; only the user-facing label changed (#1148). --- ## Explainability Each dimension score is displayed as a bar (0–100) in the UI alongside the numeric value. Risk flags provide plain-language explanations tied directly to the threshold values in this document.