Standard reports and similarity weighting

Standard reports use RentJudge’s proprietary ML similarity model to weight each comparable rental’s contribution to your rent estimate. Closer property matches receive more influence, rather than every comparable counting equally.

How it works

The model compares location, living area, bedrooms, bathrooms, building year, property type and how recently the listing was last seen. Learned coefficients combine these differences into a penalty, which becomes a similarity score. The scores are normalized into contributions that sum to 100% across the included comps.

Standard reports take a weighted average of the original listed rents. Pro uses the same similarity weights with rents additionally adjusted for property characteristics, amenities and available photo analysis. Weighting changes a comp’s contribution; it does not change the comp’s listed dollar amount. See Pro reports and adjustments.

Search radius, listing age and other comparable criteria determine which rentals are retrieved. A report searches for up to 25 qualifying comps; Standard and Pro display and calculate from every returned comp. Sparse markets or narrower criteria can return fewer. Similarity weighting is a separate step on the final comparable set.

Reading estimate weights

Estimate weight shows how much influence each comparable has in the estimate. The list, map details, shared report and PDF use the same percentages. For example, a comp with a 50% weight has half the influence in the weighted average. Small weights are normal when many comps are included.

Included comps’ weights total 100% before rounding. Excluding a comp sets its weight to zero and redistributes the weight across the remaining comps. These percentages are not accuracy, confidence or match percentages. Badges show only the percentage, with color indicating relative contribution: green is above an equal share, amber is half to one equal share, red is below half, and gray is excluded. With 10 included comps, an equal share is 10%, so green is above 10%, amber is 5–10%, and red is below 5%. Color does not measure match quality or accuracy.

Use the ⓘ beside a percentage for its explanation and available comparison facts. In Pro reports, weights apply to each comp’s adjusted rent. The report and PDF include a written explanation and the color legend.

Unknown size and building year receive learned missing-data penalties. Invalid scoring inputs or a scoring-service failure prevent new report completion. Earlier snapshots without learned weights retain their equal-average results and show equal weights with an explanation.

Offline benchmark

Experiment date: September 11, 2026. We sampled 100,000 distinct properties from rental listing data. The final held-out test included 12,023 properties with at least five eligible comparables.

The decrease was 37.4%, calculated as (533.16 − 333.80) / 533.16. This measures lower average absolute dollar error, not a 37.4-percentage-point increase in accuracy.

Coefficients were trained on older observations using Huber loss on log estimated rent versus log observed rent, with nonnegative coefficients. Hyperparameters were selected on a separate validation period. The figures above are from the untouched later test period, not training loss. Linear and quadratic variants were compared; the simpler linear model was chosen for this version.

All methods used the same nearest geographic candidates within three miles and 180 days, capped at twenty and requiring at least five. This candidate set did not apply today’s production bedroom, size or property-type filters. Subjects and their property identities were excluded from the reference pool. The current 25-comp report limit was not separately evaluated in this experiment.

Limits: this retrospective snapshot used asking rents, not verified achieved rents, and does not reconstruct every historical listing attribute. The result has not been verified on RentJudge’s current production-filtered reports, unseen future data or Pro-adjusted rents. It does not establish a gain over competitor estimates, guarantee an achievable rent or predict the error for an individual property.

See Credits and report changes for pricing and included changes.