Positive EV Betting: From Model Estimate to Auditable Decision

Build the decision in seven steps

The sequence combines expected-value arithmetic with chronological evaluation controls:

1. Define the target. Match the model outcome to the operator's market and settlement period.
2. Freeze the forecast. Save probability, model version, inputs and cutoff before checking later information.
3. Capture the price. Record source, selection, decimal price, timestamp, availability and intended stake.
4. Handle margin and costs. Use the offered price for EV; use a declared de-margin method only when comparing market probabilities.
5. Calculate EV. For a binary one-unit back bet, EV = (p * d) - 1.
6. Stress the estimate. Recalculate under plausible lower and upper probabilities.
7. Preserve the decision. Keep bets, passes, missing prices, voids and rejected stakes.

OpenStax's expected-value material supports the weighted-payoff calculation. TimeSeriesSplit documentation provides one way to preserve chronological evaluation when developing the probability model.

Worked decision record

Suppose a model gives 0.51 for over 2.5 goals and a timestamped price is 2.05.

EV = (0.51 * 2.05) - 1 = 0.0455 units

At p = 0.48, EV = (0.48 * 2.05) - 1 = -0.016 units. At p = 0.53, EV = (0.53 * 2.05) - 1 = 0.0865 units. A positive point estimate does not require action when the uncertainty range crosses zero.

Price probability and model probability are different

Raw reciprocal odds show a break-even rate for that quote. A complete market can also be normalized or adjusted with another margin-removal method. Research across 37 competitions found that probability forecasts differed by conversion method and source (Štrumbelj, 2014). Do not relabel a normalized market share as the one true probability.

Evaluate the forecast before the betting rule

First compare model probabilities with outcomes using proper scores and calibration. scikit-learn's calibration guidance explains the reliability check. Then evaluate the complete decision rule using the exact price cutoff, all qualifying rows, settlement, limits and costs. This separation reveals whether a weak return came from probability quality, price capture, execution or variance.

Common audit failures

• Prices are collected after the model cutoff or after favourable movement.
• Losing or unavailable selections disappear from the dataset.
• The threshold is chosen after inspecting the final test period.
• Average advertised odds replace accepted prices.
• Commission is omitted from exchange returns.
• A short favourable sequence is described as proof of positive EV.

Verification checklist

• Recalculate the EV row independently.
• Confirm probabilities and market outcomes are mutually consistent.
• Check the price was available for the recorded stake.
• Compare forecast quality with at least one relevant baseline.
• Report coverage, number of decisions, total stake, uncertainty and drawdown.
• Lock the rule before the next evaluation period.

Next step

Use Expected Value Betting for the next part of this topic.

Continue learning

• Next guide: Value Betting Spreadsheet
• Related guide: Finding Value in Asian Handicap Markets
• Definition: Edge in Betting

Assumptions and limitations

The example is illustrative and deliberately results in a pass. Positive expected value does not guarantee positive realised returns, and a betting rule can stop working after market, data or model changes.