Field Property Partners

Methodology

How Field Property Partners pulls data, builds the pro forma, computes returns, and produces a defensible memo. Every page-1 tile is sourced from a single canonical row indeal_metrics_snapshot; every paragraph of generated prose is validated against that same row before render.

1 · Two Modes

A single data layer feeds two underwriting engines. Property type auto-detection on deal ingest routes each address to the right engine; the user can override.

Multifamily (5+ units)

  • NOI / cap-rate framework
  • Sales comparison + DCF reconciliation
  • European-style waterfall (70/30 LP/GP, 8% pref, 20% promote)
  • 5-year annual pro forma; interest-only debt by default

Residential (1–4 units)

  • AVM / ARV framework with HouseCanary forward AVMs
  • Year-1 monthly pro forma + 5-yr annual rollup
  • 30-year amortizing P&I (residential convention)
  • Strategy classifier: buy-and-hold / BRRRR / fix-and-flip / marginal
  • Multi-scenario evaluation — 1yr flip, 5yr hold, BRRRR refi math all computed

2 · Data Sources

ATTOM Data

Property metadata, sales comps, tax records.attomdata.com

RentCast

Rent AVMs + value AVMs (fallback when HouseCanary returns no-coverage). Foundation tier.rentcast.io

HouseCanary

Residential AVMs with 1Y / 3Y forward forecasts and confidence intervals. Teams tier.housecanary.com

Attribution:HouseCanary

FRED (Federal Reserve Economic Data)

Treasury yield curve (2Y / 10Y / 30Y) and SOFR. Used as the discount-rate calibration anchor and to sanity-check the user's interest-rate assumption against the current rate environment.fred.stlouisfed.org

Anthropic Claude

Claude Sonnet 4.5 — reconciliation prose generation and listing-facts extraction. Every generated number is validated against the snapshot row before render; retries automatically if the model hallucinates a figure.

CoStar

stub

Architected for enterprise integration. Thedata_providers/costar.pystub class implements the provider interface; activation requires a CoStar enterprise contract. See the Roadmap for what the product looks like with CoStar firm-level data.

3 · Underwriting Methodology

Multifamily

  • Line-by-line 5-year pro forma (GPR, vacancy, OpEx, NOI, debt service, CFADS)
  • Sales-comp adjustment grid: paste 3-5 comps, auto-defaults for location / condition / size
  • DCF with terminal value at exit cap, discounted at exit-cap + 1.5%
  • Sensitivity battery: exit-cap × rent-growth, price × exit-cap, rate-shift × hold
  • Three-paragraph reconciliation (Comps vs DCF, Deal vs Asking, Biggest Assumption)

SFR

  • Year-1 monthly pro forma + 5-year annual rollup
  • Strategy classifier (rules-based on cash-on-cash, DSCR, rehab ratio, ARV-to-basis)
  • Multi-scenario: 1yr flip math + 5yr buy-and-hold + BRRRR refi math, all computed regardless of current hold setting
  • HouseCanary forward AVM as exit anchor; RentCast fallback when HC has no coverage
  • Three SFR-specific paragraphs (Strategy Fit, Return vs Norm, Risk Factors)

4 · Property Tax + Insurance Calibration

  • County-level property tax: 120+ counties with verified effective rates (sourced from county assessor records). Fallback chain is county → state median → 1.5% national. Examples: SF County 1.18%, Travis (Austin) 2.10%, Maricopa (Phoenix) 0.65%, Cook (Chicago) 2.25%.
  • Insurance by market: coastal FL/LA/TX get flood + wind ($350–550/mo), CA fire zones get $400/mo, mountain west $120/mo. National average $150/mo.
  • Rent control flags: SF (2.5% cap), LA RSO (4%), NYC RGB (3%), NJ municipalities (4%), Oakland (2.3%), Berkeley (65% CPI), Santa Monica (75% CPI), plus 20+ others. When triggered, rent_growth default overrides to the local cap.
  • Closing costs by state: NY mansion + transfer (5.0% over $1M), NJ (3.5% over $1M), IL Cook (3.0%), HI (3-4%), DE (4%), CA / TX / FL (2.0-2.5%), default 2.0%.

5 · Validation Layer

  • Internal consistency: every metric on memo Page 1 (cap, IRR, equity multiple, DSCR) matches the corresponding metric in the Page 6 reconciliation prose within 0.5%. Enforced by a pre-PDF assertion; prose generation retries up to 2× if the LLM introduces a non-snapshot number.
  • AVM divergence flag:when HouseCanary AVM and purchase price differ by > 10%, Page 1 surfaces a non-blocking yellow info box pointing to possible condition / ownership / encumbrance factors.
  • HouseCanary confidence flag: when the AVM confidence interval spread exceeds 25% of the midpoint, the valuation is flagged as low-confidence and the user is directed to validate with comparable sales.
  • Strategy classifier multi-scenario:for SFR, the 1-year flip scenario, 5-year buy-and-hold scenario, and the BRRRR refi math are all computed regardless of the user's current hold-period setting. Whichever scenario shows the cleanest risk-adjusted return wins the classification — BRRRR outranks flip when both qualify.
  • Production E2E suite: real listing URLs run through the live HTTPS endpoint on every release. Last sweep: 33 URL parser variants + 5 production ingest E2Es covering Zillow, Realtor.com, LoopNet, and Crexi.