Valuation accuracy

Accuracy

Tests on recent dealer sales only, each valued with its sale kept out of the evidence.

4,000 dealer cars sold 9 to 16 October 2026

Test on recent dealer sales · 9 October 2026 to 16 October 2026 · 4,000 cars · Up to 1,000 a night, in a fixed pseudo-random order, of the dealer sales recorded on 14 to 17 October 2026: each was valued at 23:50 Perth that night by the live engine, before the 00:30 refresh took it into the engine's evidence. Same rules as our regular 500-car test set (trim and series known, km and price stated, model year 2003 or later, price $5,000 or more, one sale per car, sales later found false left out). Month to date: the full October rows replace it on 1 November.

Correction, 13 October 2026. We now work out every figure the same way. For the 43 cars we valued under $15,000 the worst-10% figure moves from 19.9% to 29.1%, because a single car sits on that line; over $50,000 it moves from 11.3% to 12.4%; the overall median reads 4.22% and the worst 10% 14.28%.

500 dealer cars sold 24 to 26 September 2026

Test on recent dealer sales · 24 September 2026 to 26 September 2026 · 500 cars · Drawn from the dealer sales recorded by the 28 Sep load, after that night's engine refresh, so none of the 500 sales was in the engine's evidence; same rules as our regular 500-car test set (trim and series known, km and price stated, model year 2003 or later, price $5,000 or more, one sale per car), taken in a fixed pseudo-random order; valued 28 Sep 2026 05:01 Perth. All figures are computed from the saved per-car results with the same method as the standing-500 runs (median; 90th percentile at the nearest car); first published as 4.23 / 14.27 with a percentile one car lower. Confidence grades were not recorded for this run, so there is no confidence split.

Correction, 13 October 2026. We now work out every figure the same way. For the 43 cars we valued under $15,000 the worst-10% figure moves from 19.9% to 29.1%, because a single car sits on that line; over $50,000 it moves from 11.3% to 12.4%; the overall median reads 4.22% and the worst 10% 14.28%.

Every run on the standing 500

Not an accuracy claim: the engine had already seen these sales. Every engine change is scored on the same 500 sold cars before it goes live, and the tester also runs each morning. These are all 43 runs since the set was fixed on 28 September, newest first. They read about 0.9 of a point better than a held-out test (3.36% against 4.23% median error, both measured on 28 September), so they are kept out of the headline figures.

How to read this

1. Tests on recent dealer sales are the accuracy. Such a test values each car with its sale kept out of the engine’s evidence, which is how you use it: you value a car before it sells. The standing 500 is the fixed set every engine change is scored on; the engine has already seen those sales, so it reads better. We list every run, kept apart, and quote only the held-out figure.
2. Against the last advertised price. We see advertised prices, not what was paid, so error is measured against the last price each car was advertised at before it sold. What was negotiated after that is not visible to us. Where a dealer's ad states no price basis, we count the price as advertised, excluding government charges. That is about 6% of priced sales.
3. A median is not a promise about one car. That is why we also show the worst 10% and the share within 10%, and why every value carries a range and a confidence grade.
4. How the engine works and what we do not claim are on Data and method.