Learn
The Six Point Flip: a betting model calibration sheet for baseball and every other sport
Betting model calibration is the check that the model's percentages come true, and the Six Point Flip shows why it matters: on a -125 baseball favorite, a model that is six points too confident turns a winning bet into a losing one. This sheet walks through the flip and the calibration check, with a table you can fill in.
A baseball moneyline is a bet on a probability. If your model's probability is off by a few points, the bet can go from worth taking to worth avoiding without the pick changing at all. The Six Point Flip makes that concrete, and the calibration sheet below is how you find out whether your model has the problem.
The two moves from the video, in order:
- See the flip. On one bet, compare what the model's probability says it's worth with what a slightly lower true probability says.
- Check the percentages, not the picks. Group the model's past calls by the probability it gave and count how often each group actually won.
Who this is for
This sheet is for anyone who bets from a model's numbers, whether the model is a spreadsheet you built, a paid prediction service or one of ours. You need the model's past probabilities for a few hundred games, the results of those games, and a spreadsheet. No coding is required. If you only ever see a model's picks and never its probabilities, ask for the probabilities first, because there's nothing to calibrate without them.
Step 1. See the flip
Take a hypothetical moneyline favorite at -125. Using the moneyline formulas, that price risks 125 to win 100, so it needs to win 55.6% of the time to break even.
| Your model says (hypothetical) | Fair odds for that chance | Expected value per $100 at -125 |
|---|---|---|
| 50% | +100 | -$10.00 |
| 52% | about -108 | -$6.40 |
| 54% | about -117 | -$2.80 |
| 55.6% | -125 | about $0 |
| 56% | about -127 | +$0.80 |
| 58% | about -138 | +$4.40 |
| 60% | -150 | +$8.00 |
| 62% | about -163 | +$11.60 |
Read the 58% and 52% rows together. If the model says 58%, the bet averages about $4.40 per $100. If the team really wins 52% of the time, the same bet loses about $6.40 per $100. Six points of probability is the whole difference between a good bet and a bad one, and the pick ("this team wins") is the same in both rows. You can run any line in our expected value calculator.
Step 2. Check the percentages
Calibration is the test that the model's percentages come true. Walsh and Joshi, writing in Machine Learning with Applications in 2024, found that selecting NBA betting models by calibration rather than accuracy led to greater returns. Baseball gives you plenty of games to run the same check, since each team plays 162 in a regular season.
How to run it:
- Collect the model's past calls. For each game, the probability it gave the side you track, and whether that side won.
- Group the calls into bands, for example 50 to 54%, 55 to 59%, 60 to 64% and 65% or more.
- Count wins in each band. Divide by the number of games to get the actual win rate.
- Compare each band's average prediction with its actual win rate. Calibrated means they're close. A model that says 58% on games that win 52% is the Six Point Flip waiting to happen.
- Plot it if you like. Predicted on the horizontal axis, actual on the vertical, and a diagonal line for perfect calibration, which is the property the Walsh and Joshi paper found worth selecting models for.
The calibration sheet
| Band | Games | Average model probability | Wins | Actual win rate | Gap (actual minus model) |
|---|---|---|---|---|---|
| 50 to 54% | |||||
| 55 to 59% | |||||
| 60 to 64% | |||||
| 65% or more |
A worked band, hypothetical: 120 games in the 55 to 59% band, average model probability 57.5%, 63 wins, actual win rate 52.5%, gap -5.0 points. That model is about five points too confident in that band, so every bet it likes near -125 is closer to the 52% row than the 58% row. Rerun those bets at 52.5% in the expected value calculator and most of them turn negative.
How to fix an over-confident model
- Shrink toward the market. Blend the model's probability with the market's fair chance. Betting odds are hard to beat; Wunderlich and Memmert note they frequently outperform mathematical models, so leaning on them is a reasonable default.
- Recalibrate by band. If the 55 to 59% band wins 52.5%, map those calls down before you compare them with prices.
- Re-test on new games. A fix that only works on the games you built it on is not a fix.
What to watch out for
- Too few games per band. A band with twenty games can be off by ten points from luck alone.
- Mixing prices. Calibrate the model's probability, not the bets you happened to take.
- Only checking winners. A model can pick plenty of winners and still be wrong on the price.
- Limits. If the model does find real edges, books may limit you; Kaunitz, Zhong and Kreiner reported account limits a few months into real-money betting.
How to tell it worked
After your fix, the bands' actual win rates should sit within a couple of points of their average predictions on games the model hasn't seen. Then the expected-value table above means something for your bets. The Quant+ model scorecards show calibration for each model, and you can compare any model's probability with the market's no-vig price in our no-vig calculator.
21+. For informational purposes only. Nothing is guaranteed. If gambling stops being fun, call 1-800-GAMBLER.
Questions people ask
What does calibration mean for a betting model?
It means that when the model says 60%, those outcomes happen about six times in ten. A calibrated model's percentages can be used as prices; a poorly calibrated one's can't, even if it picks plenty of winners.
Why does calibration matter more than accuracy?
Because a bet's value depends on the probability, not on the pick. Walsh and Joshi (2024) found that choosing models by calibration rather than accuracy led to greater returns in NBA betting experiments.
How many games do I need for a calibration check?
Enough for each probability band to hold a meaningful number of games. With a few dozen games per band you see rough patterns; with hundreds you can trust the small gaps.
Does Quant+ show calibration?
Yes, every Quant+ model has a scorecard that includes calibration. Quant+ does not place bets or promise results, and it is for adults 21 and over.
Sources
- Machine learning for sports betting: should model selection be based on accuracy or calibration?, Machine Learning with Applications 16 (Walsh, Joshi), University of Bath
- Moneyline odds, Wikipedia
- 2026 Major League Baseball season, Wikipedia
- The Betting Odds Rating System, PLOS ONE (Wunderlich, Memmert)
- Beating the bookies with their own numbers, arXiv (Kaunitz, Zhong, Kreiner)
