Learn

The Four Gate Test: how to test a sports betting model before you trust it

Here is how to test a sports betting model before you trust it: pass four gates in order, the cutoff (no information from after your bet time), the time split (train on past seasons, test on later ones), calibration (probabilities that match reality) and the market test (beat the no-vig price). In one NBA study, models selected for calibration returned about +34.69 percent over a season while models selected for accuracy returned about -35.17 percent.

Updated October 22, 2026 · 8 min read

The four gate test is a checklist a betting model has to pass before you trust its numbers with real money. It is the checklist from our video on how to build a sports betting model, with the exact check for each gate.

  1. The cutoff. Every input exists before the moment you would place the bet.
  2. The time split. Train on past seasons, test on later ones, never on shuffled games.
  3. Calibration. When the model says 60 percent, those games win about 60 percent of the time, and you choose models by that, not by accuracy.
  4. The market test. The model's probabilities beat the market's no-vig price on the same games.

A model that fails any gate can still look great in a backtest. That is the whole problem: the gates exist because a leaky or overconfident model produces the most convincing charts.

Who this checklist is for

It is for anyone building their own model in a spreadsheet, Python or R, whether it predicts game winners, totals or player stats. You don't need a particular algorithm. The gates are about how you test, and they apply the same way to a logistic regression and to a neural net. If you are buying picks from someone else's model, the same four questions are the ones to ask them.

1. The cutoff

What it means. Pick the moment you would actually bet (for example, the morning of the game, or the opening line). Every feature the model sees for a game must have been known at that moment. Anything published later is off limits for that bet.

Why it matters. This mistake has a name, leakage: using information during training that would not be available when the prediction is made. It makes test results look better than they really are, and nothing in the model's code will warn you.

Common leaks in betting data:

  • The closing line or any price taken after your bet time.
  • Final injury reports and confirmed lineups when you bet the night before.
  • Season averages that include the game being predicted (a rolling average must stop the game before).
  • Team ratings recalculated at the end of the season and then applied to every game in it.

The exact check. For every column, write down when that number becomes known, as a time relative to kickoff. Then sort the columns by that time and delete every column that comes after your bet time. In code, build features with a strict filter such as games[games.date < bet_date] and never from a full-season table.

2. The time split

What it means. Train on earlier seasons and test on a later one, in order. For example, train on 2022 to 2024 and test on 2025. Never shuffle games at random into training and test sets.

Why it matters. A random split lets the model learn from games played after the ones it is tested on. The leakage article calls this time leakage, and it is one of the most common ways a sports model ends up looking better on paper than in a real season.

The exact check.

  • Keep one full season, the latest one, untouched until the very end. Don't look at it while you choose features.
  • If you need several test windows, walk forward: train through season N, test on season N+1, then add N+1 to training and test on N+2.
  • Write the date ranges of every split into your notes so you can prove later that test games always came after training games.

3. Calibration

What it means. A calibrated model's probabilities match reality: across all the games it calls 60 percent, the favorite wins about 60 percent of the time. A model can be accurate (it picks the right side often) and still be badly calibrated (its 70 percents only win 58 percent of the time), and for betting the probabilities are what you trade on.

Why it matters. In a study of NBA betting, Walsh and Joshi compared models selected by accuracy with models selected by calibration over one season, betting at published odds. The calibration-selected models returned about +34.69 percent, while the accuracy-selected models returned about -35.17 percent. That is one season of one sport, so treat it as a strong warning, not a promise, but it shows why the most accurate model is not automatically the one to bet.

The exact check.

  • Draw a reliability diagram: group your test predictions into bins (50 to 55 percent, 55 to 60 percent and so on) and plot each bin's average prediction against how often it actually won. A calibrated model sits close to the diagonal. scikit-learn's calibration guide shows how, with calibration_curve.
  • If you recalibrate (for example with isotonic or sigmoid calibration), fit the calibrator on separate data from the data the model was trained on, as the same guide recommends.
  • Report the Brier score, the mean squared error of probability forecasts. It runs from 0 (perfect) to 1, and a forecast of 50 percent on every game scores 0.25. Keep in mind the scikit-learn guide's note that a lower Brier score does not by itself mean better calibration, so look at the diagram too.

4. The market test

What it means. Compare your model's probabilities with the market's on the same games, after removing the book's margin (the no-vig price). If you can't beat the market's probabilities, your model has no reason to bet.

Why it matters. The market is a strong opponent. A PLOS ONE study of a betting-odds rating system found that ratings built from betting odds had the best predictive quality among the soccer rating systems it compared. Beating a naive baseline is easy; beating the price is the test that matters.

The exact check, with a worked example. Here is a hypothetical with four games. The model says 70, 60, 40 and 80 percent; the no-vig market says 60, 50, 45 and 70 percent; the favorites' results are win, loss, loss, win. The Brier score for the model is 0.1625 and for the market is 0.1756, so on these four games the model's probabilities were closer to what happened. Four games prove nothing; run the same comparison on a full test season, game by game, with the market price taken at the same time you would bet.

Mistakes that get past the gates

  • Re-using the test season. Every time you tweak the model after looking at the test results, the test season quietly becomes training data. Keep a final season locked away.
  • Testing against the wrong price. Compare with the price you could actually have bet at your bet time. Comparing a morning model against the closing price mixes two different moments.
  • Sizing bets on raw model output. The Kelly criterion sizes a bet from your edge, and overestimating the probability pushes the stake above the optimal size. An overconfident model makes every bet too big, which is why calibration comes before staking.
  • Judging the model on a few weeks of results. Small samples swing a lot. Judge it on the gates and on its probabilities over a full test season.

How to tell it worked

Before you bet a dollar, you should be able to show four things: a list of every input with the time it becomes known, the date ranges of your train and test seasons, a reliability diagram on the untouched test season, and a Brier score comparison against the no-vig market on the same games. If any of the four is missing, the model hasn't passed yet.

Once it has, the next step is tracking it live against the price you actually got. The full guide on how to build a sports betting model covers choosing a target and a first model, and closing line value explains how to grade live bets against the final price.

Quant+ is analytics for informational purposes. Nothing is guaranteed. 21+. If you or someone you know has a gambling problem, call 1-800-GAMBLER.

Questions people ask

What is data leakage in a sports betting model?

Leakage is using information in training that would not have been known when you placed the bet, such as the closing line or a final injury report. It makes backtests look better than real results will be.

Should I split my data randomly or by season?

By season. Train on earlier seasons and test on a later one, so the model never learns from games played after the ones it is tested on.

Is accuracy or calibration more important for betting?

Calibration. You bet on probabilities, so they need to match how often things actually happen. In one NBA study, calibration-selected models did far better than accuracy-selected ones over a season.

How do I know if my model beats the market?

Compare its probabilities with the no-vig market price on the same games, taken at the time you would bet, using a score such as the Brier score over a full test season.

How many games do I need to test a model?

At least a full season you haven't touched while building it. A few weeks of games is too small a sample to tell skill from luck.

Sources

  1. Leakage (machine learning), Wikipedia
  2. Machine learning for sports betting: should model selection be based on accuracy or calibration?, arXiv (Walsh, Joshi)
  3. Probability calibration, scikit-learn
  4. Brier score, Wikipedia
  5. Betting Odds Rating System, PLOS ONE
  6. Kelly criterion, Wikipedia