Calibration and Brier Scores for Sports Bettors: Is Your 60% Really 60%?

The quick answer

• Calibration: across many bets you rated at 60%, about 60% should win. If they win 50%, your 60s are too high.

• Brier score: the average of (your probability minus the result) squared, with a win scored 1 and a loss 0. Lower is better. Saying 50% on everything scores 0.25.

• Worked example: a bettor whose 100 bets averaged a 53.75% estimate won 47. Their Brier score was 0.2539, worse than a flat 50%, and the de-vigged market scored 0.2485 on the same bets.

• Sample size: that gap is smaller than one standard error. At the same gap it takes about 450 bets to reach two.

• Reference prices: Novig quotes prices as percentages, so its pre-game prices can be logged as probabilities next to your own.

Profit tells you how a run of bets went. Calibration asks whether the probabilities behind those bets were right, and every +EV calculation depends on them. If you write a number down before each bet you can check it later; otherwise you have only results, which carry a lot of noise over a few hundred bets.

What calibration means

A forecaster is calibrated when outcomes they rate at x% happen about x% of the time. Weather forecasting has used the idea for decades: on days a calibrated forecaster gives a 30% chance of rain, it rains on about 30% of them.

For a bettor, the forecast is your probability that the bet wins, written down before kickoff. Group bets by that number and compare each group's hit rate with the estimate. If your 60% group hits 50%, you're overconfident in that range, and every edge you computed there was overstated by about 10 points of probability.

A bettor who says 50% on every coin-flip NFL side is calibrated and has told you nothing, so calibration alone can't separate a good forecaster from a cautious one. The Brier score rewards both halves: being right on average, and putting distance between the bets that win and the bets that lose.

How we checked

The 100-bet record below is invented to show the calculation. Figures were computed to four decimal places, and the sample-size probabilities are exact, assuming independent bets with pushes left out. The Brier score follows Glenn Brier's 1950 paper in Monthly Weather Review and the reliability split follows Allan Murphy's 1973 paper in the Journal of Applied Meteorology, both checked 12 October 2026.

A calibration table from 100 bets

Here's a season of 100 straight bets, grouped by the probability the bettor wrote down for each one.

The average estimate was 53.75%, so this bettor expected about 54 wins and got 47. The gap grows with confidence: bets rated 45% and bets rated 65% won at about the same rate.

High estimates that win no more often than middling ones are what overconfidence looks like in a table. A betting record leans that way by construction, since you only place the bets where your number beat the price, and the log fills up with your most optimistic reads.

The Brier score formula

For N bets:

Brier score = (1 / N) x sum of (f - o)²

f is your probability that the bet wins, as a decimal. o is the outcome: 1 for a win, 0 for a loss.

Here's how it works on five bets:

The squared errors add to 1.1394. Divided by five, the Brier score is 0.2279.

A perfect forecaster scores 0. Saying 50% on everything scores exactly 0.25, because every bet adds 0.5² whatever happens. A score above 0.25 on two-way bets means a flat 50% would have done better. Squaring makes confident misses expensive: the 62% loss in bet 5 cost more than the 45% loss in bet 4.

For a group of identical estimates, the sum is wins x (1 - f)² + losses x f². For the 60% group above: 10 x 0.16 + 10 x 0.36 = 5.2.

Comparing your score with the de-vigged market

A Brier score means little in isolation, because 0.24 can be good on coin-flip sides and poor on heavy favorites. The useful comparison is the market's probability for the same bets, with the margin removed. Our walkthrough of how to de-vig betting odds covers that step.

For the 100 bets above, suppose the de-vigged market had these sides a little lower than the bettor did, which is why they looked like value:

Divide by 100. The bettor scores 0.2539 and the market scores 0.2485. The market's average probability on these bets was 49.1% against an actual 47.0%, and the bettor's was 53.75%. In every group, the market's lower number scored better than the bettor's higher one.

One way to put the comparison in a single figure is a skill score: 1 minus your Brier score divided by the market's. Here it's 1 - 0.2539 / 0.2485 = -2.2%. A positive number means you've beaten the reference.

Murphy's 1973 split writes a Brier score as reliability (calibration error) minus resolution plus uncertainty. The bettor's reliability term is 0.0061 and the market's is 0.0007; resolution (0.0013) and uncertainty (0.2491) are shared because the groups don't change. So 0.0061 - 0.0013 + 0.2491 = 0.2539, and the whole gap comes from calibration.

The reference has to be a price you'd trust. A sharp, liquid line close to game time works better than an early number at a single book. Prices quoted as percentages save a conversion step. On Novig, a 40% price is the probability itself, and $100 at 40% returns $250. Novig's pre-game straight trades carry no fee for either side, so the logged price is the price paid. Live straights carry a 0.03 taker coefficient (0.06 for live NCAAF, NHL and MLB), and makers pay nothing on any trade type.

Why 100 bets can't settle it

Take the 60% group. If those 20 bets really were 60% propositions, 10 or fewer wins would still happen 24.5% of the time. A 95% band around a 60% hit rate is about ±21.5 points at 20 bets, ±9.6 at 100 and ±4.8 at 400.

The whole season has the same problem. If every estimate had been correct, 47 or fewer wins from these 100 bets would happen about 10.3% of the time. That's unlikely without being proof.

The Brier gap is thinner still. The bettor trails the market by 0.0054 per bet, and the standard error of that gap is 0.0057, so the difference is under one standard error. At the same gap it would take about 450 bets to reach two standard errors. Our piece on variance and sample size runs the same arithmetic on profit.

What to look for

• Record: write your probability before every bet, along with the price you took and the de-vigged reference price at the same moment.

• Buckets: group by estimate in ranges wide enough to hold 20 or more bets, and read the gap column before the profit column.

• Direction: hit rates that stay flat while your estimates rise mean the extra confidence isn't adding information.

• Benchmark: score the market on the same bets. A bettor with no edge should expect to trail it, and over a large enough sample a persistent gap points at the estimate.

• Costs: log the fee or vig on each bet. With no fee, the logged price and the paid price match.

Calibration FAQ

What is a good Brier score for sports betting?

There's no fixed target. On two-way sides near even money, a flat 50% scores 0.25, so beating that is the minimum. Beating the de-vigged market on the same bets is the stronger test.

Is Brier score or log loss better for betting?

Both are lowest when you report the probability you actually believe. Log loss averages minus the log of the probability you gave the actual result, so a confident miss costs much more than it does under Brier. The five bets above score 0.6463 in log loss. Brier is easier to read and split into parts; log loss punishes overconfidence harder.

Where can I get a reference probability to score against?

From a sharp, liquid market with the margin removed. Novig shows prices as percentages, and pre-game straight trades there carry no fee, so its pre-game price can go straight into the log as a probability.

Does good calibration mean I'm profitable?

No. Calibration measures the estimates. Profit also depends on the prices you took and their costs, and a calibrated bettor who only bets where the market agrees has no edge to collect. Closing line value is a separate check on price, covered in our closing line value explainer.

Sources

All checked 12 October 2026.

Brier, "Verification of Forecasts Expressed in Terms of Probability," Monthly Weather Review 78(1), 1950, 1-3, doi 10.1175/1520-0493(1950)078<0001:VOFEIT>2.0.CO;2

Murphy, "A New Vector Partition of the Probability Score," Journal of Applied Meteorology 12, 1973, 595-600

Novig fee schedule: support.novig.com/en/articles/16195057-fees-on-novig

Novig odds calculator: tools.novig.com/odds-calculator