Valuation methodology
How Twellie prices a home — and how we compare to the rest of the industry.
No black box. We publish what the report uses, how confidence bands are formed, and where the output must be verified by appraisal, inspection, and source freshness. Updated June 2026.
The three approaches every valuation uses
Most credible US property valuations combine some subset of three classic approaches: comparable sales, cost, and income. Twellie uses that structure as decision support, not as a substitute for a licensed appraisal.
Approach 1
Sales Comparison
Find recent sales of similar homes nearby. Adjust each one for differences (more sqft, newer roof, no garage, etc.) and take a weighted average. The dominant approach for residential.
Used by: Every appraiser, every AVM. The Fannie Mae 1004 / new UAD 3.6 form is built around this.
Approach 2
Cost
What would it cost to rebuild this home from scratch today (replacement cost), minus depreciation, plus the value of the land? Used mostly for unique/new properties without good comps.
Used by: Insurance, new construction. Less common for resale unless there are no comps.
Approach 3
Income
What rent could this home command? Multiply annual rent by the local Gross Rent Multiplier (GRM), or use a discounted-cash-flow / cap-rate model. Standard for investment properties.
Used by: Investors, lenders, iBuyers. Cross-checks comp-anchored values.
Twellie uses four models, not one.
A single approach is fragile: comp-anchoring breaks down in rapid markets, regression weakens on unusual properties, and last-sale data goes stale. So Twellie runs four signals in parallel, then widens confidence when they disagree.
| Model | Weight | What it does | When it's right / wrong |
|---|---|---|---|
| Sales Comparison USPAP-grade | 45% | Weighted median $/sqft of similar comps within 0.5–2 mi, then 8 per-feature $ adjustments (size, beds/baths, lot, age, garage, pool, condition, market time). | ✓ Strong when 5+ recent comps exist · ✗ Breaks down in rapid markets and unique properties. |
| Hedonic Regression ML on features | 25% | Multivariate model on 14 features (sqft, beds, baths, year, lot, walk score, school rating, etc.). Currently uses national-median coefficients; refits per-CBSA when MLS data is wired. | ✓ Catches feature mispricings the comp set hides · ✗ Needs training data (we use public-records baselines). |
| HPI-Adjusted Last Sale FHFA index | 15% | Subject's last recorded sale × FHFA House Price Index for the metro since that sale. Effectively asks: "if everything else was equal, what would inflation alone make this worth?" | ✓ Strong for unrenovated owner-occupied homes · ✗ Wrong when major remodel happened post-sale. |
| Income (GRM) Investor cross-check | 15% | Comparable rentals × Gross Rent Multiplier for the submarket. Confirms the price isn't insane from an investor underwrite (cap-rate sanity check). | ✓ Anchors against the rental market · ~ Reduced weight on luxury / non-investible. |
How the four numbers become one
Weighted by model suitability and available data. Large outliers can be rejected or down-weighted, and the confidence interval widens when the signals disagree. Tight agreement means a narrower range; thin or conflicting evidence means the buyer should be more conservative.
What sets Twellie apart.
Most consumer AVMs ship a single number with no audit trail. Twellie's report shows every model, every adjustment, every weight — so you can act on it, not just look at it.
Capability scorecard
Engine status
The live report separates what is actually configured from what is still demo or roadmap: comps/listing source, HPI support, photo condition analysis, flood data, rental inputs, and back-test status.
Four-model ensemble
Sales-comparison + hedonic regression + HPI-adjusted last sale + income (GRM). The report shows which signals ran, which were thin, and how disagreement affects the confidence range.
Photo condition grading
When current listing photos are available, the condition layer grades visible rooms and turns obvious condition signals into adjustment notes. It cannot replace an inspection or detect hidden defects.
Per-adjustment audit trail
Every $ delta carries a label, an amount, and a one-sentence rationale. Disagree? You can re-do the math on the page. Most consumer AVMs ship a black-box headline; we ship the work.
Confidence interval — explicit
The width of the band comes from inter-model spread: tight agreement → narrow CI → high confidence. The headline number is one point in that range, not the whole story.
USPAP-informed adjustments
The feature adjustments (sqft, beds/baths, lot, age, garage, pool, condition, market time) are appraisal-informed and surfaced line by line so buyers can inspect the reasoning.
50-state coverage
155M+ US residential addresses with state-specific tax rules, transfer-tax estimates, exemption logic, and closing timelines built in. No metro-only restrictions.
The data behind every number.
A valuation is only as good as the inputs. Here are the sources we read for every report — every live valuation surfaces the source per-adapter, so you know what's behind the number.
Geocoding
US Census Geocoder
Address → lat/lng for every US residential property.
House Price Index
FHFA HPI (CBSA-level)
Federal House Price Index for inflation-adjusting the subject's last recorded sale.
Mortgage rates
Freddie Mac PMMS
Live 30/15-year fixed rates for true cost-of-ownership modelling.
Flood + climate
FEMA NFHL · climate models
Flood Hazard Layer for SFHA designation; climate risk overlay by ZIP.
Listings + comps
Listing/comps adapters + public records
Property attributes, recent sale evidence, and photos where available. Each report labels whether the source was live, hybrid, or demo.
AI vision
photo condition model
Per-room condition grading mapped to NAR Cost-vs-Value $ deltas.
Walkability
Walk Score · GreatSchools
Walk/transit/bike scores and school ratings for the neighborhood profile.
Tax + deed records
County records · ATTOM
Assessed values, last-sale prices and dates, exemption history.
Rental comps
RentCast · submarket-median rent
Income-approach inputs for the GRM cross-check.
How you can verify our number on your own.
Don't take our word for it. The fastest sanity check on any AVM: run the same address through 3 sources and look for spread.
Step 1: Spread check
Run the address through Redfin, Zillow, and Twellie. If all three are within ±5% of each other, the number is solid. If one is wildly different — that's the one to question.
Step 2: Open our adjustment ledger
Every $ delta we apply has a label, an amount, and a one-sentence rationale. Disagree with one? Manually adjust the price by that delta. The math is on the page so you can re-do it.
Step 3: Cross-check with the appraisal
For purchases involving a mortgage, your lender orders a licensed appraisal. Compare that number to ours — if they're within 5%, you have triangulation. The lender's appraisal is the legal benchmark; Twellie's job is to make sure your offer math holds up before you write it.
How we keep getting more accurate.
A valuation engine isn't a static thing. Three continuous-improvement loops compound over time.
Licensed comp-feed expansion
Roadmap work is to add more licensed listing and closed-sale feeds, then label freshness and source coverage per report so buyers know when evidence is strong versus thin.
Per-CBSA hedonic regression
A separate hedonic model fit for each of the top 50 Core-Based Statistical Areas — covering ~80% of the US population. Local features (water view in Seattle, lot size in Texas) get the right weight in the right market.
Continuous back-test loop
Every prediction is scored against the eventual sale price ~60 days after closing. The loop detects bias by submarket and feeds calibration updates back into the ensemble — the same instrumentation pattern lender-grade AVMs run.
A note on what Twellie is. Twellie is a hybrid AVM with USPAP-informed methodology — not a USPAP-compliant licensed appraisal. For purchases involving a mortgage, your lender will order a licensed appraisal and that report is the legal benchmark. Twellie's role is to make sure your offer math holds up before you write it.
Methodology last updated June 1, 2026. Questions? hello@twellie.com