Aarhus RE Scanner — Automated Property Valuation Model
A solo-built automated valuation model (AVM) for Aarhus residential real estate, benchmarked directly against the accuracy figures published by the Danish tax authority's own model-development process — the one at the center of Denmark's 2013 ejendomsvurderingsskandale. Five reverse-engineered and public data sources feed a 5-model stacking ensemble, validated on a strict temporal holdout so the reported accuracy reflects genuine forward-looking generalization, not a favorable random split.
Base Models
5
Hit Rate (±20%)
82.2%
Data Sources
7
Holdout Test Rows
6,458
The Model
A hedonic log-price model: 5 base learners (Ridge, ElasticNet, LightGBM, XGBoost, CatBoost) combined via a Ridge meta-learner trained on out-of-fold predictions. Data is joined from Datafordeler BBR (structural attributes), OSM Overpass (spatial amenities), Statistics Denmark (macro/fiscal context), and two undocumented sources reverse-engineered from scratch — a historical sale-price API and an energy-certificate lookup, neither of which has an official public endpoint.
Validation is a single temporal holdout: every model is trained only on transactions before a fixed cutoff date and tested only on transactions after it, never seen during training or tuning — the same discipline used across every project on this site. A/B feature and architecture experiments are accepted only when they improve the net holdout RMSE summed across all three property segments (condo, house, apartment), no per-segment cherry-picking.
Benchmarked Against the Government's Own Model Development
Denmark suspended its public property valuations in 2013 after a damning Rigsrevisionen audit and spent the following years building a statistical replacement — documented in a 2014 expert-committee report (Engberg-udvalget) and a 2016 refinement (Skatteministeriet / ICE). Both publish the accuracy metric Danish valuation law actually targets: the percentage of assessments landing within ±20% of the real sale price. Pulled directly from both primary-source PDFs and compared against this model's own genuine held-out test predictions:
| Property type | SKAT, old system (2014) | Engberg prototype (2014) | ICE refined (2016) | This model (2026) |
|---|---|---|---|---|
| Houses | 62.9% | 67.5% | ~71% | 81.3% |
| Condos | — | 83.4% | ~85% | 83.7% |
Both government figures are pre-launch, development-stage numbers (2014/2016), not the system's real-world post-2019 production accuracy, which isn't publicly available — the honest framing is "beats every documented iteration of the government's own model development," not "beats the current system."
Try the Search
Loading snapshot…