Real Estate Valuation Model Performance Evaluation
Summary
This process involves the systematic evaluation and sanity checking of the Sokrates real estate valuation model, specifically focusing on its performance across different property segments in the Icelandic capital area. The analysis tracks model accuracy against official assessments (fasteignamat) and investigates the impact of market cooling and House Price Index (HPI) reinflation logic.
Details
The evaluation focuses on the model’s ability to predict property values relative to the official fasteignamat (official assessment), which is often a lagging indicator. The analysis distinguishes between two primary segments: Apartments (fjölbýli) and Detached houses (einbýli).
Model Performance by Segment
- Apartments (fjölbýli): The model demonstrates healthy performance in this segment. Predictions show a mean error between -5.8% and +3.1% relative to fasteignamat, with a standard deviation of approximately 10-14%. The model tends to predict slightly below fasteignamat on average, which is considered reasonable given that official assessments can overshoot during market shifts.
- Detached Houses (einbýli): The model faces significant challenges in this segment, showing a systematic underprediction (mean error of -14.6% to -15.5%) and high variance (standard deviation of ~18-21%). Specific outliers, such as properties in Glæsibær (-70%) and Garðaflöt (-48%), highlight the model’s struggle with heterogeneous properties. The high variance is attributed to missing features such as lot utilization, renovation status, and neighborhood micro-effects that are not captured by basic structural signals like size, age, and location.
HPI Reinflation and Segment Selection
A critical part of the evaluation involved verifying the House Price Index (HPI) reinflation logic. The pipeline uses sub-indices from HMS (Húsnæðis- og mannvirkjastofnun) to adjust historical prices to current values (reinflation).
The investigation confirmed that the segment selection logic is functioning correctly:
- Capital Apartments: Uses index 111.5 (11.5% appreciation since the Jan 2024 base).
- Capital Houses: Uses index 112.6 (12.6% appreciation).
- Rural Properties: Uses index 116.4 (16.4% appreciation).
While initial sanity checks questioned why all capital area properties received the same HPI regardless of postal code, it was confirmed that the HMS index is designed to be capital-wide rather than granular to the postal code level.
Market Correction and Temporal Trends
The analysis revealed a significant market shift between 2024 and 2026. In 2024 and 2025, the median sale price for detached houses was approximately 13% above fasteignamat. By 2026, this premium crashed to +4.6%, with the bottom quartile (p25) of houses selling at -5.1% below assessment.
The model’s perceived “bias” of -15% is partially explained by this market cooling; the model, trained on historical data where houses sold at a higher premium, is predicting values based on historical patterns that the 2026 market no longer supports. This suggests that the model’s performance is reasonable when accounting for the rapid downward market correction, though it still requires better feature engineering to handle property heterogeneity.