Published accuracy experiments

Results are internal, reproducible measurements. Whether any outside researcher has reproduced them is shown below. Failed results remain public. Each population and time window has its own criteria; results do not transfer between them.

Independent replication

No outside researcher has reproduced these results or assessed the methodology yet. Until one does, treat every figure on this page as reproducible internal validation, not independent confirmation.

Ascent publishes a replication kit for outside quantitative researchers: one command that re-checks every report, an independent Python re-derivation of every published figure, a method specification for reimplementing the valuations, and a signed record for their conclusions. Unfavorable conclusions are published like favorable ones.

How to replicate these studies · Signed records are checked by npm run validate:valuation -- --check.

Maryland · revised model

9.5%median error on 2,812 unseen Montgomery & Prince George’s, MD (revised) sales95% CI 9.0%–10.0%

52.0%within ±10% of the recorded price79.3% within ±20%

80.2%of test sales inside the calibrated 80% intervalthe raw comparable envelope held 70.7%

0.8 ptslower median error than median $/sqft of the same sales7.8 pts better than a trailing ZIP median

Pre-registered criteria · Montgomery & Prince George’s, MD (revised)

8 of 9 met

Met: Holdout sales valued94.13% target ≥ 90%
Met: Median absolute percentage error9.52% target ≤ 10%
Met: Share within ±10% of sale price51.98% target ≥ 50%
Met: Coefficient of dispersion (COD)14.8 target ≤ 15
Not met: Price-related differential (PRD)1.042 target 0.98–1.03
Met: Absolute median signed error1.30% target ≤ 3%
Met: Beats median $/sqft of the same comparables-0.31 pts target 95% upper bound below 0
Met: Beats trailing ZIP median price-6.97 pts target 95% upper bound below 0
Met: Calibrated 80% interval coverage80.17% target 75%–85%

These figures use the Montgomery & Prince George’s, MD (revised)-calibrated settings, which could value 94.1% of the holdout sales. With Ascent's general default settings, the median error on the same sales was 10.2%.

Revised experiment: consideration at least $50,000, no bedroom or bathroom inputs, construction quality adjustment capped at $75,000. Three previously unscored months, July–September 2026, with recording lag. Price-related bias still failed. This is a different cohort from the original 13.85% result; it is not a paired improvement estimate. The original failure remains published. Read both experiments. No calibrated range is released.

Maryland SDAT/MDP recorded sales of at least $50,000 (Public Domain public records). Rates were fitted on Jan 2022–Dec 2024, search filters chosen on Jan 2025–Sep 2025, the interval calibrated on Oct 2025–Jun 2026, then Jul 2026–Sep 2026 was scored once. Thresholds were fixed before scoring and have not moved. Report sha256-1c420bbcfa82… can be re-run with npm run validate:valuation -- --verify.

Also published: King County, WA, Jan 2015–May 2015, 10.6% median error on 6,975 sales, 5 of 9 criteria met. Results do not transfer between markets.

Maryland · original failed experiment, unchanged

13.8%median error on 12,831 unseen Montgomery & Prince George’s, MD sales95% CI 13.5%–14.2%

39.2%within ±10% of the recorded price63.5% within ±20%

79.4%of test sales inside the calibrated 80% intervalthe raw comparable envelope held 67.0%

3.1 ptshigher median error than median $/sqft of the same sales3.7 pts better than a trailing ZIP median

Pre-registered criteria · Montgomery & Prince George’s, MD

4 of 9 met

Met: Holdout sales valued93.63% target ≥ 90%
Not met: Median absolute percentage error13.85% target ≤ 10%
Not met: Share within ±10% of sale price39.20% target ≥ 50%
Not met: Coefficient of dispersion (COD)9913.8 target ≤ 15
Not met: Price-related differential (PRD)98.794 target 0.98–1.03
Met: Absolute median signed error1.58% target ≤ 3%
Not met: Beats median $/sqft of the same comparables3.49 pts target 95% upper bound below 0
Met: Beats trailing ZIP median price-3.20 pts target 95% upper bound below 0
Met: Calibrated 80% interval coverage79.39% target 75%–85%

These figures use the Montgomery & Prince George’s, MD-calibrated settings, which could value 93.6% of the holdout sales. With Ascent's general default settings, the median error on the same sales was 10.1%.

Montgomery and Prince George’s only. Bedrooms and bathrooms are unavailable; this model uses size, assessor construction quality grade, location and sale timing. Quality grade is not interior condition. Current parcel records retain only the latest three transfers, so older history is incomplete. The $/sqft baseline uses the same raw eligible sale pool, including sales the engine may reject after adjustment. These are reproducible internal results, not independent certification. The fitted model was worse than untuned settings on this holdout. Five retained transfers have consideration at or below $1,000, contributing extreme ratios to the dispersion and price-bias failures. No records were removed after scoring. The model did not meet the acceptance criteria; every miss is shown above.

Montgomery: 7,465 held-out sales; 94.6% valued, 14.1% median error, 38.6% within ±10%. Descriptive slice of the same pooled model; no separate county acceptance claim.

Prince George’s: 5,366 held-out sales; 92.3% valued, 13.5% median error, 40.1% within ±10%. Descriptive slice of the same pooled model; no separate county acceptance claim.

Maryland SDAT/MDP recorded sales in Montgomery and Prince George’s (Public Domain public records). Rates were fitted on Jan 2021–Dec 2023, search filters chosen on Jan 2024–Jun 2024, the interval calibrated on Jul 2024–Jun 2025, then Jul 2025–Jun 2026 was scored once. Thresholds were fixed before scoring and have not moved. Report sha256-483c1495df5e… can be re-run with npm run validate:valuation -- --verify.

Also published: King County, WA, Jan 2015–May 2015, 10.6% median error on 6,975 sales, 5 of 9 criteria met. Results do not transfer between markets.

Maryland method · four other counties, registered before download

12.3%median error on 19,039 unseen Anne Arundel, Baltimore County, Frederick & Howard, MD (external) sales95% CI 12.0%–12.5%

42.5%within ±10% of the recorded price70.1% within ±20%

77.3%of test sales inside the calibrated 80% intervalthe raw comparable envelope held 82.5%

0.3 ptslower median error than median $/sqft of the same sales8.9 pts better than a trailing ZIP median

Pre-registered criteria · Anne Arundel, Baltimore County, Frederick & Howard, MD (external)

5 of 9 met

Met: Holdout sales valued96.31% target ≥ 90%
Not met: Median absolute percentage error12.26% target ≤ 10%
Not met: Share within ±10% of sale price42.49% target ≥ 50%
Not met: Coefficient of dispersion (COD)18.6 target ≤ 15
Not met: Price-related differential (PRD)1.065 target 0.98–1.03
Met: Absolute median signed error3.00% target ≤ 3%
Met: Beats median $/sqft of the same comparables-0.08 pts target 95% upper bound below 0
Met: Beats trailing ZIP median price-8.54 pts target 95% upper bound below 0
Met: Calibrated 80% interval coverage77.25% target 75%–85%

These figures use the Anne Arundel, Baltimore County, Frederick & Howard, MD (external)-calibrated settings, which could value 96.3% of the holdout sales. With Ascent's general default settings, the median error on the same sales was 13.1%.

External test of the revised Maryland method in four counties Ascent does not serve. The protocol, bars and analysis were committed and pushed before these counties’ records were downloaded; the data was locked before any valuation, and the holdout was scored once. The method was refitted only on these counties’ earlier sales. It did not meet acceptance, so it does not support the method beyond Montgomery and Prince George’s, and no calibrated range is released anywhere in Maryland. Registration, results and analysis.

Met exactly at the bar: absolute median signed error (3.00%). Any worse would have failed.

Maryland SDAT/MDP recorded sales in Anne Arundel, Baltimore County, Frederick and Howard (Public Domain public records). Rates were fitted on Jan 2021–Dec 2023, search filters chosen on Jan 2024–Jun 2024, the interval calibrated on Jul 2024–Jun 2025, then Jul 2025–Jun 2026 was scored once. Thresholds were fixed before scoring and have not moved. Report sha256-60f04f4d65eb… can be re-run with npm run validate:valuation -- --verify.

Also published: King County, WA, Jan 2015–May 2015, 10.6% median error on 6,975 sales, 5 of 9 criteria met. Results do not transfer between markets.

DC condominiums · separate population test

13.0%median error on 1,434 unseen Washington, DC condominiums sales95% CI 12.2%–14.0%

41.2%within ±10% of the recorded price68.1% within ±20%

77.0%of test sales inside the calibrated 80% intervalthe raw comparable envelope held 78.9%

0.2 ptslower median error than median $/sqft of the same sales18.4 pts better than a trailing ZIP median

Pre-registered criteria · Washington, DC condominiums

4 of 9 met

Met: Holdout sales valued96.23% target ≥ 90%
Not met: Median absolute percentage error12.95% target ≤ 10%
Not met: Share within ±10% of sale price41.16% target ≥ 50%
Not met: Coefficient of dispersion (COD)17.8 target ≤ 15
Not met: Price-related differential (PRD)1.055 target 0.98–1.03
Met: Absolute median signed error2.98% target ≤ 3%
Not met: Beats median $/sqft of the same comparables0.61 pts target 95% upper bound below 0
Met: Beats trailing ZIP median price-16.35 pts target 95% upper bound below 0
Met: Calibrated 80% interval coverage76.96% target 75%–85%

These figures use the Washington, DC condominiums-calibrated settings, which could value 96.2% of the holdout sales. With Ascent's general default settings, the median error on the same sales was 17.6%.

Condo-specific study; studios included, condition adjustment disabled. Missing floor, view, parking, HOA finances and assessments. Outcomes in the same building can be correlated; sale-level confidence intervals may be too narrow. Internally reproduced, awaiting independent professional review. No calibrated range is released.

District of Columbia qualified condominium sales (CC-BY-4.0 public records). Rates were fitted on Jul 2022–Jun 2024, search filters chosen on Jul 2024–Dec 2024, the interval calibrated on Jan 2025–Sep 2025, then Oct 2025–Sep 2026 was scored once. Thresholds were fixed before scoring and have not moved. Report sha256-06e88b2bfbcf… can be re-run with npm run validate:valuation -- --verify.

Also published: King County, WA, Jan 2015–May 2015, 10.6% median error on 6,975 sales, 5 of 9 criteria met. Results do not transfer between markets.

△ ascent

Opening Ascent…

Return to Ascent

Published accuracy experiments

Independent replication

How to replicate these studies · Signed records are checked by npm run validate:valuation -- --check.

Maryland · revised model

9.5%median error on 2,812 unseen Montgomery & Prince George’s, MD (revised) sales95% CI 9.0%–10.0%

52.0%within ±10% of the recorded price79.3% within ±20%

80.2%of test sales inside the calibrated 80% intervalthe raw comparable envelope held 70.7%

0.8 ptslower median error than median $/sqft of the same sales7.8 pts better than a trailing ZIP median

Pre-registered criteria · Montgomery & Prince George’s, MD (revised)

8 of 9 met

Met: Holdout sales valued94.13% target ≥ 90%
Met: Median absolute percentage error9.52% target ≤ 10%
Met: Share within ±10% of sale price51.98% target ≥ 50%
Met: Coefficient of dispersion (COD)14.8 target ≤ 15
Not met: Price-related differential (PRD)1.042 target 0.98–1.03
Met: Absolute median signed error1.30% target ≤ 3%
Met: Beats median $/sqft of the same comparables-0.31 pts target 95% upper bound below 0
Met: Beats trailing ZIP median price-6.97 pts target 95% upper bound below 0
Met: Calibrated 80% interval coverage80.17% target 75%–85%

Also published: King County, WA, Jan 2015–May 2015, 10.6% median error on 6,975 sales, 5 of 9 criteria met. Results do not transfer between markets.

Maryland · original failed experiment, unchanged

13.8%median error on 12,831 unseen Montgomery & Prince George’s, MD sales95% CI 13.5%–14.2%

39.2%within ±10% of the recorded price63.5% within ±20%

79.4%of test sales inside the calibrated 80% intervalthe raw comparable envelope held 67.0%

3.1 ptshigher median error than median $/sqft of the same sales3.7 pts better than a trailing ZIP median

Pre-registered criteria · Montgomery & Prince George’s, MD

4 of 9 met

Met: Holdout sales valued93.63% target ≥ 90%
Not met: Median absolute percentage error13.85% target ≤ 10%
Not met: Share within ±10% of sale price39.20% target ≥ 50%
Not met: Coefficient of dispersion (COD)9913.8 target ≤ 15
Not met: Price-related differential (PRD)98.794 target 0.98–1.03
Met: Absolute median signed error1.58% target ≤ 3%
Not met: Beats median $/sqft of the same comparables3.49 pts target 95% upper bound below 0
Met: Beats trailing ZIP median price-3.20 pts target 95% upper bound below 0
Met: Calibrated 80% interval coverage79.39% target 75%–85%

Montgomery: 7,465 held-out sales; 94.6% valued, 14.1% median error, 38.6% within ±10%. Descriptive slice of the same pooled model; no separate county acceptance claim.

Prince George’s: 5,366 held-out sales; 92.3% valued, 13.5% median error, 40.1% within ±10%. Descriptive slice of the same pooled model; no separate county acceptance claim.

Also published: King County, WA, Jan 2015–May 2015, 10.6% median error on 6,975 sales, 5 of 9 criteria met. Results do not transfer between markets.

Maryland method · four other counties, registered before download

12.3%median error on 19,039 unseen Anne Arundel, Baltimore County, Frederick & Howard, MD (external) sales95% CI 12.0%–12.5%

42.5%within ±10% of the recorded price70.1% within ±20%

77.3%of test sales inside the calibrated 80% intervalthe raw comparable envelope held 82.5%

0.3 ptslower median error than median $/sqft of the same sales8.9 pts better than a trailing ZIP median

Pre-registered criteria · Anne Arundel, Baltimore County, Frederick & Howard, MD (external)

5 of 9 met

Met: Holdout sales valued96.31% target ≥ 90%
Not met: Median absolute percentage error12.26% target ≤ 10%
Not met: Share within ±10% of sale price42.49% target ≥ 50%
Not met: Coefficient of dispersion (COD)18.6 target ≤ 15
Not met: Price-related differential (PRD)1.065 target 0.98–1.03
Met: Absolute median signed error3.00% target ≤ 3%
Met: Beats median $/sqft of the same comparables-0.08 pts target 95% upper bound below 0
Met: Beats trailing ZIP median price-8.54 pts target 95% upper bound below 0
Met: Calibrated 80% interval coverage77.25% target 75%–85%

Met exactly at the bar: absolute median signed error (3.00%). Any worse would have failed.

Also published: King County, WA, Jan 2015–May 2015, 10.6% median error on 6,975 sales, 5 of 9 criteria met. Results do not transfer between markets.

DC condominiums · separate population test

13.0%median error on 1,434 unseen Washington, DC condominiums sales95% CI 12.2%–14.0%

41.2%within ±10% of the recorded price68.1% within ±20%

77.0%of test sales inside the calibrated 80% intervalthe raw comparable envelope held 78.9%

0.2 ptslower median error than median $/sqft of the same sales18.4 pts better than a trailing ZIP median

Pre-registered criteria · Washington, DC condominiums

4 of 9 met

Met: Holdout sales valued96.23% target ≥ 90%
Not met: Median absolute percentage error12.95% target ≤ 10%
Not met: Share within ±10% of sale price41.16% target ≥ 50%
Not met: Coefficient of dispersion (COD)17.8 target ≤ 15
Not met: Price-related differential (PRD)1.055 target 0.98–1.03
Met: Absolute median signed error2.98% target ≤ 3%
Not met: Beats median $/sqft of the same comparables0.61 pts target 95% upper bound below 0
Met: Beats trailing ZIP median price-16.35 pts target 95% upper bound below 0
Met: Calibrated 80% interval coverage76.96% target 75%–85%

Also published: King County, WA, Jan 2015–May 2015, 10.6% median error on 6,975 sales, 5 of 9 criteria met. Results do not transfer between markets.