Learn
The Honest Percentage Test: an NBA betting model checklist in four checks
This NBA betting model checklist is the Honest Percentage Test: four checks that tell you whether a model's percentages can be trusted before you bet on them. Calibration, sample, many tries and the schedule, in that order, with a scorecard you can fill in for any model.
Most model reviews start and end with the record. The Honest Percentage Test starts from a different idea: a betting model is only as good as its percentages, because every price comparison and every stake you size depends on them. Here are the four checks from the video, in the same order, then the scorecard.
| # | Check | The question it answers |
|---|---|---|
| 1 | Calibration | When the model says 60%, do those teams win about six in ten? |
| 2 | Sample | Is the record long enough to mean anything? |
| 3 | Many tries | How many models or rules were tested before this one? |
| 4 | Schedule | Does the model know who is on the second night of a back-to-back? |
Who this is for
This checklist is for anyone deciding whether to trust an NBA model with real money: one you built, one you pay for, or one of ours. You need the model's past probabilities (not just its picks), the results, and the closing prices if you have them. It takes about an hour the first time and a few minutes after that.
Why percentages beat picks
Walsh and Joshi, in a 2024 paper in Machine Learning with Applications, ran NBA betting experiments and found that selecting models by calibration rather than accuracy led to greater returns. Calibration is like a weather forecast that's honest about rain: when it says 60%, it rains on about six days out of ten. A model with honest percentages can be compared with prices. One with dishonest percentages can't, however many winners it picks.
Check 1. Calibration
- Take every past game where the model gave a probability for a side.
- Group the games into bands (50 to 54%, 55 to 59%, 60 to 64%, 65% and up).
- For each band, divide wins by games and compare with the band's average prediction.
- Pass if every band with enough games is within a few points of its prediction. The Walsh and Joshi results are the reason this check comes first.
A band that says 60% and wins 52% is the most expensive failure there is, because it makes the model look like it has an edge exactly where it doesn't. Odds are a strong benchmark here: Wunderlich and Memmert note that betting odds are frequently found to outperform models, so compare the model's calibration with the market's no-vig chances too.
Check 2. Sample
A full NBA regular season has 1,230 games, since each of the 30 teams plays 82. At -110 the break-even win rate is 52.38% (moneyline odds). The table shows, for a hypothetical model whose true skill is 55%, how often its record would still sit below break-even.
| Bets | Chance the record is below 52.38% (true skill 55%, hypothetical) |
|---|---|
| 100 | about 30% |
| 500 | about 12% |
| 1,230 (every game of one season) | about 3% |
So a model's first hundred bets can easily look like failure, or like genius, for reasons that have nothing to do with skill. Pass this check only when the record covers enough bets for its claimed edge, or when it is backed by closing-line results, which you can grade on every bet with our closing line value calculator.
Check 3. Many tries
You'd think a backtest that looks great is proof. But if you test 20 rules that have no edge at the usual 5% significance level, the chance that at least one looks significant is one minus 0.95 raised to the 20th, about 64%. Test 50 and it is about 92%. This is the arithmetic of repeated tests, and it is the check most model sellers skip.
What to ask:
- How many models, features or thresholds were tried before this one?
- Was the final model tested on games it never saw (walk-forward, season by season)?
- Did the result hold in each season, or only in total?
Pass if the model was chosen before the test games and holds up walk-forward. Our backtest lab reports bootstrap intervals for exactly this reason.
Check 4. Schedule
Rest changes NBA games. In a study of the 1,230 regular-season games of 2016-17, Yang and colleagues found that points in the paint and three-point scoring were significantly lower in back-to-back games than in games with one or two days of rest. A model that doesn't know who is on short rest is missing something real.
Pass if the model includes rest days (or back-to-back flags) for both teams, and if its calibration holds on back-to-back games specifically.
The scorecard
| Check | What you looked at | Pass / fail | Notes |
|---|---|---|---|
| 1 Calibration | bands vs actual win rates | ||
| 2 Sample | number of bets, closing-line results | ||
| 3 Many tries | how the model was chosen, walk-forward results | ||
| 4 Schedule | rest features, back-to-back calibration |
A model you trust should pass all four. One that fails calibration should not be used for prices at all, whatever else it passes.
What to watch out for
- Records without probabilities. If a service shows picks but not percentages, you can't run check 1.
- Total-only results. A great total can hide one lucky season and three flat ones.
- Late news. Injury and rest news can land after a model runs; check whether it updates.
How to tell it worked
You know the test is doing its job when it changes a decision: a model you were about to trust fails a check, or a quiet model passes all four. Every Quant+ model publishes a scorecard with calibration and walk-forward results by season, so you can run this test on ours as well.
21+. For informational purposes only. Nothing is guaranteed. If gambling stops being fun, call 1-800-GAMBLER.
Questions people ask
Why not just pick the model with the best record?
Because records over a few hundred bets are mostly luck, and a model can pick winners while its percentages are wrong. Walsh and Joshi (2024) found that choosing NBA models by calibration rather than accuracy led to greater returns.
How many NBA games are in a season?
Each of the 30 teams plays 82 regular-season games, so there are 1,230 games in total.
What is the many tries problem?
If you test many rules or models, some will look good by chance. With 20 independent tests of rules that have no edge, there is about a 64% chance at least one passes a 5% significance test.
Does Quant+ publish this for its own models?
Yes. Every Quant+ model has a scorecard with calibration, beats-the-close and walk-forward results. Quant+ does not place bets or promise results. 21+.
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
- Influence of Match Congestion on Performances in the National Basketball Association, Frontiers in Psychology (Yang et al.)
- 2025-26 NBA season, Wikipedia
- Moneyline odds, Wikipedia
- Multiple comparisons problem, Wikipedia
- The Betting Odds Rating System, PLOS ONE (Wunderlich, Memmert)
