Plain-English account of where the numbers in an Eight Furlong single-horse report come from, what they can and can't tell you, and where the model is limited. Linked from the footer of every report so the analytical prose can stay focused on the horse.
Every percentile, every "above-cohort-median" comparison, every empirical anchor in a report is computed against the same 14,654 yearlings across 38 major Australian sales 2019–2026. By sale family:
| Sale family | Coverage | Cohort lots |
|---|---|---|
| MM Gold Coast Flagship | 2019–2026 (most complete) | 5,909 |
| Inglis Premier (Melbourne) | 2022–2025 | 2,736 |
| Inglis Australian Easter | 2020–2025 | 2,327 |
| MM Perth Premier | 2020–2026 | 1,328 |
| MM Adelaide Premier | 2020–2025 | 1,129 |
| MM Gold Coast Yearling (Book 2/3) | 2019–2023 | 992 |
| Inglis Classic | 2021, 2022 only | 233 |
Sales not yet in the cohort (single-horse reports for these lots score against the same cohort as everything else — the underlying horse comparison is unchanged):
Coverage audit: single-horse reports cover approximately 79% of catalogued lots within the sales we crawl. The 21% of catalogued lots not in the cohort tend to be higher-priced (mean $222k vs cohort median $120k) and systematically returned less per dollar paid — so any model claim about market inefficiency is conservative relative to the true market.
Each report's central scorecard scores twelve biomechanical features, six static body ratios plus six dynamic motion features:
| Feature | Source | Favourable when |
|---|---|---|
| Chest depth / body length | Static | lower (less stocky) |
| Shoulder angle | Static | lower (more sloped) |
| Hip angle | Static | lower (more sloped) |
| Tail angle (resting) | Static | lower |
| Body length (motion-frame) | Dynamic | lower (relative to height) |
| Neck joint angle (motion) | Dynamic | higher |
| Hip joint left / right (motion) | Dynamic | lower |
| Shoulder joint left (motion) | Dynamic | lower |
| Fore-hind phase left / right (motion) | Dynamic | higher (better coupling) |
| Tail range of motion | Dynamic | lower |
Each feature is independently scored against the cohort. The net score is the count of favourable leans minus the count of unfavourable leans. A horse with a flat scorecard (most features in the cohort's 25–75 percentile middle band) gets a net of 0 — this is a clean read of "physically average", not a failure to detect.
When a sire has at least 5 mature cohort progeny (≥5 career starts + stud-filtered), the sire-context section above the summary populates with three blocks:
For sparse-cohort sires (fewer than 5 progeny past the maturity gate — Zacinto and other first-crop / cooling sires fall in this bucket), the sire-context section explains the gap rather than fabricating blocks. The 12-feature scorecard is unaffected — it scores against the full 14,654-row cohort.
We do not currently fit per-sire predictive models. That would require approximately 400+ mature cohort progeny per sire — only a handful of Australian sires have reached that volume in our cohort.
Each report's "Maximum defensible bid" section quotes career-prizemoney percentile ceilings from a validated subset of 1,872 horses in the middle (net −2 to +2) biomech tier:
These ceilings cover racing return only — stud value, breeding income, and residual sale value are excluded by design. Add them to the ceiling on your own terms.
Out-of-sample backtest (train pre-2021, test 2021+): binomial test on the 75th-percentile ceiling, p < 10⁻¹⁸⁰.
Source videos vary by sale house and year — typically 25 fps, occasionally 50 fps (most newer Inglis sales). The pipeline reads frame rate from the source file and computes velocities accordingly. Per-feature n columns in the cohort data are raw frame counts, so a 50-fps source naturally shows double the frame count for the same wall-clock walking section. The metric values themselves (median angle, median ratio, etc.) are frame-rate-independent and remain directly comparable across the cohort.
Limitations we're aware of — disclosed so you can apply judgement on top of the model:
Every cohort price and buyer attribution in our analysis is verified against Magic Millions' live catalogue at audit time. Every sire / dam name and DOB joins live to racing-db at report-build time, so career records and prize money update automatically as horses race.
Cohort source-of-truth: gs://mm-parade-videos/barn_processed/parade_path_c/parade_path_c_joined_v2.json. Open-source biomech extraction code is in the eightfurlong-parade GitHub repo (private; available on request).