Why probability without a price is not a decision
A forecast answers how often an outcome should occur. A price determines whether that forecast has any decision value. Treating them as the same thing is the fastest route to a misleading betting product.
Start with a probability, not a pick
If a model assigns a 60% probability to an outcome, its fair decimal price is 1 / 0.60 = 1.67. That is not an instruction. It is only the price at which the model’s expected return is zero before uncertainty, limits, execution and model error.
A binary “Over” label discards most of the useful information. FORM/PRICE therefore stores the full probability, its model artifact, the exact data cutoff and the price snapshot used for comparison.
Remove the bookmaker margin
Raw implied probabilities from both sides of a market normally sum to more than 100%. Comparing a model directly with that number overstates the disagreement. For a two-way market we proportionally remove the margin, then compare against the median of identical Over/Under lines from multiple bookmakers.
Expected value depends on the attainable price
At model probability 60%, decimal odds of 1.80 imply an expected value of 0.60 × 1.80 − 1 = +8%. At odds of 1.55, the same forecast implies −7%. The probability did not change; the decision did.
This is why every research row carries a bookmaker, price, vendor update time and capture time. A price observed after the decision—or one already stale when captured—is forbidden evidence.
Closing-line value before short-run ROI
Small betting samples can show impressive ROI by chance. A stronger diagnostic asks whether the recorded entry price consistently beats the last temporally valid price from the same bookmaker before kickoff. FORM/PRICE defines decimal CLV as entry odds / closing odds − 1.
CLV is not proof of profit, but it is harder to fake with a few fortunate results. Our evidence gate requires a positive lower 95% bootstrap bound, not merely a positive average.
What must happen before anything is public
- 500 settled target-horizon probability observations;
- 150 selections passing the frozen candidate gate;
- at least 80% same-bookmaker closing-price coverage;
- Brier skill above 2% and expected calibration error at or below 3%;
- a positive lower 95% confidence bound for mean CLV;
- independent methodology and jurisdiction-specific legal review.