Military-Base Housing: The Data-Driven Thesis

The Underlying Asset

BAH has three properties that make it commercially interesting:

• Predictable — rates are set per MHA × paygrade × dependent-status × year and published in advance. Forward-looking, not lagging.
• Demand-anchoring — near a base, a meaningful slice of rental demand is effectively underwritten by the federal government via BAH.
• Scattered + access-hostile — the source data is fragmented. The moat is assembly, not the raw facts.

Four Angles

Angle 1 — BAH-Anchored Rental Underwriting ("Rent Floor")

An address near a base → resolve MHA → show BAH by paygrade + adjacent base branch/type/troops → underwriting view: BAH-implied rent band vs. asking rent.

• Feasible on free data? Yes for BAH/base; the ZIP→MHA crosswalk is public but must be sourced separately.
• Risks: Regulatory (needs disclaimer if advice-adjacent); data-quality (BAH resets annually — must surface year+source); BRAC concentration.
• Standalone monetization: High — converts a raw rate into an underwriting decision for SFR investors and RE syndicators.

Angle 2 — PCS-Cycle Timing Signal

PCS (Permanent Change of Station) season concentrates housing turnover near bases in roughly the mid-year window. A demand-index by market/month.

• Feasible? Partially — base/troop side yes; seasonality is directional, not precise.
• Recommendation: Fold into Angle 1 as a feature; not viable as standalone.

Angle 3 — Public-Equity / Base-Exposure Screen

Map publicly traded SFR/defense-adjacent REIT operators to base concentration via troop_count as BAH-demand proxy.

• Feasible? Medium-low — company portfolio data is the binding constraint (10-K property schedules, unstructured).
• No tickers or returns asserted — category-level only until real filing data is ingested.

Angle 4 — The Sourced Dataset Itself, as a Product

Package the assembled store (installations + BAH series + troop counts), every row sourced or explicit NULL, as a queryable screener / downloadable dataset / API.

• Feasible? Yes — zero additional data beyond what Y ships.
• Moat: "We did the painful aggregation of bot-walled, scattered DoD sources into one clean, NULL-honest, sourced store."

Ranked Shortlist → Future Tickets

All economic characterizations are [ASSUMPTION] pending real comp/market research. No figures fabricated.

1. Closure Risk (BRAC)

US base closures run through the Base Realignment and Closure (BRAC) process — independent commission rounds in 1988, 1991, 1993, 1995, and 2005. A closed or realigned base can collapse the BAH-anchored tenant base that the entire MILBASE thesis depends on. This is the single largest tail risk.

No fabricated closure probabilities. Instead, a reasoned indicator framework — directional, evidence-based, labeled [JUDGMENT]:

Output rule: every base gets a closure-risk tier (Low / Moderate / Elevated) with a one-line reason citing which indicators fired — never a fabricated percentage.

2. Geographic / Physical Risk

FEMA National Risk Index is the recommended spine — one free dataset gives a composite + per-hazard score per county, joinable on installation county FIPS.

3. Area Trajectory (Up vs. Down)

4. Known Data Gaps

• Troop count time series — current schema holds a point value; force-trend (strongest closure indicator) needs history.
• Mission / tenant-unit per installation — needed for single-mission concentration; not yet in schema.
• Regional economic dependence — derivable from BEA/BLS but not yet sourced.

Sources (all free, authoritative)

• DoD / BRAC Commission final reports (1988, 1991, 1993, 1995, 2005) — public domain
• FEMA National Risk Index; FEMA NFHL — free, public domain
• USGS seismic hazard; NOAA climate/hurricane — free, public domain
• Census ACS / PEP / Building Permits Survey — free, public domain
• BLS LAUS & QCEW — free, public domain
• FHFA HPI; Zillow ZORI — free

A defensible 0–100 ranking of each base as a rental-investment target. Decomposed into three sub-scores: Yield, Stability, and Risk. This is a research spec — implementation is a separate ticket.

1. Feature Set

Numerical

Categorical → Tiered

Tier→multiplier values are [ASSUMPTION] — chosen monotonic and round for transparency, tunable once real data exists.

2. Normalization

• Continuous features: min–max to [0,1] across the populated base universe (not z-score). Robust min–max: clip to 5th/95th percentile before scaling.
• Directionality: invert cost-like features so higher = better (e.g. rent-to-BAH: lower is better → use 1−norm).
• Missing values: do not impute. A NULL feature is excluded; sub-score renormalized over present features; coverage% reported.

3. Formula

4. Missing Inputs (→ Data Tickets)

1. Who Uses It

2. The One-Screen Deliverable

┌───────────────────────────┬──────────────────────────────┐
│  LEFT (Explorer)          │  RIGHT (Map)                  │
│  • search/sort base list  │  • US map, size = troop_count │
│  • BAH-over-year chart    │  • click dot ↔ list (postMsg) │
│  • RE panel: price/rent/  │                               │
│    price-to-rent ratio    │                               │
├───────────────────────────┴──────────────────────────────┤
│  DECISION PANEL (selected base)                           │
│  SCORE 0–100  ·  Yield __ | Stability __ | Risk __        │
│  Risk flags: closure / hazard / trajectory (tiers+why)    │
│  Coverage __%  ("NULL — reason" when data thin)           │
└───────────────────────────────────────────────────────────┘

The screen answers one question end-to-end: revenue (BAH) vs. cost (RE panel) → ratio → score → the catch (risk flags) — one decision path, no dead ends, honest NULLs.

3. MVP vs. Later

MVP (buildable on current/near-term tickets)

• Map + searchable/sortable base list (shipping)
• BAH chart with 14-year history (shipped)
• RE panel: price, rent, price-to-rent (G-re-panel + Y-city-econ)
• Composite score + Yield/Stability/Risk vector + coverage%
• Risk tier flags with reasons — honest partial coverage, NULL-with-reason

Later (post-MVP backlog)

• Validated thesis results surfaced in-app (THESIS-VALIDATION, after backtest runs)
• Watchlist / alerts on score or BAH changes
• Dataset/API product (Opportunity Angle 4)
• Public-equity base-exposure screen (Opportunity Angle 3)

4. Decision Path

1. Map = discovery. Big dots = large demand. Find submarkets.
2. BAH chart = revenue durability. Is the rent floor rising?
3. RE panel = cost. Price, market rent, price-to-rent headline.
4. Score vector = synthesis. Yield + Stability + Risk gate.
5. Risk flags = disqualifier check. Elevated closure → exit.

5. Non-Goals

• Not investment advice — data + transparent score, with disclaimer.
• Not a transaction/marketplace.
• No fabricated coverage — sparse data shows as NULL+reason, by design.
• No CDN/online deps — all offline-served.

1. The Landscape

A. BAH Calculators / Lookups (the crowded part)

B. BAH Maps / Visualizations (sparse)

• mpyne BAH map (github.com/mpyne-navy) — open-source map visualizing BAH geographically. Closest visual analog to MILBASE's map. Gap: visualization only — no RE overlay, risk, or screen.

C. SFR / Military-Housing Investing (separate world)

• Institutional SFR REITs — own/operate SFR at scale; some Sunbelt/base-adjacent exposure. [CATEGORY]
• Privatized base housing (MHPI operators) — on-base housing under DoD contracts. [CATEGORY]
• Generic RE screeners — comps/rent/yield, not BAH-aware.

These have the investment lens but no BAH anchor and no base-closure/risk model.

2. The Honest Verdict

3. Whitespace (where to push)

• The screen/ranking ("N of 459 by BAH-yield, risk-gated") — nobody packages this. Highest-differentiation surface.
• Time-aligned BAH-vs-local-RE (2013–26) — calculators are point-in-time; the trend join is unoccupied.
• Closure/risk-adjusted view — absent everywhere.
• Avoid competing as "a better BAH calculator" — that lane is saturated and not the edge.

Sources: military.com · veteran.com · collegerecon.com · navyfederal.org · travel.dod.mil · github.com/mpyne-navy. Category-level claims labeled [CATEGORY]; no figures, tickers, or features fabricated.

Honest Framing

1. Inputs

2. Model Formula

Band for display: 0–33 Low · 34–66 Moderate · 67–100 Elevated — same 3-tier vocabulary as RISK.md, now with a documented score behind it.

3. Mandatory Labeling

4. Gaps → Data Tickets

• troop_count time series — force-trend input; until then that input is NULL and index renormalizes (lower confidence, flagged).
• BRAC round membership table — small Y/research ticket from public DoD reports.

1. Ranked Signal List

2. Graph Composition — What Goes Together

Only co-plot series that share a unit and a meaningful comparison. Ratios, dollars, and counts do not belong on one axis.

3. Ratio-Rank Rules

• Rank is N of M where M = bases with a real value (NULLs excluded from M, shown "unranked — no data", never last-by-fiat).
• One rank per signal; card shows rank + percentile; master chart sortable.
• Never invent an order for NULLs.

Six named-view derived formulas that power the curated scatter comparisons. Exact fields from milbase.db; units; sane ranges; data status. NULL propagates — never imputed.

Rules

• NULL propagates: any NULL input → result NULL + reason; never 0, never imputed.
• Out-of-range values surface as data-quality flags, not silently clamped.
• Global-scale invariant applies to any cross-base comparison.
• Named-views show the formula + source/year inline (auditability).

DSCR (Debt Service Coverage Ratio) = NOI / annual debt service. The lender test for a leveraged SFR buy — how many times rental income covers the loan payment. DSCR < 1.0 means the property does not self-fund the loan.

Formula (exact)

Interpretation

• DSCR = 1.0 → breakeven: rental income exactly covers debt service.
• DSCR ≥ 1.20–1.25 [ASSUMPTION] = common lender minimum. Rendered as the practical "financeable" band.
• Breakeven rent / breakeven price: solve DSCR = 1.0 for each, holding the other — the actionable underwriting output.

BAH-Floored Variant

A second DSCR uses min(median_rent, BAH_E5dep) as income — the conservative, thesis-relevant variant: "what if only the BAH floor rents?" Primary for the investment thesis; never overclaims the full market rent.

Rate Source

FRED MORTGAGE30US — Freddie Mac 30-yr fixed; free fredgraph.csv, no API key required, official weekly data. Latest value stamped with observation date.

Rules

• Any NULL input → DSCR = NULL + reason; never imputed (HARD RULE).
• All five assumptions (vacancy, opex, LTV, spread, lender-min) are labeled, tunable inputs — not hard constants.
• Coverage% reported; "DSCR pending data" shown where area_econ absent.

BAH "growth" can be measured over four time windows. Default locked at trailing-5yr CAGR; all four available as a toggle. Canonical series: paygrade='E5', with_dependents=1.

Window Definitions (exact)

NULL if required prior year absent for that MHA — no fabrication, no proxy year.

Pros / Cons

Recommendation

Primary = trailing-5 CAGR. The thesis is a multi-year buy-and-hold demand-floor argument; a 5yr CAGR is the honest middle — stable enough to rank on, not distorted by one reset, lag acceptable for an asset held years.

Show trailing-3 alongside as the acceleration/deceleration companion. Never display a window unlabeled.

Officer-Average Variant

Officer rates read upper-tier demand. Label distinctly from the E5 default; never blend on one axis (HEURISTICS axis-compat rule).

Deeper than rent/price alone — a livability and demand-depth profile for the feeder and adjacent towns around each base. Answers: is the area attractive and growing independent of the base itself?

Geography: reuses the RE-DATA-SPEC feeder-set (feeder_city ∪ ≤25mi Gazetteer places, capped, tie-broken by distance then FIPS). Same join keys — one geo effort, many metrics.

Metrics and Sources (all free)

Demand-Depth Profile

• Growing & diversified (pop ↑ + low unemployment + multi-industry) = demand cushion beyond the base → lower closure-sensitivity [JUDGMENT].
• Base-dependent monoculture (employment concentrated in defense) = amplifies BRAC tail risk.
• Supply pressure (permits ↑↑ vs population flat) = rent-growth headwind.

Each component cites source, year, and coverage%. NULL where unsourced — never imputed.

Handoff

Spec for a surround(base_id, metric, value, year, source) long table — yellow writes per the spec. Feeds the investment thesis (livability/demand depth) and Risk Framework trajectory vector.

Local-Area Economics: Free Authoritative Sources

Establishes the profit input — what housing actually costs near each base — to compare against BAH (the demand-anchored revenue floor). Research only; no fabricated figures.

Recommended source stack

Verification finding: HUD FMR and Census ACS are free but require free registration — keyless requests return auth errors (first-hand verified 2026-05-16). Plan a credential step before the Y scrape ticket.

Derived signals (formulas)

• rent-to-BAH = market_rent / BAH_rate — <1.0 means BAH exceeds market rent (structural landlord margin); >1.0 means thin/negative cushion
• price-to-BAH = median_home_value / (BAH_rate × 12) — lower = faster BAH-financed payback
• BAH-implied gross yield = (BAH_rate × 12) / median_home_value
• BAH-implied cap rate ≈ (BAH_rate × 12 × (1 − opex_ratio)) / median_home_value — [ASSUMPTION] opex_ratio exposed as tunable

Geography join keys

All external sources key on ZIP / county-FIPS / CBSA. Critical dependency: a ZIP→MHA crosswalk is the single highest-leverage missing join — it unlocks rent, price, and underwriting in one step. Recurring dep across AREA-ECON, SCORING, UNDERWRITING, NEIGHBOR-ECON (9× flagged as of 2026-05-17).

Scoring: Exact Implementable Spec

Every judgment call from SCORING.md is locked here. Green / yellow can implement with zero discretion. Constants are [ASSUMPTION] but concrete and final for v1.

Locked decisions

Feature list

Formula

Normalization: robust P5/P95 clip → (xc−lo)/(hi−lo); direction-− features invert with 1−n. Weights are [ASSUMPTION] — renormalize over present features only. NULL score (with reason) if any sub-score below 60% coverage floor.

Historical Thesis Validation — Design (Pre-Registration)

Thesis under test

Buying residential property in markets adjacent to bases with a high BAH-to-price ratio outperforms a baseline, on a risk-adjusted basis, over multi-year holds.

Hypotheses (pre-registered)

• H1: Top-quartile-by-bah_yield base markets (at time t) beat an equal-weight all-base portfolio over horizon h.
• H0: No outperformance after controlling for region and national housing cycle.
• Falsification (committed up front): if H1 does not beat all three baselines net of region/cycle controls across a majority of rolling windows → thesis rejected. No goalpost moving after seeing data.

Design

• Signal at t: bah_yield = BAH_E5dep_t × 12 / median_home_value_t — rank, take top quartile.
• Annual rebalanced cohorts; horizons: 1, 3, 5-year rolling holds.
• Return proxy: FHFA HPI appreciation + min(BAH, market_rent) income; opex_ratio [ASSUMPTION] swept (0.35–0.50), never a single constant.

Baselines (must beat all three)

1. All-base-markets equal weight (isolates selection)
2. National housing (FHFA HPI / Case-Shiller — isolates cycle)
3. Random base-market bootstrap draws (isolates luck)

Key confounders

Does the Thesis Generalize to All Government Buildings?

Tested honestly — not force-fit. Research only; no fabrication. Verdict is allowed to be — and largely is — "distraction."

The military edge requires both:

• (A) Published, location-indexed, guaranteed housing floor — BAH is a per-paygrade cash housing allowance, set by MHA, federally backed. Rare and differentiating.
• (B) Concentrated, stable, relocation-driven tenant demand — thousands on-station, PCS-cycle turnover, single dominant employer.

B-only (generic stable employer) is a commodity signal with no edge. A is the rare one.

Class-by-class test

Honest finding

The strongest civilian analog of the BAH mechanic is not a government building class — it is the HUD Housing Choice Voucher / Fair Market Rent system: a published, location-indexed, government-backed rent payment to landlords. Structurally closest to BAH. Already in scope via AREA-ECON.md.

Recommendation

Do not open a "government buildings" product line. The defensible expansion is the HUD voucher/FMR family (already underway), optionally annotated with OPM FedScope single-employer concentration as a secondary stability factor — not a standalone product. This was worth testing precisely so it can now be deliberately set aside.

Personnel-by-Paygrade Estimation Model

Formula

Confidence bands

Presentation rule: always rendered as "≈ modeled estimate (method: troop_count × DoD rank distribution; not measured)" with source + confidence band inline. Feeds SCORING only as a modeled feature, weighted below measured fields.

DB grounded (as of 2026-05-16): 84/459 installations have sourced troop_count; 375 NULL = free-data ceiling (DoD doesn't publish clean per-installation counts). See Methodology page for full ceiling detail.

Deployment-Tempo Estimation Model

Why it matters: time-deployed proxies tenant turnover/absence near a base → feeds RISK (area trajectory/stability) and the rent-floor durability argument in OPPORTUNITY and underwriting.

Formula

Confidence bands

Sources: GAO reports on deployment/dwell time · CRS reports on military OPTEMPO/end-strength · DoD posture statements (all public). Implementation = later yellow/green ticket with mandatory "modeled / historical-average" label.

Base-Town Area Sentiment: Methodology

What we're measuring

Public perception of the town/area a base sits in (not the military) — livability, housing, safety, schools, economy — as a risk/quality modifier feeding RISK (trajectory) and SCORING (Stability).

Signal design

Rule: every sentiment claim is anchored to a quantitative public stat where one exists. Public sources only (Census ACS/PEP, BLS LAUS, FBI UCR, public local news, public forums — ToS-respecting); no auth-walled or private data.

Source classes

Follow-on: P/Y-MILBASE.sentiment-pull — execute the bounded public pull under explicit human "go", rate-limited, ToS-respecting, sources logged. Feeds RISK §3 trajectory + SCORING Stability once a real pull exists.

Neighboring-Area Economics

Extends AREA-ECON.md from the base ZIP to the feeder area — the adjacent ZIPs/towns that actually house service members. Underwriting on the base ZIP alone is biased; members spread across a commuting radius.

Feeder set definition (locked)

For each base: feeder_set = {feeder_zip} ∪ {ZIPs in feeder_county} ∪ {ZIPs adjacent to feeder_zip}. Adjacency via Census ZCTA boundaries (free, public domain). DB hooks: installations.feeder_city, feeder_county, feeder_zip.

Blended price-to-rent (the deliverable signal)

The delta between blended_p2r and base-ZIP-only p2r is the output of value — it shows how much the feeder-area picture differs from the centroid.

Output spec

Specs a neighbor_econ(base_id, blended_rent, blended_value, blended_p2r, coverage_pct, source, year) view — yellow writes; pink specifies. Feeds SCORING Yield (replaces base-ZIP p2r with blended) and G-ratio-rank.

Free Signal APIs & Datasets

Reference table of all free (no/low-key) APIs usable to estimate real estate, market/equity, and demand signals. Verification status labeled per row — no sample values invented.

Legend: V✓ first-hand verified · V-key free key required (keyless rejected) · V✗env keyless but this env's fetch tool was blocked · K stable public knowledge

Real estate

Market / equity

Demand / personnel

Recommended stack

• RE: FHFA + Zillow CSV (keyless) for momentum; HUD FMR + Census ACS (key) for levels — system of record.
• Market: Stooq/Yahoo for prices (keyless, direct client); SEC EDGAR (keyless + UA) for fundamentals; FRED (key) for macro.
• Demand: BLS (optional key) + Census BPS + DoD/DMDC.

Base Founding Year: Sourcing Spec

Defines the sourcing methodology and discipline for installations.established_year + established_src. Bulk per-base population is a yellow data ticket — pink specifies, yellow writes. No fabricated years.

Source hierarchy (per base, highest-trust first)

1. Official .mil base fact sheet / installation history page — primary
2. DoD / service historical office publications
3. Wikipedia infobox "Built/In use" — only with its own cited reference; record the underlying source, not "Wikipedia" alone
4. Reputable historical references (service museums/registries)

Hard rules

• No fabrication, no estimation. No credible source → established_year = NULL, established_src = NULL. Partial real > fake.
• Definition: established_year = year the installation began military operation. Renames do not reset the year (e.g., Fort Liberty was formerly Fort Bragg — the clock starts at Bragg's founding).
• Joint bases: use the earliest constituent installation's operation year; note the merger year in source string.
• Conflicting sources: take the official .mil value; flag established_src with "(conflicting sources; .mil used)".

Display: years_around = current_year − established_year (NULL if year NULL). Feeds the All-Bases scatter "years around" sortable axis; NULLs sort last.

Rent / Price Engine: Exact Source + Method Spec

Pink writes the spec; yellow scrapes per it. Implementable with zero judgment calls. Real source per row or NULL — no fabrication.

Feeder-town set per base (deterministic)

Metrics per place, per year 2013–2026

Priority: ACS for level → Zillow for years ACS lacks → HUD FMR as county fallback. Stamp the actual source used per row. Years with no source → NULL row, not fabricated.

Target table (yellow writes)

Time alignment: 2013–2026 matches the BAH series span so price_to_rent is joinable to bah_rates.year for blended ETF/ratio-rank signals. Any base/year missing on either side → excluded from blended signal, disclosed, never guessed.

Ship report fields (yellow): # bases with ≥1 year, median coverage%, per-source row counts, NULL count, years covered, weight-source used.

The product is the assembly + provenance discipline, not the raw facts. Public data, painfully aggregated, every row sourced or explicit NULL — sold as a clean, queryable, attributed dataset.

Three Export Tiers

Formats

• Bulk: versioned CSV + SQLite snapshot + JSON (data.json contract, extended)
• Manifest: row counts, coverage%, per-table as_of, source list, schema version
• API (later ticket): read-only REST; every record carries its source or explicit null

Provenance Contract (the moat)

• Every fact row: non-empty source OR explicit null + *_src=null
• manifest.coverage_pct per table published with the data — buyers see exactly what's real
• Versioned + changelog; never silently mutate a prior release

Licensing

• Federal source data (HUD/Census/FHFA/DoD) = public domain → redistributable
• Zillow/Redfin-derived signals = excluded from redistribution; Pro tier flags these fields as non-redistributable
• LICENSE + SOURCES file enumerating every upstream + its terms ships with every export

Near a base, BAH is a federally-set, published rent floor for a large tenant slice. Convert that into a one-screen underwriting view.

Input

• Address or ZIP or base select
• Resolves → MHA via ZIP→MHA crosswalk (dependency: backlog #1)
• Paygrade + dependent-status selector (default E-5, w/dep — canonical)

Output (six blocks, one screen)

Hard Rules

• Every number traces to a real source or shows "no data" — never a fabricated estimate
• Sparse inputs → explicit NULL, not a guess; degrades to "data pending," not fake
• Not advice. Disclaimer banner required on every render
• Single global scale for any BAH visual (GLOBAL-SCALE INVARIANT)

A time-boxed feasibility spike: can we map public companies/REITs to base-anchored, BAH-backed tenant demand using milbase.db + free SEC filing data?

What We Have (Free)

• Base side: locations, troop_count, branch — solid, in milbase.db
• Company side: SEC EDGAR full-text search + filings API (free) — 10-K/10-Q property schedules, but unstructured prose/exhibits

Binding Constraint

Mapping a public operator's properties to base proximity requires parsing property lists from filings. No free, structured, comprehensive property-by-issuer dataset exists.

Go / No-Go Criteria

• GO if ~10-issuer manual probe shows property locations reliably extractable from EDGAR filings AND base-proximity produces a differentiated signal vs. a naive Sunbelt proxy
• NO-GO if extraction is bespoke per issuer (no pattern) or signal collapses into "REITs in growth metros" (no military-specific alpha)

Raw per-MHA BAH growth treats a 500-troop post and a 50,000-troop installation equally. For a demand thesis, the question is how fast the BAH-backed tenant-dollar pool grows — which is troop-scaled.

Recommended Formula (troop-weighted BAH CAGR)

Why It Matters

An area where the large bases' BAH is rising fast has a deepening guaranteed-rent pool — the core demand signal for scoring, not distorted by tiny posts.

NULL / Coverage Discipline

• troop_coverage% = (troop-present bases) / total, reported per area
• Zero troop_count in area → metric = NULL + reason (no fabrication)
• 84/459 current ceiling — dominant gap; troop enrichment is the follow-on ticket