homeval-comps-adjustments-and-avms

Home Valuation and Market Research: Comps, Adjustments, Regression and AVMs

Part 3 of 7 of the Home Valuation and Market Research reference (plugin home-valuation), covering §10–§15. Sibling skills: homeval-concepts-methods-and-market-structure (§0–§4), homeval-market-analysis-and-timing (§5–§9), homeval-buyer-playbook (§16–§20), homeval-seller-playbook (§21–§25), homeval-diligence-risk-and-investment (§26–§30), homeval-reference (§31–§35). Code: scripts/homeval.py, scripts/test_homeval.py.

Currency: The method is stable (it is the core of the Appraisal Institute / USPAP approach). The AVM error figures are Zillow's published numbers as of Aug 8, 2026 and will refresh.

⚠️ Scope. This is how to do a careful informal valuation. It is not a USPAP-compliant appraisal and is not a substitute for one where a lender, court or tax authority requires one. The worked example in §15 uses a synthetic market with a known true price function, so the tools' accuracy can be verified; real markets are noisier and have unobserved condition and quality.

The three ideas:

1. ⚠️ The best comp is the one that needed the least adjustment, not the one that sold highest or lowest. Every adjustment is an opinion that injects error; minimise the number and size of them (§10–§11).
2. ⚠️ Adjustments come from the market, not from the contractor. A $40,000 kitchen does not add $40,000. Derive adjustments from paired sales or regression (§11, §13).
3. ⚠️ Report a range, not a number. If your comps-based value, regression and best AVMs agree within ~3–4%, say so with moderate confidence. If they spread 10%+, the honest answer is "it depends on condition/terms; here is how to resolve it" (§15.4).

---

§10. Selecting Comparable Sales

10.1 The priority-ordered rules

Rank candidates; do not just filter. Hard guardrails (relax only with a written reason) then ranked similarity.

Criterion	Standard starting guardrail	Widen when…	⚠️ Pitfall
Status	Closed arm's-length sales	Support only with active/pending (§10.3)	List prices are not comps. They are asks
Recency	≤ 90 days ideal; ≤ 6 months acceptable	Thin market: ≤ 12 mo with time adjustment	Stale comps in a moving market are the most common error (§12)
Distance	≤ 1 mile (urban/suburban)	Rural: 3–10+ miles	Don't cross school-zone, flood-zone, busy-road, HOA or municipal-tax lines without adjusting
Size (GLA)	± 10–20%	Large/unique homes	GLA measured differently (use ANSI Z765 logic); below-grade ≠ GLA
Style / design	Same (ranch with ranch)	Few comps	Cross-style comps need larger adjustments
Age / quality	± 10–15 years and same quality tier	Thin market	A renovated 1970s house is closer to newer than to its unrenovated twin
Condition	Same band (5-point scale you apply consistently)	Always adjust if different	Flips as comps for unrenovated homes = overvaluation
Beds/baths/garage/lot	± 1 bed / ± 1 bath / same garage count	–	Count above-grade only
Property rights / type	Same (fee simple SFR vs SFR)	Never mix	Condo vs SFR, leasehold vs fee, new vs resale: different markets
Transaction type	Exclude REO/short sales/relatives/estate quick-sales unless the market is distressed	Distress-heavy local market	Non-market sales understate value

10.2 Ranking score

homeval.score_comps() computes a transparent similarity score: normalised gaps in distance, recency, GLA, age, beds, baths, condition, lot and style, with heavier weight on GLA, condition and location. Keep the score components visible so you can explain why a comp ranked where it did; an unexplainable rank is a black box you will not defend in negotiation.

10.3 How many, and the supporting cast

• Minimum: 3 closed comps; good practice: 5–6, plus supporting evidence: (a) active listings (the competition and the ceiling), (b) pending listings (current demand, price per pending date), (c) expired/withdrawn listings in the last 6 months (prices the market rejected). A price above the cheapest comparable active listing that is better is a price that will not sell.
• Bracket the subject: ideally comps both above and below on key features (size, condition, price). A set all smaller than the subject forces extrapolation.
• Use the same tie-breaker an appraiser does: which would a buyer of this house pick instead? That is the substitution principle.

10.4 Data hygiene

Verify each comp with at least two sources (MLS or portal + county recorder). Check: sale date vs recording date, concessions (sellers paid X), financing (seller financing, assumption, buydown), GLA vs permitted (unpermitted additions), lot (usable vs gross), condition from photos (kitchen/baths/flooring/roof), days on market, price history (cuts, relists).

---

§11. The Adjustment Grid

11.1 Direction (get this right or every number is backwards)

Adjust the comp toward the subject.

• Comp is superior to the subject on a feature → subtract (negative adjustment).
• Comp is inferior → add (positive adjustment).
• Example: comp has a third garage bay the subject lacks → subtract the value of a bay. Comp is 150 sf smaller → add 150 × marginal $/sf.

11.2 Order (USPAP/URAR sequence)

1. Transactional adjustments first (applied in order, each on the cumulative price): property rights, financing terms, conditions of sale, expenditures immediately after purchase (deferred maintenance).
2. Market conditions (time), applied to the adjusted price (§12).
3. Then property characteristics: location, site, view, design/quality, age/condition, GLA, rooms, garage, extras.

11.3 Where the numbers come from (in descending reliability)

Method	How	When to use
Paired sales	Two sales identical except one feature; the price gap ≈ the feature's value	Rich markets; best evidence
Regression (hedonic)	Coefficients on log price (§13)	≥ ~50 local sales; controls many features at once
Matched-pair inference from listings	Compare listings pre/post renovation, or same model homes with/without a feature	Cross-check
Cost-to-cure	Cost of fixing a defect (not of building the feature new)	Deferred maintenance, defects
Depreciated cost	Replacement cost new − depreciation	Only when market evidence is absent; it overstates
Survey/rule-of-thumb (Cost vs Value etc.)	Industry averages	Last resort; ⚠️ these are survey-based opinions (§23)

11.4 Typical magnitudes (starting hypotheses ONLY; replace with local evidence)

Feature	Typical treatment	⚠️ Caveat
GLA	~30–50% of the comp's average $/sf for each marginal sf (marginal sf < average sf)	Using the average $/sf over-adjusts; smaller differences → smaller error
Bathroom	A bath of a type (full vs half) is worth ~low-to-mid five figures in typical markets; use local pairs	Count above-grade full/half consistently
Bedroom	Often near zero after adjusting GLA; bed count matters at thresholds (3 vs 4)	Adjusting both GLA and beds double-counts
Garage bay	Low-to-mid four to low five figures	Attached vs detached, door height, workshop
Condition	~2–5% of price per condition step on a consistent scale	Largest and most subjective adjustment; document with photos
Lot	$/sf for excess usable land, not gross	Slope, flood, easements, shape
Pool	Often near zero or negative in many climates	Region-dependent; verify locally
Age/updating	Small per-year effect, large step effects from systems (roof, HVAC)	Overlaps with condition; avoid double counting
Location	Paired sales across the boundary (school, road, flood)	Often the largest single factor

homeval.adjust_comps() defaults are placeholders (e.g., $7,500/bath, $8,000/garage bay, condition at 3% per step, GLA at 40% of $/sf). Override each with local evidence.

11.5 Size limits (quality control)

Traditional underwriting guidance flags net adjustments > ~15% and gross adjustments > ~25% of a comp's price (⚠️ guidelines, not law). adjust_comps() marks OVER-ADJUSTED above these; drop such comps unless nothing else exists, and say so.

11.6 Cash-equivalent adjustments (financing and concessions)

If a seller paid concessions that were reflected in the price, the cash-equivalent price is lower. Simple approach: subtract seller-paid concessions (and the value of any below-market seller financing/buydown) from the sale price before other adjustments. Document it; many "paid-over-list" sales evaporate after this step.

---

§12. Market-Conditions (Time) Adjustment

Problem: a comp that sold 5 months ago reflects the market 5 months ago. In a market moving +/−0.5% per month that is ±2.5%; during a rate shock (§7) it can be larger.

Methods (use at least two):

1. Paired resales: the same house selling twice (repeat sales) is the cleanest signal.
2. Regression on sale date (log price per sf on months, time_adjustment_pct_per_month()): noisy with fewer than ~30 sales.
3. Index: metro Case-Shiller / FHFA HPI / Zillow ZHVI / Cotality HPI monthly changes.
4. Live evidence: pending list-to-contract prices and active listing price cuts show where the market is now; if pending sales are below recent closings, drift is negative even if the index is lagging.
5. Sanity cap: ±0.5%/month is a plausible range outside crises; extrapolating 2–3%/month from a seasonal bump is a classic error.

⚠️ Do not apply an annual appreciation rate mechanically in a market with a rate shock. Re-estimate after any move of ~0.5+ points in rates.

---

§13. Hedonic (Regression) Cross-Check

Model: ln(price) = β0 + β1·ln(sqft) + β2·beds + β3·baths + β4·ln(lot) + β5·age + β6·condition + β7·garage + β8·months_since_sale + ε.

Why log price: effects become percentages; heteroscedasticity falls; coefficient on ln(sqft) is an elasticity (in the example: 0.84 vs a true 0.85: a 1% bigger house costs ~0.84% more, not 1%).

When to run it: ≥ ~50–100 sales in your market area in the last 6–12 months, same property type. Add neighborhood/school-zone dummies if you can.

Checks (do all):

• Hold-out test: fit on 80%, predict 20%; report median absolute % error (MdAPE).
• Residual plot: look for clusters (a street or condition class the model misses), heavy tails, trend with price.
• Robust standard errors (the toolkit uses HC0); sign and size sanity on every coefficient (age negative? condition positive?).
• Outliers: examine, don't auto-drop (they may be flips or distress).
• What it cannot see: condition beyond your scale, finishes, view, layout, deferred maintenance, lot quality. Condition is the dominant omitted variable in real data, so treat a regression as a cross-check on comps, never the sole answer.

Interpretation: price_band and hedonic_predict() return p10–p90 using the model's residual spread. approx_mdape_pct = 0.6745 × residual sd × 100 ≈ the typical absolute % error.

---

§14. AVMs: Zestimate, Redfin Estimate and Friends

14.1 Published accuracy (the honest numbers)

• Zillow (own published, accuracy data refreshed Aug 8, 2026): median error 1.78% for on-market homes and 7.20% for off-market homes. (Other Zillow-cited snapshots in 2026 ranged 1.9–2.4% on-market and 6.9–7.5% off-market. ⚠️ Trade press repeats different vintages; use the live Zillow page.)
• Redfin Estimate: reported at ~1.98% on-market and ~7.66% off-market (older published figure).
• By geography: off-market median error varies widely by state (e.g., ~5.3% in Colorado vs ~12.7% in Vermont in Zillow's Apr 2026 table; ⚠️ third-party summary); rural/thin markets: 8–12%+ is reported.
• Why on-market is 4× better: the estimate incorporates the list price, description and photos, i.e., it partly copies the human CMA. An off-market Zestimate is the real model.

14.2 What "median error" means

Half of estimates are worse than the median; the tail is much worse. On a $430,000 home, 7.2% = ±$31,000 for the median case. Zillow publishes the share within 5%, 10% and 20% of the sale price: check the within-10%/within-20% shares for your metro to see tail risk. Do not interpret "±median error" as a confidence interval for your house; it's roughly a 50% interval (§15.4).

14.3 How to use AVMs well

1. Collect 3–5 (Zillow, Redfin, Realtor.com, Cotality/CoreLogic, a county value, HouseCanary or bank tool). Their spread is an uncertainty signal: wide spread → model doesn't know your house.
2. Correct the inputs: beds, baths, GLA, lot, upgrades, year, condition. Claim your home on Zillow and edit facts; Redfin lets you update details.
3. AVMs lag turning points and cannot see: kitchen quality, layout flaws, view, road noise, deferred maintenance, additions. Treat AVMs as the first screen, not the answer.
4. Never use an AVM alone to price, offer or refinance. Use it to decide whether to invest a weekend in a real analysis.
5. LLMs are not valuation tools. A language model with no comps data will fabricate a plausible number. Use an LLM for structure (checklists, drafting, code) but feed the numbers from real sales data. Same standard for any "AI CMA" product: ask what comps and what error rate.

---

§15. Worked Example (sandbox-verified) and the Recipe for Any House

15.1 Setup (synthetic market with known truth)

220 synthetic closed sales (price function with known elasticities and a +0.15%/month market drift), noise 4.5% log-sd. Subject: 2,000 sf, 4 bd, 2.5 ba, 9,500 sf lot, built 2004, condition 4/5, 2-car garage, two-story. Ground-truth value today: $261,623. (The tools do not see this.)

15.2 Steps and results (run python3 scripts/test_homeval.py)

1. Market stats (last 90 days): n = 49, median price $220,500, median $/sf $126, median DOM 49, 24.5% over list, 26.5% with price cut.
2. Time drift (regression on date): +0.227%/month (truth +0.15%; noisy by construction, so capped to ±0.5%).
3. Rank comps → 71 within guardrails (≤1 mile, ≤6 months, GLA within 20%); use top 6.
4. Adjustment grid (comp → subject):

Comp	Sale	Time	GLA	Baths	Garage	Cond.	Net adj.	Adjusted	Net %	Gross %
116	269,000	+1,022	−268	0	0	−8,101	−7,347	261,653	−3	3
62	250,400	+391	+1,371	+7,500	0	0	+9,262	259,662	+4	4
186	256,700	+1,493	+933	0	−8,000	−7,746	−13,320	243,380	−5	7
8	227,000	+2,392	+5,183	0	+8,000	0	+15,575	242,575	+7	7
110	234,800	+3,004	−4,259	+7,500	+16,000	0	+22,245	257,045	+9	13
28	263,100	+1,412	−4,339	0	0	+7,935	+5,009	268,109	+2	5

5. Reconcile (weight by 1/(1+gross%/10) and similarity; use the 3 lowest-gross comps: 116, 62, 28): $262,615 → +0.38% vs truth. Range of best: $259,662–$268,109 (spread ~3%). Unweighted median of all six: $258,354.
6. Hedonic regression (n = 220, R² = 0.948, log-residual sd 0.045): sqft elasticity 0.838 (truth 0.85); point $257,605, p10–p90 $243,293–$272,759, implied MdAPE ~3.0%; −1.54% vs truth. (Drift coefficient 0.06%/mo vs truth 0.15% shows why §12 says use multiple methods.)
7. Triangulate (60% comps / 40% regression): $260,611 (−0.39% vs truth).
8. Band for an off-market AVM at 7.2% median error: $241,847–$279,375, wider than the 3% spread of the best comps, the point of doing real comps.

15.3 What the example teaches

• Comps with gross adjustments ≤5% (116, 62, 28) are the evidence; comp 110 (13% gross, three adjustments) is supporting, not core.
• Condition and garage drive the largest errors; document them with photos.
• The regression was less accurate than good comps (−1.5% vs +0.4%) but stable, and it flagged the right range.
• The AVM-style band was ~4× wider than the comp spread: a quantification of why careful work beats a single automated number when the subject is comparable and the data is good.

15.4 Recipe: valuing any real house (do these in order)

INPUTS : address, GLA, beds/baths, lot, year, condition notes, upgrades+dates, permits
1  Pull closed sales: same property type, last 6 mo (12 if thin), 1 mile (widen stepwise), GLA ±20%.
2  Verify each comp with two sources; log concessions/financing; photograph-based condition score (1–5, consistent).
3  Rank with score_comps(); keep guardrail-passing, top 5–6 plus 2–3 active/pending as supporting.
4  Estimate time drift 2 ways (repeat sales/regression/index) and from live pending/active evidence; cap at ±0.5%/mo.
5  Build the grid with adjust_comps(); replace placeholder amounts with LOCAL paired-sale/regression numbers.
6  Drop OVER-ADJUSTED comps; reconcile() with top 3 lowest-gross; record spread.
7  If ≥50 local sales: hedonic_fit()/hedonic_predict(); hold-out MdAPE; compare with step 6.
8  Collect 3–5 AVMs; use their spread as an uncertainty indicator, not as the answer.
9  Check live competition: best 3 active homes that beat the subject on price/feature; adjust expectation downward if they undercut.
10 State value as a RANGE + confidence + assumptions + what would change it.

Confidence statement template: "Opinion of value (informal, not an appraisal): $X–$Y, most likely about $Z, effective [date]. Based on N closed comps (list), adjusted to cash-equivalent, time-adjusted at [rate] and cross-checked with a regression (MdAPE [x]%) and AVMs ($a–$b). Largest uncertainty: [condition/layout/time]. A change of ±[k]% in [assumption] moves value by ±$m."

15.5 Tool calls (copy-paste)

import pandas as pd, homeval as hv
TODAY = pd.Timestamp("YYYY-MM-DD")
subj = hv.Subject(sqft=..., beds=..., baths=..., lot_sqft=..., year_built=..., condition=..., garage=..., style="...", lat=..., lon=...)
ranked = hv.score_comps(subj, comps_df, TODAY)               # columns in docstring
good = ranked[ranked.within_guardrails].head(6)
drift = hv.time_adjustment_pct_per_month(comps_df)            # cap to ±0.5
adj = hv.adjust_comps(subj, good, TODAY, monthly_drift_pct=max(min(drift,0.5),-0.5),
                      bath_adj=LOCAL, garage_bay_adj=LOCAL, lot_per_sqft=LOCAL)
rec = hv.reconcile(adj, scores=good.score)
m = hv.hedonic_fit(comps_df, TODAY); pred = hv.hedonic_predict(m, subj)
tri = 0.6*rec["estimate"] + 0.4*pred["point"]