From Raw Data to a Priced Position
There is no oracle here, and no insider whispering in an earpiece. There is a disciplined seven-stage pipeline that converts publicly available information into calibrated probabilities, compares those probabilities against what the market is charging, and acts only where the gap is wide enough to be worth the risk. Below is every stage of it, documented rather than dramatised.
- 01≈ 340 raw fields per contestRaw material
Ingestion
Every morning the pipeline pulls box scores, snap and minute counts, pace metrics, injury designations, travel and rest schedules, weather at kickoff, and the opening-to-current line history from multiple books. Nothing enters the model until it has passed a staleness check and a cross-source consistency test.
- 02Recency half-life: 21 daysFrom noise to signal
Feature engineering
Raw fields are useless in their native form. They are normalised into z-scores against league and positional baselines, decayed exponentially so last week outweighs last October, and adjusted for opponent strength so a productive night against a porous defence is not mistaken for form.
- 038 weighted feature familiesWhat actually matters
Signal weighting
Each engineered feature carries a fitted weight. The weights are not intuitions — they are the coefficients that minimised out-of-sample log loss across the training window, and they are published in full rather than hidden behind a black box.
- 04Brier score 0.2378Turning score into probability
Calibration
The composite signal passes through a logistic link fitted against closing lines. This is the step most public models skip: a directional lean is not a probability, and betting a lean without knowing its magnitude is how bankrolls die.
- 05Release threshold: edge > 1.2%Edge, not opinion
Market comparison
The calibrated probability is compared against the vig-adjusted implied probability of the posted price. If the model does not beat the number by a meaningful margin, no play is released — regardless of how confident the projection looks in isolation.
- 06Max exposure: 3.0uSurvival first
Stake sizing
Approved selections are sized by quarter-Kelly and hard-capped at three units. Fractional Kelly deliberately sacrifices theoretical growth for drawdown tolerance, because the fastest way to lose a real edge is to be over-levered during an ordinary losing week.
- 07Weekly refit cadenceThe loop closes
Grading and retraining
Every released selection is settled, scored for calibration error, and folded back into the training window. Closing-line value is tracked separately from win rate, because beating the close is the leading indicator that an edge is real rather than lucky.
Published Feature Weights
The coefficients that drive every release. No hidden layer, no undisclosed override.
Where the process is verified
A pipeline is only as credible as its audit. Ours was replayed across 100,000,000 settled selections at 57.35% blended accuracy and +2.9% return on turnover.