DEV Community: Quant+

Backtest a betting idea before you trust it

Quant+ — Wed, 14 Oct 2026 20:23:07 +0000

Originally published at quantplus.win.

Sports betting backtesting software runs a betting rule over past games using only the prices and information available at bet time, then tells you whether the result could be luck. The quant+ Backtest Lab blocks look-ahead in the form, reports ROI with a bootstrap interval, closing line value and drawdown, and scores each rule season by season with walk-forward testing.

Sports betting backtesting software runs a betting rule over past games, using only the prices and information that existed when each bet would have been placed, and tells you whether the result could be luck. The quant+ Backtest Lab does that for NFL, college football, NBA, college basketball, MLB and NHL sides, totals and moneylines. It blocks look-ahead in the form itself, reports intervals instead of a single number, and scores the rule season by season with walk-forward testing.

Most backtests that look great are wrong in one of three ways. They peek at the future, they test too many ideas, or they ignore how small the sample is. This page covers how the lab handles each.

How you define a strategy

A backtest in the lab is a rule you build from a form, not code:

Scope: leagues, season range, optional date range, regular season or playoffs.
Market: spread, total or moneyline.
Side: home, away, favourite, underdog, over, under, with the line move or against it.
When the bet is placed: when lines open, or at kickoff (the closing price).
Which price: the consensus, one named book, or the best available.
Filters: week, venue, favourite or underdog, divisional game, neutral site, rest advantage, spread and total ranges, how far the line or price moved, temperature, wind and roof.
Staking: flat units, a percent of bankroll, or fractional Kelly.

Two presets show the idea: "Home dogs +3 to +7" and "Unders in 15+ mph wind". Change one field and run it again.

How the lab prevents look-ahead bias

Look-ahead bias means the backtest uses something you couldn't have known at bet time. In betting it usually hides in one of these:

Using the closing line to choose a bet, then pretending you got the opening price.
Filtering on line movement while betting at the open, before the move happened.
Using game-day weather for a bet placed on Tuesday.

Machine learning has the same trap. The scikit-learn documentation warns that ordinary cross-validation on time-ordered data would lead to training on future data and evaluating on past data, which is why time-series splits train only on what came before.

In the quant+ lab, every input is tagged with when it becomes known. The schedule and the line are known at the open. Line movement, weather and roof status are only known near kickoff. If you choose "bet when lines open", the form marks movement and weather filters as "close only" and won't run them. The price taken can never come from before the moment your rule could have fired. That check is enforced by the engine and covered by tests, so it doesn't depend on you remembering.

Why one ROI number isn't enough

A rule that won 56 percent of 80 bets sounds good. It mostly tells you that 80 bets is a small number.

The lab reports:

Profit and ROI with a 95 percent bootstrap interval: the range of results you'd plausibly see if the same edge produced a different run of games. If that range comfortably includes zero, you haven't found anything yet.
A p-value against break-even, so you know how surprising the result is if the rule had no edge at all.
Win rate against the break-even rate for the prices you actually took.
Closing line value, when the rule bets before the close. A rule that wins without beating the close is usually a lucky rule.
Max drawdown and the longest losing and winning streaks, plus the equity curve.
Walk-forward by season: the rule is scored on each season using an expanding window of earlier seasons, the way you would actually have used it.

The interval and p-value are hidden on very small samples. Showing a confidence interval on a dozen bets would be false precision.

The multiple-testing trap

The quickest way to fool yourself is to try twenty filters and keep the best one. If you test 20 unrelated rules that each have no edge, the chance that at least one passes a 5 percent significance test by luck alone is 1 − 0.95^20, about 64 percent (our arithmetic).

Finance has studied this for years. Bailey, Borwein, López de Prado and Zhu found that standard defences such as hold-out samples tend to be unreliable and inaccurate in the context of investment backtests, and proposed measuring the probability that a backtest is overfit. Betting rules suffer the same problem with far smaller samples.

What we suggest:

Write the rule down before you run it, with the reason you expect it to work.
Count every variation you tried. The lab keeps "How this run was made" (engine version, seed, resamples, de-vig method, data sources) with each run, so you can retrace it.
Trust walk-forward over the full-sample number. If the rule only works in one season, it doesn't work.
Check the close. A real edge usually shows up as positive closing line value.

Where the data comes from

The lab uses historical closing lines from nflverse for the NFL, opening and closing main lines from OddsPapi across the six leagues, college lines from CollegeFootballData and CollegeBasketballData, and quant+'s own price collector, which stores every change on the board. Each run lists the sources it used. NFL history goes back furthest; other leagues have shorter histories, and the lab tells you when a sample is thin.

Player props are not in the lab yet. Prop prices and box scores are on the props board, but we won't publish prop backtests until the historical prices are deep enough to be honest about.

Limits that apply to every backtest

Past prices aren't future prices. Markets adapt. A 2024 study of NFL money lines from 2007 to 2016 found inefficiencies, but noted they were unpredictable and varied by season.

You may not get the stake. Books limit accounts that keep beating them; researchers betting a market-based strategy had stakes capped within months, one to $10.45 on a $50 bet. A backtest assumes you got on in full.

A model can be worse than the market and still useful, or better on paper and useless. Hubáček and Šír showed a model can find value while less accurate than the market if it's decorrelated from it. The backtest is where you check which one you have.

quant+ never places bets or connects to sportsbook accounts. Read the risk disclosure before acting on any result.

Run your first backtest

Pick one idea you already believe in. Run it with walk-forward on, bet at the open, and look at three things: the interval, the closing line value and the worst drawdown. Then log any live bets in the journal so you can compare the real thing with the test. The quant+ guides explain closing line value and fair odds if those are new, and the +EV scanner uses the same de-vig methods as the lab.

Start free to open the Backtest Lab, or see it on the home page.

Quant+ is analytics for informational purposes. 21+. If you or someone you know has a gambling problem, call 1-800-GAMBLER.

Questions people ask

What is the best software for backtesting sports betting strategies?

Look for software that stops you using information from after the bet, shows a confidence interval rather than one ROI number, scores rules season by season and reports closing line value. Without those, a backtest mostly measures luck and look-ahead.

How many bets do you need for a backtest to mean anything?

More than most people think. A few dozen bets can swing wildly by chance. Look at the interval: if the plausible range of ROI includes zero, the sample hasn't shown an edge yet.

What is look-ahead bias in sports betting?

It is using information you couldn't have had when the bet was placed, such as filtering on line movement or weather while assuming you got the opening price. It makes almost any rule look better than it is.

Is it illegal to use AI or software for sports betting research?

Using analytics software to research prices is generally legal where sports betting is legal, but sportsbooks set their own terms and can limit accounts. Check your state's rules and each book's terms.

Can I backtest player props in quant+?

Not yet. The lab covers spreads, totals and moneylines. Prop backtests will come once historical prop prices are deep enough to test honestly.

Sources

TimeSeriesSplit, scikit-learn documentation
The probability of backtest overfitting, Journal of Computational Finance (eScholarship, Lawrence Berkeley National Laboratory)
An Examination of the Money Line Market for National Football League Games, Journal of Economic Insight
Beating the bookies with their own numbers, and how the online sports betting market is rigged, arXiv
Beating the market with a bad predictive model, arXiv

Quant+ provides sports market analytics for informational and educational purposes. It does not offer picks or guarantees. 21+. If you or someone you know has a gambling problem, call 1-800-GAMBLER.

Sports betting models that publish a fair price

Quant+ — Wed, 14 Oct 2026 17:51:56 +0000

Originally published at quantplus.win.

A sports betting model should output probabilities and fair odds for every game, not picks, and it should prove itself with calibration and results against the closing line. quant+ models publish a fair spread, total and moneyline (and prop probabilities at any line), show that price next to the board, and keep a public scorecard with walk-forward results and calibration charts.

A sports betting model should output a price, not a pick. Give it a game and it should tell you the probability of each outcome and the fair odds that go with it. Then you compare that fair price with what the books are offering and decide for yourself. quant+ models work that way: each one publishes a fair spread, total and moneyline (or, for props, the probability of the over at any line), and its scorecard shows how well those prices have held up against the closing line.

If a model can't show you its calibration and how it does against the close, you're being asked to trust it on faith.

What a model output should look like

Here is what the quant+ game models produce for every game they cover:

Team ratings (overall, offense and defense).
A predicted margin and total.
Fair prices: the spread, total and moneyline probabilities those predictions imply.

Prop models produce the probability of the over at any line, so a 6.5 strikeout prop and a 7.5 alternate get their own fair prices instead of one guess at the median.

On each game page, the model's fair price sits in its own column next to the board. The cell turns green when the best available price is better than the model's fair price. That's the whole output. No "best bet of the day", no confidence stars, no units.

How a model price turns into an edge

Take a moneyline where the model says the home side wins 55 percent of the time. Here is the arithmetic (an illustration, not a result).

Step	Model view	Market view
Fair win probability	55.00%	50.00%
Fair American odds	about -122	+100
Best price available	-110	-110
EV per 100 staked at -110	0.55 × 90.91 − 0.45 × 100 = +5.00	0.50 × 90.91 − 0.50 × 100 = −4.55

Same bet, two opposite answers. Everything depends on whose probability is closer to the truth. Markets are hard to beat as forecasters: a PLOS ONE study found betting odds predicted results better than various quantitative models. So when a model disagrees with the market by 5 points, the default assumption should be that the model is wrong, until its track record says otherwise.

Calibration matters more than accuracy

Accuracy asks how often the model picked the winner. Calibration asks whether its probabilities mean what they say. The scikit-learn documentation puts it simply: among the games a well-calibrated model gives a probability close to 0.8, about 80% should actually win.

For betting, calibration is the one that pays. A model that picks winners often but calls every favourite 75 percent will keep telling you short prices are good value. Walsh and Joshi tested this directly on NBA seasons with published odds and found that choosing models by calibration rather than accuracy led to greater returns in their betting experiments (Machine Learning with Applications, 2024).

A quick check you can run on any model: take the 200 games it priced near 60 percent. About 120 should have won. If 100 won, the model is overconfident there, and every "edge" it shows in that range is inflated.

What the quant+ model scorecard shows

Every model has a detail page with:

Margin error and moneyline Brier score (the average squared gap between the forecast probability and what happened; lower is better).
Beats the close: how its prices compared with the closing line.
Walk-forward by season: each season scored by a version trained only on earlier seasons.
Calibration charts for moneyline, spread and total.
Against the spread and totals results on games where the model disagreed with the close, shown with a 95 percent interval and the break-even line.
Team ratings and a versions table, so you can see what changed between releases.

Prop models are graded on held-out games against a naive baseline: the player's last-ten average. If a prop model can't beat that, it shouldn't be trusted on anything.

We publish these numbers even when they aren't flattering. A model that roughly matches the close is still useful as a second opinion and a sanity check. It is not a reason to bet more.

Using a model that's less accurate than the market

A model doesn't have to beat the market everywhere to be useful. Hubáček and Šír showed that a model can be less accurate than the market and still find value, if its errors are decorrelated from market prices. The reverse warning matters too: a model that mostly copies the market adds nothing, because it will agree with the books exactly where they're right and wrong.

In practice:

Start with the market's fair price. The +EV scanner does this for every line.
Use the model as a filter, not a trigger. A price that is good against the market and the model is more interesting than one that is good against only one of them.
Check the scorecard for that market. A model calibrated on NFL sides may be weak on totals.
Log every bet and watch your closing line value. If your model-backed bets don't beat the close over a few hundred bets, the model isn't helping.

Coverage and limits

Leagues: NFL, college football, NBA, college basketball, MLB and NHL. Game models go live league by league once each one has enough graded results to publish a fair scorecard; NFL comes first. Prop models cover the main stat markets (yards and receptions in football, points, rebounds and assists in basketball, strikeouts and total bases in baseball, shots on goal in hockey).

What the models don't do: they don't pick bets, they don't place bets and they don't connect to any sportsbook account. They can be wrong for weeks at a time, and injury news can make a pre-game price stale within minutes. Read the risk disclosure before you use any of it.

If you want to test a rule built on model output before using it, the backtest lab runs it on past seasons with no look-ahead. The quant+ guides cover fair odds and closing line value first.

Start free to open the model pages, or see the terminal on the home page.

Quant+ is analytics for informational purposes. 21+. If you or someone you know has a gambling problem, call 1-800-GAMBLER.

Questions people ask

Do sports betting models actually work?

Some do in some markets, many don't, and you can only tell from an honest track record. Look for calibration charts, results scored only on games the model hadn't seen, and a comparison with the closing line. Win-loss records over a few weeks tell you almost nothing.

What is the most accurate sports betting model?

Accuracy is the wrong yardstick. Published research on NBA betting found that selecting models by calibration led to better betting results than selecting by accuracy. Ask whether a model's 60 percent games win about 60 percent of the time.

Can I use AI picks instead of a model?

A pick hides the probability, so you can't tell whether the price is any good. A model that publishes a fair price lets you compare it with every book and check its record yourself.

How do I test a sports betting model?

Score it only on games it was not trained on, season by season, and compare its prices with the closing line. Check calibration in each probability range and count how many rules you tried, because testing many filters on the same data produces lucky winners.

Does quant+ place bets based on its models?

No. quant+ is analytics only. It shows model prices beside the market and never places bets or connects to sportsbook accounts.

Sources

The Betting Odds Rating System: Using soccer forecasts to forecast soccer, PLOS ONE
Probability calibration, scikit-learn documentation
Machine learning for sports betting: Should model selection be based on accuracy or calibration?, University of Bath (Machine Learning with Applications)
Beating the market with a bad predictive model, arXiv