📦 TRAINING SET
forward-collected from the live board · AU/NZ onlyEvery tracked race is snapshotted at fixed marks before the jump (T‑120 … T‑2 min) — price, tote share, Betfair weight-of-money, form, tips — and labelled after the race with what actually happened: did the price firm (shorten ≥8% to the jump), and did the runner win. This grows daily; the model needs weeks of it.
Coverage per capture mark: how many runner-snapshots exist at each minutes-before-jump bucket, and how many are already labelled with a result. The far-out buckets (T‑120/T‑90) fill slowest — a race is only tracked once it enters the 2-hour window.
📡 LIVE REFLECTION
recent races — predicted firmer vs who actually closedEach row is a settled race: the runner the firm-score rated highest before the race (at the earliest capture mark), the price then vs where it closed at the jump, and whether it actually firmed / won. Green move = shortened as predicted.
🔮 HEURISTIC TRACK RECORD
the rule-based firm-score, graded on real outcomesEvery heuristic prediction is logged at the time it's made and scored once the race resolves. Lift = tier firm-rate ÷ overall firm-rate — above 1× means the score finds firmers better than chance. This is the baseline the ML model has to beat.
How the score is built (v1.1) — a weighted blend, per runner, field-relative, all knowable before the jump:
| FACTOR | WEIGHT | SIGNAL |
|---|---|---|
| Market respect | 0.35 | implied probability from the best price, favourite-normalised |
| Recent form | 0.25 | TAB form rating, min-max within the field |
| Money flow | 0.25 | Betfair weight-of-money (backers stacking the queue) blended with tote pool-share change since open |
| Consensus | 0.15 | tipped runner / expert best-bet |
| × Momentum | 0.75–1.10 | graded by the size of the pool-share move — a whisper nudges, a surge saturates |
🧠 MODEL
LightGBM · walk-forward validated · offlineThe model trains offline on the collected snapshots and is judged against backing the market's opening price out-of-sample (time-ordered split, isotonic-calibrated probabilities, half-Kelly ROI + closing-line-value backtest). It ships to the board only when it beats that baseline — until then the heuristic stands in.