We index public listings, normalise them against closed transactions, and train gradient-boosted models on year, model, kilometres, condition, location, and seasonality. Here's exactly what goes in and what comes out.
§ 01
Input signals
Year, model, variant, kilometres, declared condition (5 classes), district, ownership history, paperwork status, and listing recency. Image-derived condition scoring is rolling out in Q2 2026.
§ 02
Training data
Roughly 50,000 closed transactions sourced from dealer partnerships and verified private-sale reports, plus 380,000 active listings. We retrain monthly to capture seasonality and supply shocks.
§ 03
Known limits
We can't see mechanical condition. We can't price modifications. We struggle with imports under 100 transaction samples. We flag predictions with low confidence rather than guess.