How we measure accuracy
Accuracy claims in this industry are usually marketing. Ours are a measurement protocol — and the results will be published on this page, regenerated weekly, once the evaluation window is long enough to be honest.
The protocol
- Forward-in-time holdout. The model is trained on sales up to a cutoff date. It is then evaluated on every qualifying sale that happened after that cutoff — lots it never trained on, predicted as they would have been before the sale.
- No look-ahead. The comparable sales the model retrieves for each prediction are restricted to sales that occurred strictly before that lot's own sale date.
- All lots in scope, not a curated subset. Every realized sale above a $100 floor in the window is included. Exclusions, if any, are stated on the page.
- Named baselines, same lots. Every headline metric is shown next to (a) a naive median of the same year/make/model and (b) the auction house's own per-lot estimate, evaluated on identical lots.
- Weak segments shown. Error by price decile, by make/model, by damage type, by title brand, by auction house — including the unflattering cells, each with its sample size.
- Every number carries its n and its evaluation window. No exceptions.
What will be published
| Metric | What it tells you |
|---|---|
| Median absolute % error (MdAPE) | The honest central error for right-skewed prices — less flattering than mean-based metrics, which is the point. |
| Within-10/20/30% hit rates | The share of estimates that landed within X% of the realized price. |
| Interval coverage | Whether the stated 80% price range actually contained the realized price 80% of the time — the promise a probability range makes. |
| Segment drill-downs | Error by price band, vehicle segment, damage, title, and auction house, with sample sizes. |
| Recent held-out sales | A rolling table of recent lots: the model's estimate next to what the lot actually sold for. |
What you will not find here
- Hand-picked success stories without the full-population numbers beside them.
- In-sample or random-split metrics dressed up as accuracy.
- Accuracy claims on thin samples — small cells show their n and nothing else.
- Raw auction listing data. The page reports model outputs, realized prices, and error statistics only.