The Brier score is one way to evaluate consistently defined probability forecasts preserved before an event. A journal also needs separate decision, receipt, settlement, and correction fields so later knowledge does not overwrite the original record.
Keep five records separate
| Layer | Minimum fields | Freeze point |
|---|---|---|
| Forecast | Event definition, probability, evidence, method version, data cutoff | Before viewing the outcome |
| Decision | Action or no action, reason, price threshold, maximum loss, stop conditions | Before placement |
| Contract | Accepted event, market, selection, period, price, stake, receipt ID, rules | At acceptance |
| Settlement | Status, gross return, fees, net result, settled time, source | When reconciled |
| Process and wellbeing | Rule adherence, pressure or impairment flag, effect on time, money, sleep, work, or relationships | At decision and review |
Do not overwrite a pre-event value with a later interpretation. Add a correction row containing the old value, new value, reason, evidence, and timestamp.
Set up the journal
- Give every decision a stable ID, including no-bet decisions.
- Define the event and forecast target precisely.
- Save probability, evidence, and data cutoff before the event.
- Record the maximum possible loss across related open positions.
- Attach the accepted receipt and governing rule.
- Reconcile settlement from the account record.
- Review rule adherence before discussing the result.
- Re-queue any decision made after a stop condition for a wellbeing review, not performance optimisation.
Worked probability entry
Suppose four illustrative binary forecasts are 0.70, 0.60, 0.40, and 0.30. Outcomes are 1, 0, 1, and 0. The Brier components are:
- (0.70 - 1)^2 = 0.09;
- (0.60 - 0)^2 = 0.36;
- (0.40 - 1)^2 = 0.36;
- (0.30 - 0)^2 = 0.09.
Mean Brier score = (0.09 + 0.36 + 0.36 + 0.09) / 4 = 0.225. Lower is better for the same target and scoring convention. The original probability-forecast verification paper defines the score. Four observations are far too few for a stable conclusion; keep calibration, discrimination, baseline comparison, and uncertainty separate.
Scikit-learn's calibration documentation explains that a calibration curve groups predictions into bins and compares mean predicted probability with the observed positive fraction. Report the binning rule and case coverage; do not create a universal minimum sample rule or use one score to claim expertise.
Review without outcome bias
The original outcome-bias research shows why knowing an outcome can alter evaluation of the decision. Hide the result while rating whether the evidence cutoff, forecast, contract, exposure, and stop conditions were followed. Then reveal and reconcile the outcome as a separate step.
A journal is not a safety device by itself
A complete spreadsheet can coexist with gambling harm. If the journal shows borrowing, hidden activity, repeated limit breaches, chasing, or effects on essentials, sleep, work, study, health, or relationships, stop the performance review. The NHS gambling-harm page provides practical controls and treatment and support routes.
Next step
Use How To Track Your Bets for the next part of this topic.
Continue learning
- Next guide: Betting Routine and Discipline
- Related guide: Chasing Losses
Assumptions and limitations
The worked forecasts are hypothetical. The page does not recommend gambling, stakes, or a performance threshold. Scores are comparable only for consistently defined targets, horizons, and samples. Records can be incomplete or selectively maintained, and past performance does not establish future results.
GambleAware
GamCare
Gambling Therapy
GAMSTOP