Property Value Assessment Using Quantile Regression Forests in Real Estate Markets
Annotatsiya
Automated valuation models (AVMs) conventionally report a single point estimate of property value, discarding the fact that valuation uncertainty itself varies enormously across properties: a tract home in a subdivision with dozens of recent comparable sales is a fundamentally easier valuation problem than an architecturally unique custom home with almost no direct comparables, yet a point-estimate model reports both with the same implicit confidence. Quantile Regression Forests (QRF), introduced by Meinshausen (2006), extend Random Forests from estimating the conditional mean to estimating the full conditional distribution, recovering any desired quantile from the weighted empirical distribution of leaf-co-occurring training observations across the forest. This paper applies QRF (500 trees, minimum leaf size 10) to 84,600 residential sale transactions across a regional metro market (Single-Family Detached: 50,760; Townhouse/Condo: 25,380; Multi-Family: 8,460) over a 5-year window, using 18 structural, locational, and market features, with a temporal 80/20 train-test split. The median-quantile point accuracy (MdAPE 7.8%; 75.6% of predictions within 10% of actual sale price) is close to a Gradient Boosting point-estimate benchmark (7.6%) and outperforms a standard Random Forest (8.4%) and OLS hedonic regression (12.4%). The 90% prediction interval has an empirical coverage of 88.4% against a nominal target of 90%. Most importantly, there is an expected, systematic, and monotonic relationship between mean interval width and comparable-sales density, with a 2.85-fold difference between the interval width for high and architecturally unusual properties where there were fewer than 2 close comparables, indicating that QRF's uncertainty estimates reflect actual valuation difficulty, not uniform, uninformative bands. Mortgage-underwriting risk triage achieved an 80.0% market standard of 34.4% of listings sold at or above asking price within 14 days of flagging in cases of full manual appraisal, compared to just 70.0% of the cases flagged by an interval-width-based routing application, while reducing mean underwriting time by 58.3%. The current scope of reliable use of the model was bounded by each of these four reasons for spatial extrapolation failure in newly developed areas, staleness due to the renovation quality, staleness due to the market-regime, and a measurable income-related fairness gap.
Hali tarjima qilinmagan