What was observed
Photo-score distribution and anchors
Finish: 1 very basic/poor; 2 dated budget; 3 ordinary functional; 4 modern good quality; 5 exceptional premium materials and fittings. Upkeep: 1 serious visible disrepair; 2 worn/damaged; 3 serviceable with wear; 4 well maintained; 5 freshly finished/no visible wear. Half-points are allowed. These are anchored categories, not percentiles; scores are concentrated in a few categories.
Model and validation
Log monthly rent = intercept + regularised property features + a bilinearly interpolated location grid. The spatial penalty is the squared second differences in x and y plus twice the squared mixed differences: a discrete thin-plate penalty. Curvature penalties are normalised for grid spacing; the 125 m display cells do not provide extra fitted resolution. There is no explicit subway-distance feature. Huber weighting reduces extreme-advert influence.
Grid spacing, ridge strength and spatial smoothness are selected within grouped three-fold validation using a joint parameter search. The headline accuracy uses separate fitting, calibration and test groups randomly distributed across the study area; nearby training comparables are allowed. Validation uses nested five-fold grouped cross-validation. The displayed listing estimates and calculator share the model refitted on all groups except the reserved calibration groups. Adverts sharing an address, near-identical coordinates or photo asset are grouped.
Unknown is a learned missing-data category, not a neutral physical state. Supplying floor or area can therefore move a prediction either way. Training-only validation chooses between a shared area slope and bedroom-specific reference sizes with slopes shrunk toward a shared slope. The area-unreported adjustment is shown separately from physical floor area.
The 80% calculator interval uses a finite-rank-corrected calibration quantile of absolute log residuals. Market drift and correlated listings weaken exchangeability; it is empirical guidance, not a guaranteed coverage statement. Displayed listings use the same interval scale as the calculator. For listings used in fitting, this is a model reference range rather than held-out coverage evidence.
Interpretation limits
These are asking prices, including adverts marked let agreed—not completed tenancy prices. The location field also absorbs unmeasured property composition. Photos are selected by agents and cannot establish noise, damp behind furniture, neighbours, management quality or scams.
Floor area and numeric EPC are extracted from advert text and linked EPC/floorplan documents where legible. Linked certificates can be old; no current-registry validation is implied. Council tax, location and dwelling size are correlated; individual effects can be unstable. Advert update date is a proxy for price date, not tenancy date. No causal interpretation is intended.
The snapshot comes from Rightmove only. Room-only, student-hall, short-term and ambiguous-price adverts are filtered; conservative deduplication can merge distinct units with identical advert metadata. Mid-market rent is tagged from explicit advert wording and included as a model indicator; only a few examples are available. Untagged does not establish unrestricted market rent. Areas above 50 + 40 × bedrooms m² are treated as missing, while reported values are preserved. Known document/flat mismatches also exclude document-derived numeric EPC. Read the source advert before relying on a flagged listing.