How Prediqta works
Prediqta is a gradient-boosting AI valuation engine trained on 200,000+ real Hungarian residential transactions. Every valuation ships with a confidence band and explainable factors.
Model architecture
The core engine is a gradient-boosted decision tree ensemble (XGBoost-class), combined with a geographic embedding layer that encodes location value independently of unit attributes. A meta-model calibrates the ensemble output per segment (new-build vs. resale, urban vs. rural).
Training data
- Recorded sales since 2010, nationwide
- Daily-refreshed listing-portal signal (ingatlan.com and peers)
- KSH (Hungarian Central Statistical Office): district price index, demographics, income
- HÉSZ zoning, transit accessibility, POI density
Training and validation cadence
The main model retrains monthly on fresh transactions; listing signal updates weekly as a delta. Hold-out validation on every train run with an 80/10/10 train/val/test split. Model versions are pinnable for the duration of an audit cycle.
Accuracy by segment
Accuracy varies by transaction volume in the segment. The platform-wide average is 7.5%.
- New-build Budapest apartments: 3–4%
- Resale Budapest apartments: 6–7%
- Regional cities (Debrecen, Szeged, Győr, Pécs): 8–9%
- Rural property: 10–12% (wider confidence band)
- Platform-wide average: 7.5%
Known limitations
- Atypical layouts get a wider confidence band
- Sparse-transaction locations (rural) trail recent market shifts
- Standalone garden valuations are less accurate (fewer comparables)