Validation Report
We tested the adult-weight estimator on dogs that were excluded from curve creation. This page reports the results—including where the model is weakest.
Model version 3.0 · Analysis generated October 7, 2026
Across 335,780 held-out age observations, the central estimate's median absolute error was 9.5%. The displayed p10–p90 variation band contained 80% of observed young-adult outcomes.
Study split
Eligible dogs were divided deterministically by anonymized PET_ID. Dogs whose ID modulo 5 equaled zero formed the 20% holdout group; none of their puppy measurements contributed to the curve percentiles. The other 80% formed the training group.
Performance by puppy age
The model becomes more informative as the puppy gets older. This is why an early estimate should never be presented with the same certainty as an estimate near one year.
| Age at measurement | Held-out observations | Median absolute error | 80th-percentile error | Variation-band coverage |
|---|---|---|---|---|
| 8–19 weeks | 98,537 | 16% | 31.6% | 79.8% |
| 20–31 weeks | 78,658 | 11.5% | 22.4% | 80.2% |
| 32–51 weeks | 59,591 | 8.2% | 16.3% | 80.2% |
| 52–76 weeks | 98,994 | 5.8% | 12.1% | 79.9% |
| All evaluated ages | 335,780 | 9.5% | 21.1% | 80% |
Performance by adult-size group
| Adult-size group | Held-out observations | Median absolute error | 80th-percentile error | Variation-band coverage |
|---|---|---|---|---|
| Toy | 152,934 | 10.4% | 22.3% | 80% |
| Small | 32,416 | 9.1% | 20% | 79.6% |
| Medium | 36,602 | 9.6% | 21.1% | 80.2% |
| Large | 49,753 | 8.2% | 19.1% | 80.1% |
| Very large | 51,926 | 8.8% | 19.8% | 79.8% |
| Giant (exploratory) | 12,149 | 8.4% | 19.7% | 80.5% |
What each metric means
- Median absolute percentage error
- Half of evaluated estimates were closer to the dog's observed young-adult outcome than this percentage, and half were farther away.
- 80th-percentile absolute error
- Eighty percent of evaluated central estimates had an absolute percentage error at or below this value.
- Variation-band coverage
- The percentage of observed outcomes falling inside the range produced from the training distribution's 10th and 90th percentile growth fractions. This is not a clinical confidence interval.
What this evaluation does not prove
- It is a retrospective holdout drawn from the same source network and era as the training data, not an external prospective validation.
- The source population consisted of owner-reported purebred dogs; mixed-breed performance was not tested.
- Young-adult outcome was defined from eligible measurements at 1.75–2.25 years, not lifetime mature weight.
- Dogs with sufficient adult follow-up may differ from dogs without adult records.
- The over-40 kg group is exploratory because the paper found giant-breed trajectories too heterogeneous for one formal clinical category.
- Error for an individual dog may be much larger than the medians shown here.
Audit the numbers
Download the curve table as CSV, the full results as JSON, and the data README. The complete selection rules and formulas are on the methodology page.
The source data is available under CC BY 4.0 from the University of Liverpool Data Catalogue. The associated peer-reviewed paper is Salt et al. (2017), PLOS ONE.