Break-Even Win Rate Calculator

By Alex Thornton · author of the MIT-licensed teachersbettextbook library · Last reviewed 24 July 2026

Two halves. The first is one division: the win rate a price demands, on a sportsbook line or on a prediction-market contract after fees. The second is the part almost nobody publishes — how many settled bets a record needs before that win rate means anything at all.

A break-even win rate is the share of bets that must win for a price to return exactly what it risks. It equals one divided by the decimal odds, so a −110 price (decimal 1.909) needs 52.38% of those bets to land. Points above that rate are edge; points below it are a slow loss.

q = \dfrac{1}{d} = \dfrac{1}{1 + b}, \qquad q_{\text{ml}<0} = \dfrac{|\text{ml}|}{|\text{ml}| + 100}, \qquad q_{\text{ml}>0} = \dfrac{100}{\text{ml} + 100}

where:

$q$ — the break-even win rate, a probability between 0 and 1. It is numerically identical to the price’s vig-inclusive implied probability.
$d$ — decimal odds, total return per unit staked including the stake. −110 is $d = 1.909$ .
$b = d - 1$ — net profit per unit staked. −110 is $b = 0.909$ .
$\text{ml}$ — the American moneyline, in whole units either side of ±100.

On a prediction market the fee moves the same figure, because the fee is paid whether or not the contract resolves your way:

q_{\text{fee}} = P + r\,P\,(1 - P), \qquad r = 0.07\ \text{(taker)}, \quad r = 0.0175\ \text{(maker)}

$P$ is the contract price in dollars — a 50¢ contract is $P = 0.50$ . See the Kalshi fee calculator for the fee itself and the round-trip case.

Worked example — the numbers in the form below

A −110 price, and a 500-bet record with 270 wins:

d = 1 + \tfrac{100}{110} = 1.9091 \quad\Rightarrow\quad q = \tfrac{1}{1.9091} = 52.38\%

\hat{p} = \tfrac{270}{500} = 54.00\%, \qquad \hat{p} - q = +1.62 \text{ points}

z = \dfrac{0.5400 - 0.5238}{\sqrt{0.5238 \times 0.4762 / 500}} = \dfrac{0.0162}{0.02234} = 0.72, \qquad p_{\text{2-sided}} = 0.47

The record is above break-even and the 95% interval on it runs 49.62% to 58.32%, which still contains 52.38%. A two-sided test returns p = 0.47 — about what a fair coin returns. 500 bets is not a sample; it is an anecdote with decimal places.

Closing line value grades the same bets on a fraction of the sample — open the CLV calculator →

Break-even at the prices you actually see

One division, run across the board. Nothing here depends on the sport, the book, or the bet type — only on the price.

American	Decimal	Break-even win rate	Profit per $100
−200	1.500	66.67%	$50.00
−150	1.667	60.00%	$66.67
−130	1.769	56.52%	$76.92
−120	1.833	54.55%	$83.33
−110	1.909	52.38%	$90.91
−105	1.952	51.22%	$95.24
+100	2.000	50.00%	$100.00
+120	2.200	45.45%	$120.00
+150	2.500	40.00%	$150.00
+200	3.000	33.33%	$200.00
+300	4.000	25.00%	$300.00

The standard −110 line is the one worth memorising: 52.38%, not the 52.4% that gets quoted. The difference is 0.019 points — on 10,000 bets of $100 it is the gap between returning exactly the $1,000,000 staked and returning $364 more, which is a rounding artefact wearing the costume of an edge.

The gap between break-even and fair is the vig

A break-even win rate is a statement about a price, not about a team. On a two-way market both sides carry one, and the two do not add to 100%:

q_A + q_B = 1 + \rho

At −110 on both sides, $q_A = q_B = 52.38\%$ and the pair sums to 104.76% — a $\rho = 4.76\%$ overround. De-vig it and the fair probability of each side is exactly 50%. So the 2.38 points between 50% and 52.38% is not a property of the game; it is the book’s margin on that side, expressed as win rate.

This is why a break-even win rate is the wrong number to compare a model against. A model outputs a fair probability, and fair probabilities belong next to other fair probabilities: de-vig the line first, then compare. Comparing a model’s 53% to a raw 52.38% counts the house margin as if it were edge and manufactures a bet out of nothing. The vig calculator reports $\rho$ directly.

Break-even after prediction-market fees

Exchanges charge on the fill rather than in the line, which changes the shape of the number without changing the idea. Kalshi’s standard taker fee is $0.07 \cdot C \cdot P(1-P)$ , so per contract the break-even rises by $0.07\,P(1-P)$ — largest at 50¢, where the contract is most uncertain.

Contract price	Price implies	Taker break-even	Maker break-even
10¢	10.00%	10.63%	10.16%
25¢	25.00%	26.31%	25.33%
50¢	50.00%	51.75%	50.44%
75¢	75.00%	76.31%	75.33%
90¢	90.00%	90.63%	90.16%

Put the two schedules side by side at the same fair probability. A coin-flip market costs 2.38 points of break-even at a −110 sportsbook line and 1.75 points as a Kalshi taker — or 0.44 points as a maker, if the order rests and fills. That is the exchange’s structural advantage stated in the only unit that matters here, and it is a smaller number than the marketing on either side usually implies. Books vs. exchanges works through the rest of the comparison.

How many bets before a record means anything

This is the half of the question that gets skipped, and it is the half that decides whether the first half was worth computing. A win rate is a sample mean of a Bernoulli variable, and Bernoulli variables are noisy: at $q \approx 0.52$ a single bet has a standard deviation of $\sqrt{q(1-q)} = 0.4994$ , which is 49.94 probability points. One bet tells you nothing. The question is how many make it something.

Testing an observed rate against break-even is a one-sample proportion test:

z = \dfrac{\hat{p} - q}{\sqrt{q\,(1-q)/n}}, \qquad \hat{p} = \dfrac{w}{n}

and the honest summary of a record is not the point estimate but the interval around it. The calculator uses the Wilson score interval, which behaves near 0 and 1 where the textbook normal interval breaks:

\dfrac{\hat{p} + \frac{z_{\alpha/2}^{2}}{2n} \;\pm\; z_{\alpha/2}\sqrt{\dfrac{\hat{p}(1-\hat{p})}{n} + \dfrac{z_{\alpha/2}^{2}}{4n^{2}}}}{1 + \frac{z_{\alpha/2}^{2}}{n}}

Running it forward instead — how many bets a given true edge needs before the test can see it — is the standard sample-size formula, at 95% confidence and 80% power:

n \;\approx\; \dfrac{\Big[\, z_{\alpha/2}\sqrt{q(1-q)} \;+\; z_{\beta}\sqrt{p(1-p)} \,\Big]^{2}}{(p - q)^{2}}, \qquad z_{\alpha/2} = 1.960,\ \ z_{\beta} = 0.842

The denominator is the killer. Edge enters squared, so halving the edge quadruples the sample:

True win rate at −110	Edge over 52.38%	ROI per bet	Bets to detect it
53.4%	+1.02 pts	+1.9%	18,845
54.0%	+1.62 pts	+3.1%	7,462
55.0%	+2.62 pts	+5.0%	2,849
57.4%	+5.02 pts	+9.6%	773
60.0%	+7.62 pts	+14.5%	334

ROI per bet is $p\,(1+b) - 1$ at $b = 0.909$ , the same arithmetic the expected value calculator runs. “Bets to detect it” is a two-sided 95% test at 80% power, rounded up to a whole bet — the same figure the panel above prints for a record at that rate.

A 2.6-point edge — genuinely good, a 5% ROI — needs 2,849 settled bets. At 20 bets a week that is nearly three years. A flat one-point edge needs 19,570, which is longer than most bettors stay in the game. The arithmetic is not pessimism; it is the reason serious measurement moved off win rate entirely.

Why closing line value gets there faster

Same skill, different measurement. A win rate observes one bit per bet — won or lost — and pays the full Bernoulli variance for it. Closing line value observes a continuous number per bet: the gap in probability points between the price taken and the de-vigged closing line. Nothing is thrown away, so the standard error collapses.

Put concrete numbers on it, over a 100-bet sample. On win rate, the standard error at $q = 0.5238$ is $0.4994/\sqrt{100} = 4.99$ points, so the 95% interval on the record spans roughly ±9.8 points — wide enough to contain almost any hypothesis. On CLV, if a bettor’s per-bet closing-line gap has a standard deviation of 3 probability points (an order of magnitude that is plausible inside a single sport and market, and measurable from the bettor’s own log rather than assumed), the standard error is $3/\sqrt{100} = 0.30$ points and the same 100 bets pin the mean to about ±0.6 points.

That is the whole argument for the CLV-based report card: it is not a different claim about skill, it is the same claim measured with an instrument carrying roughly one-sixteenth the noise per observation. The dispersion figure above is an assumption, stated as one — the honest version is to measure it from a real log, which is exactly what a bet archive is for.

Second example: a hot streak that is not evidence

60 bets at −110, 38 wins — a 63.3% record, more than 10 points above break-even.

\hat{p} = \tfrac{38}{60} = 63.33\%, \qquad z = \dfrac{0.6333 - 0.5238}{\sqrt{0.5238 \times 0.4762/60}} = \dfrac{0.1095}{0.06447} = 1.70

p_{\text{2-sided}} = 0.089, \qquad \text{Wilson 95\%} = [50.68\%,\ 74.38\%]

A double-digit edge over 60 bets still fails a conventional two-sided test, and the interval reaches from barely-break-even to 74%. The point is not that the bettor is bad. It is that 60 bets cannot tell the difference between a 50.7% bettor and a 74.4% one, so no honest verdict fits on that sample.

What a break-even win rate is not

It is not defined across mixed prices. A single break-even rate exists for one price at one stake. A book of bets at −110, +150 and −300 has no single break-even win rate; the correct condition is $\sum_i s_i\,p_i b_i = \sum_i s_i\,(1 - p_i)$ across the stakes $s_i$ , which only collapses to a single win rate when every bet shares a price and a size.
It is not a fair probability. It is the vig-inclusive number — see the vig section above.
It is not affected by pushes. A push returns the stake and belongs in neither column. Grading a record, pushes come out of $n$ entirely rather than counting as half a win.
It does not survive the parlay. Multiplying legs multiplies the decimal odds, so the break-even rate of a four-leg parlay is the product of four break-even rates — four −110 legs break even at 7.53%, not 52.38%. The margin compounds as well, but multiplicatively in one-plus-the-overround rather than as a product of margins: four 4.76% legs accumulate $1.047619^{4} - 1 = 20.45\%$ . The parlay calculator shows the compounding.

Common mistakes

Quoting 52.4%. The number is 52.381%. The rounded version overstates the requirement by 0.019 points, which is $364 of the wrong sign on 10,000 bets of $100 — small in dollars, and the tell of a number copied rather than computed.
Comparing a model probability to the raw break-even rate. That treats the book’s margin as a target rather than a cost. The comparison the math supports is model probability against the de-vigged fair probability.
Reading a 500-bet record as settled. At 54% it returns p = 0.47 against break-even. Interval, not point estimate.
Ignoring the fee on an exchange. A 50¢ contract does not break even at 50%. As a taker it needs 51.75%, and a round trip pays that twice.
Averaging win rates across prices. A 55% record on plus-money and a 55% record on −300 are opposite results. The figures only combine when they are grouped by price, or when the record is graded on expected value rather than on wins.
Using the sample-size table as a licence to wait. It says how long a win rate takes to become evidence, not how long an edge takes to be real. CLV answers the second question in a fraction of the sample.

Check this number yourself

The break-even rate is one function call in the MIT-licensed teachersbettextbook package — the vig-inclusive implied probability is the same number — and the fee-adjusted version has its own.

pip install teachersbettextbook

>>> from teachersbettextbook import moneyline_to_prob, kalshi_breakeven
>>> moneyline_to_prob(-110)
0.5238095238095238
>>> kalshi_breakeven(0.50)               # 50-cent contract, taker
0.5175
>>> from math import sqrt
>>> q, n, w = moneyline_to_prob(-110), 500, 270
>>> (w/n - q) / sqrt(q*(1-q)/n)          # the z the page reports
0.7248824356...

Same inputs, same digits, no account required. More on what is open and what is not: open source.

Break-even is the floor; everything interesting is measured from it. De-vig the line to see how much of that floor is margin, expected value to price a gap above it, Kelly to size one, and closing line value to find out whether it was ever there.

Frequently asked questions

What is a break-even win rate?

A break-even win rate is the share of bets at a given price that must win for the money staked to come back exactly. It is one divided by the decimal odds, which is the same number as the price’s vig-inclusive implied probability. At −110 that is 52.38%.

What is the break-even win rate at −110?

52.38%. A −110 price is decimal 1.909, and 1 / 1.909 = 0.5238. Staking 100 units on each of 1,000 bets at −110 and winning 523 of them returns 99,845 units against 100,000 risked; 524 wins returns 100,036. The break-even count sits between them, at 523.81.

How many bets does it take to know whether you beat the break-even rate?

More than most records contain. Detecting a true 55% win rate against a 52.38% break-even at conventional 95% confidence and 80% power takes 2,849 settled bets. A flat one-point edge takes 19,570. A 500-bet record at 54% carries a 95% interval of 49.6% to 58.3%, which still contains break-even.

What is the break-even win rate on a prediction market?

The contract price plus the fee. On Kalshi’s standard schedule a taker pays 0.07 × P × (1 − P) per contract, so a 50¢ contract breaks even at 51.75% rather than 50%, and a 25¢ contract at 26.31%. A resting maker order pays a quarter of that, putting the 50¢ break-even at 50.44%.

Does beating the break-even rate prove an edge?

Not on its own, and usually not on the sample sizes people have. A 270–230 record over 500 bets at −110 is 54%, 1.62 points above break-even, and a two-sided test returns p = 0.47 — roughly what a coin does. Closing line value measures the same skill on far fewer bets because each bet contributes a continuous number instead of a win or a loss.