
- Calibrating means comparing the probability you estimated with the real frequency of hits.
- The Brier score sums up your error in a single number, and a table by bands tells you where you go wrong.
- The AI organizes and calculates what you paste in, but you verify the calculations: it can get them wrong.
Save this article: every time you close a block of bets, come back, copy the prompt you need and pick up your measurement where you left off.
Your progress is saved in this browser.
Why calibrate your probabilities with a record?
Winning a bet does not prove your estimate was good.
Losing it does not prove it was bad either.
Only many bets together show whether your numbers look like reality.
Calibrating is exactly that: comparing what you said would happen with what happened.
If you estimated 70% on a hundred occasions, about 70 should have come in.
If 55 came in, your 70% was really a 55%.
This article does not explain the topic from scratch.
If you do not have your record set up yet, start with prompts to audit your betting history.
Here we go one step further: measuring the quality of your probabilities.
Golden rule: give the AI only data that you recorded yourself.
In every prompt we ask it not to invent rows and to write “I don’t know” if information is missing.
Even so, redo each calculation in your spreadsheet.

Before you paste: what data each row of your spreadsheet needs
- The probability you estimated before the match: as a percentage, written down before you knew the result.
- The outcome as 1 or 0: 1 if the bet came in, 0 if it did not. Voided bets stay out.
- The odds you took: the ones you saw at your bookmaker at that moment.
Without the probability written down beforehand, no calibration is possible.
Do not reconstruct it from memory: it comes out biased toward what happened.
If you do not write it down today, start doing it from your next bet.
Prompt 1: organize and clean your record to measure it
This is the starter prompt.
It checks that your table has the right format and tells you what is missing.
It does not calculate anything yet: it only prepares the ground.
I am giving you my bet history in table format: [PASTE YOUR ROWS WITH DATE, MARKET, ODDS, ESTIMATED PROBABILITY AND RESULT]. Check that the probability is between 0% and 100% and that the result is 1 or 0. List the rows with missing, repeated or odd data, giving the row number. Do not correct or invent any data. If something is not understandable, write “I don’t know”.
What it returns
A list of doubtful rows, with their number.
It also warns you if there are probabilities written in different formats, such as 0.6 and 60%.
It is the base for the other four prompts.
How to verify it
Open three of the rows it flagged and check them against your original record.
Also make sure it did not let any obviously wrong row slip by.
You do the final cleanup in the spreadsheet.
Prompt 2: the Brier score of your record
The Brier score sums up in a single number how far your probabilities were from reality.
For each bet you subtract the outcome (1 or 0) from your probability and square it.
Then you average everything.
With this clean table: [PASTE YOUR CHECKED RECORD]. Calculate the Brier score: for each row, (probability − outcome) squared, with the probability in decimal format. Show me the column of squared errors row by row and the final average. Tell me how many rows you used. Compare it with the value you get by always saying 50%, which is 0.25. Do not give an opinion on whether I should bet.
What it returns
A number between 0 and 1, with the detail of each row.
The closer to 0, the better.
If you go above 0.25, your probabilities do worse than always saying 50%.
How to verify it
Recalculate two or three rows with a calculator.
Then add up all the squared errors and divide by the number of rows.
If your average does not match the AI’s, yours is the one that counts.
An example with ten bets.
The figures are made up to show the calculation, they are not real data.
| Bet (example) | Your probability | Outcome | Squared error |
|---|---|---|---|
| 1 | 70% | 1 (came in) | 0.0900 |
| 2 | 60% | 1 (came in) | 0.1600 |
| 3 | 55% | 0 (did not come in) | 0.3025 |
| 4 | 80% | 1 (came in) | 0.0400 |
| 5 | 65% | 0 (did not come in) | 0.4225 |
| 6 | 50% | 1 (came in) | 0.2500 |
| 7 | 75% | 1 (came in) | 0.0625 |
| 8 | 60% | 0 (did not come in) | 0.3600 |
| 9 | 70% | 1 (came in) | 0.0900 |
| 10 | 55% | 0 (did not come in) | 0.3025 |
The sum of the squared errors comes to 2.08.
Divided by ten bets, the Brier score of the example is 0.208.
It is below 0.25, but with ten bets you cannot conclude anything.
In the example, the average probability was 64% and 6 of 10 came in, that is 60%.
It is a small difference, but with so little data it may be pure chance.
In how many bets you need to know whether you are winning we look at how much noise there is in small samples.
Prompt 3: the calibration table by probability bands
The Brier score gives you an overall grade, but it does not tell you where you go wrong.
The table by bands groups your bets according to the probability you estimated.
For example, from 50% to 60%, from 60% to 70% and so on.
With my clean record: [PASTE YOUR CHECKED RECORD]. Build a table by bands of estimated probability: 50-60%, 60-70%, 70-80% and 80-100%. For each band show me: number of bets, average estimated probability, how many came in, real hit percentage and the difference in percentage points. If a band has fewer than [MINIMUM NUMBER OF BETS, FOR EXAMPLE 20] bets, mark it as “small sample”. Do not draw conclusions from the marked bands.
What it returns
A table with one row per band.
The most important column is the difference between what was estimated and what was real.
It also warns you which bands have too little data to say anything.
How to verify it
Count the bets in one band by hand and check that they match the table.
Check that the sum of bets across all the bands equals the total number of rows.
If a row is missing, the AI left it out.
| Band (example) | Bets | Est. prob. | Came in | Real % | Difference |
|---|---|---|---|---|---|
| 50-60% | 40 | 55% | 22 | 55.0% | 0 points |
| 60-70% | 60 | 65% | 36 | 60.0% | 5 points lower |
| 70-80% | 50 | 75% | 32 | 64.0% | 11 points lower |
The numbers in the table are an example to show the calculation.
In the 70-80% band, 32 out of 50 is 64%, against the 75% estimated.
In that example, your most confident bets were the ones that came in the least.

Prompt 4: diagnosing overconfidence or underconfidence
With the table built, it still has to be read.
This prompt helps you see the pattern, without giving betting recommendations.
It also reminds you how much noise there is in each band.
This is my calibration table by bands: [PASTE THE TABLE]. Tell me in which bands my real percentage was lower than the estimated one and in which it was higher. For each band calculate the approximate margin of error with the formula square root of (p × (1 − p) / n), where p is my real percentage and n the number of bets. State whether the difference is larger or smaller than that margin. Write which conclusions are firm and which are only a clue. Do not tell me whether I should bet or how much.
What it returns
A reading band by band.
It indicates whether you were too optimistic, too pessimistic or even.
And it separates what is a pattern from what may be chance.
How to verify it
Redo the margin calculation for one band.
With p equal to 60% and n equal to 30, the margin is around 9 points.
With n equal to 100, it drops to about 5 points, and with n equal to 400, to about 2.5.
| Bets in the band | Approx. margin (p = 60%) | What you can say |
|---|---|---|
| 30 | ± 9 points | Almost nothing firm |
| 100 | ± 5 points | Only large differences |
| 400 | ± 2.5 points | Medium differences already count |
The table is an approximate calculation, useful as an order of magnitude.
It is not a statistical guarantee.
When a difference is small compared with the margin, do not take it as a pattern.
Prompt 5: compare your probability with the one in the odds and plan the adjustment
The odds also hide a probability: 1 divided by the odds.
Odds of 1.80 imply 55.6%; odds of 2.10, 47.6%.
This prompt compares which of the two measures, yours or the odds’, came closer to what happened.
With my clean record: [PASTE YOUR CHECKED RECORD WITH THE ODDS TAKEN]. Calculate the implied probability of each of the odds as 1 divided by the odds, without adjusting for the bookmaker’s margin, and make clear that it is an approximation. Calculate the Brier score using that implied probability and compare it with mine. Tell me in how many bets my probability was closer to the outcome than the implied one. Propose a single test adjustment, such as subtracting [POINTS] points from my probabilities in the high bands, to evaluate with new bets. Do not give me a betting verdict.
What it returns
Two Brier scores to compare, yours and the odds’.
If yours is worse, the odds were a better predictor than you.
Also an adjustment idea to try, without applying it all at once.
How to verify it
Convert three sets of odds to probability with the odds converter.
Check that the AI made clear it did not remove the margin, a topic we explain in implied probability.
Do not apply the suggested adjustment until you have evaluated it with new bets.
Calibrated is not the same as profitable.
You can have good probabilities and still lose if the odds already reflected them.
To go from probability to decision, use the expected value calculator.
How to use the 5 prompts in order: step by step
The limits of calibrating with AI
- The AI does not see your real record: only what you paste in, and it can make mistakes when calculating.
- With few bets, the Brier score and the bands are pure noise.
- If you wrote the probability down after the match, the measurement is useless.
- A good score does not guarantee profits: it also depends on the odds.
- Past results do not guarantee future ones.
Calibrating is a slow habit, not a shortcut.
It is for getting to know yourself as a bettor, not for predicting the next match.
If you want to measure someone else’s prediction service, see how to audit a tipster with AI.
What Gambeta does with its own probabilities, without promises
Gambeta publishes predictions with its own model.
Each one is recorded with its odds before the match.
You can review them in the public record, win or lose.
If you want to see how it works, go to the guide to free AI for betting, with free predictions every day.
A prediction is not a promise.
No bet is safe.
18 and over only.
Bet only money you can afford to lose, set a limit before you start and stop if you stop enjoying it.
More at responsible betting.
What to read next
To turn this analysis into a short routine, every Sunday, see 5 prompts to close out the betting week with AI.
And to understand why a bad run proves nothing, read losing streaks: how many are normal.
If you want a prompt made to measure, build it with the prompt builder.
And for the rest of the markets, go back to the index of betting prompts by market.
Frequently asked questions
What does it mean to calibrate a probability in betting?
It means checking whether your estimates match reality.
If you said 60% on many bets, about 6 in 10 should have come in.
If far fewer came in, you were overconfident.
What is the Brier score?
It is the average squared error between your probability and the outcome.
The outcome is 1 if it happened and 0 if it did not.
The closer to 0 the better; always saying 50% gives 0.25.
How many bets do I need to calibrate myself?
For the overall score a few dozen are enough, but with a lot of noise.
For the table by bands you need quite a few more, because they are spread across groups.
With little data, take any conclusion as a clue, not as a fact.
Can the AI calculate the Brier score?
Yes, if you give it the data and ask for the calculation step by step.
But it can make arithmetic mistakes or invent rows.
Redo the calculation in a spreadsheet before trusting the number.
Does a good Brier score mean I make money?
No.
You are measuring whether your probabilities are good, not whether you found odds with value.
You can be well calibrated and still lose if the odds already reflected those probabilities.
What do I do if I find out I am overconfident?
Try lowering your probabilities in the high bands and see whether the score improves with new bets.
Do not change everything at once.
Write down the adjustment and evaluate again when you have more data.