Betting Model

What a model produces

A betting model turns information available before an event into a probability for a defined outcome: the home team wins, the favorite covers −3.5, the game goes over 47.5. Some models project a margin or total first and convert it to a probability; others estimate the probability directly. Either way, the output has to be a probability before it can be compared with a price.

A useful model description states the target market, the data used, when each input was known, and how the model will be judged. A model that can’t be described that plainly is hard to trust and harder to fix.

From a projection to a decision

Suppose a model projects the home team to win by 4.0 points and the market has the home team at −3.5, −110. Treating the final margin as roughly normal with a standard deviation of 13.5 points, close to how NFL results have historically scattered around the spread, the chance the home team wins by more than 3.5 is about 51.5%.

The break-even rate at −110 is 52.38%. So a model that likes the home side by half a point still points to no bet, because the projected edge is smaller than the cost of the price. Real NFL margins also cluster on 3 and 7, so the normal curve is a first approximation that needs adjusting near key numbers.

Compare with the market, not just the result

The fairest benchmark for a model is the no-vig market price at the time you would bet. If your model says 58% and the market’s no-vig figure says 50%, the likeliest explanation is that the model is missing something, such as an injury, a lineup change, or a data error. Large disagreements deserve a check before they become bets.

A model that mostly agrees with the market adds little. One whose prices the market keeps moving toward is showing real information, and tracking closing line value reveals that much sooner than wins and losses do.

Test on data the model has never seen

Split the data by time. Build the model on earlier seasons, then score it on a later season without changing anything. Two common measures:

• Brier score: the average of (forecast − outcome)², where the outcome is 1 or 0. A 60% forecast scores 0.16 if the event happens and 0.36 if it doesn’t. Lower is better. Compute the same score for the no-vig market on the same games; a model that can’t beat it has nothing to add.
• Calibration: group forecasts into ranges, such as 55–60%, and check that events in each range happen at about that rate.

Sample size limits what any test can show. A record of 500 bets won at 55% has a standard error of about 2.2 percentage points, so a 95% interval runs from roughly 50.6% to 59.4%, a range that includes win rates below the 52.38% break-even.

The two failures that sink most models

Leakage means using information that wasn’t available when the bet would have been placed: final injury reports, closing lines, or season averages that include the game being predicted. It makes backtests look far better than live results.

Overfitting means tuning settings until the model matches past results, noise included. Keep a simple baseline, such as the no-vig market or a basic team rating, and require the model to beat it on untouched data. The Excel model guide walks through a workbook built this way.