The three parts of value
- Probability: your pre-event estimate that the defined outcome occurs.
- Price: the gross decimal return available for one unit staked.
- Expected value: the probability-weighted average payoff implied by those two inputs.
OpenStax defines probability as a numerical description of uncertainty and expected value as a probability-weighted average. In betting, both ideas must refer to the exact same event and settlement rule.
Worked example
Suppose a model estimates 47% for a regulation-time home win and a bookmaker displays decimal odds 2.25.
- Raw break-even probability = 1 / 2.25 = 0.4444, or 44.44%.
- Estimated probability edge = 47% - 44.44% = 2.56 percentage points.
- One-unit win profit = 2.25 - 1 = 1.25 units.
- EV = (0.47 * 1.25) - (0.53 * 1) = 0.0575 units.
The arithmetic says the bet is positive EV if 47% is a reliable estimate. It does not validate the estimate. At 44%, EV = (0.44 * 1.25) - (0.56 * 1) = -0.01 units.
What value is not
| Observation | Why it is insufficient |
|---|---|
| The team is favourite | Favourite status says nothing about whether the price is generous |
| The odds are high | A large payoff may still be too small for the chance of winning |
| The bet won | One outcome cannot validate the pre-event probability |
| The price shortened later | Movement is a later observation with its own benchmark limits |
| Another bookmaker is shorter | Sources can differ in margin, timing, limits and forecast quality |
Research comparing methods for deriving forecasts from odds found that normalization, Shin probabilities and source choice can produce different estimates (Štrumbelj, 2014). Therefore, a de-margined market probability is a named estimate, not an unknowable objective truth.
How to make the probability defensible
Freeze the model before the match, retain every forecast, evaluate later fixtures in chronological order, and compare with a relevant baseline. Check calibration by probability band: forecasts near 40%, 50% and 60% should be assessed across many comparable cases. scikit-learn's calibration documentation explains this reliability relationship.
Also record an uncertainty interval or sensitivity range. If reasonable alternative assumptions move the estimate from 43% to 48%, a price that breaks even at 44.44% is not a robust opportunity. A practical rule is to pass when the decision changes under modest, plausible input variation.
The pre-event checklist
Use this checklist as a local release rule after checking probability reliability through calibration analysis:
- Does the model target match the visible market and settlement period?
- Were all inputs available at the recorded cutoff?
- Is the quoted price executable for the intended stake?
- Has margin or commission been handled consistently?
- Does the EV remain positive under a declared probability sensitivity test?
- Will every qualifying decision be recorded, including passes and rejected bets?
Continue learning
- Next guide: What Is Vig? Bookmaker Margin in Plain Language
- Related guide: Expected Value in Betting
- Definition: Value Bet
Assumptions and limitations
The examples are illustrative and ignore operator-specific rules, limits, voids and taxes. A positive model expectation is not a guarantee of realised profit. For the full calculation, continue to expected value in betting; for a reproducible record, use the value model spreadsheet specification.
GambleAware
GamCare
Gambling Therapy
GAMSTOP