How we measureand what we evaluate against.

Target definition

The engine predicts the market price of an equivalent vehicle at selling level: the last observed price of a listing before it disappears from the market, after its full sequence of price drops. It is the best observable proxy for the closing price at market scale.

In-house sold tracking

We detect sales with our own tracking: every listing is monitored daily and, when it disappears in a sustained way, it is recorded as a sale event together with its complete price-drop curve and time on market. This signal is calibrated per market to each country’s typical stock rotation.

Honesty about the asking-to-close gap

A listing’s initial price is not the price the deal closes at; the gap varies by country, segment, and vehicle age. We do not ignore it: the training pipeline explicitly models the price trajectory down to the listing’s disappearance, and evaluation is done against that last price — not the first one.

Blind valuation

In the backtest, the model values each vehicle without seeing the listing price: it receives only brand, model, trim, year, mileage, fuel, and power — exactly what it would receive from a client. The prediction is then compared with the last price observed before the sale.

Stratified sampling

The evaluation sample is stratified over recorded sales to reflect the real composition of the market — by brand, segment, and fuel type — instead of concentrating on the easy vehicles.

Published metrics

We report MAPE with a confidence interval, bias (mean signed error), the share of valuations within ±10%, ±15%, and ±20% of the real price, and R². No single metric tells the whole story; we publish the set.

Continuous validation with a regression guard

After every retraining, each country’s accuracy report is regenerated and an automatic guard compares it with the previous cycle: if a country degrades beyond the threshold, an alert fires and the deployment is reviewed before the new model is served.

Official sources and sealed catalogs

The new-price catalog is built per market from official sources — public registries and bodies such as the BOE in Spain, OFV in Norway, or Skatteverket in Sweden — plus sealed historical catalogs by trim. We do not depend on a single origin per country.

Canonical normalization

Brand, model, and trim are normalized to a canonical catalog shared across the 13 countries, with prices in local currency and EUR. Every trim is anchored to its new price, making depreciation curves comparable across markets.

Why it matters

Anchoring gives the model a stable reference when market data is scarce — rare trims, small markets — and prevents a local bias in active stock from dragging the valuation.

No personal data

We do not collect, store, or sell personal data of buyers or sellers. The object of analysis is the vehicle and its price — not people.

No literal listing content

Our data products deliver structured, aggregated market information — prices, specifications, movements — never the literal content of source listings (texts, photographs).

Coverage varies by country

Data depth is not uniform: large markets carry more signal than small ones, and per-country accuracy varies accordingly. Per-country model cards exist precisely so that difference is visible before you sign.

More uncertainty on older vehicles

Above ~8 years of age, real market dispersion grows — condition, history, and maintenance outweigh the spec sheet — and model error grows with it. We reflect that in the intervals rather than hiding it in the average.