Computer Picks

Sports Betting Models: How They Work and How to Judge One

8 min read
How sports betting models work

Sports betting models are sets of rules and math that turn past results into a forecast for the next game, usually a projected score or a win probability, and then compare that forecast with the betting line. When the model's number and the sportsbook's number are far enough apart, the model picks a side. That's the whole idea, and it's also why most of them lose: the line is already a very good forecast, and beating it by enough to pay the bookmaker's cut is hard.

We run a free sportsbook with its own computer picks model for football and basketball, so this guide uses ours as the worked example. It covers the main types of models, what goes into them, how a rating becomes a pick, how a model is tested, and how to tell a good one from a lucky one by its record.

What a sports betting model actually does

Every model, from a spreadsheet to a machine learning system, does three jobs. It measures how good each team is. It uses those measurements to predict the outcome of a game it hasn't seen. And it compares that prediction with the price on offer.

The third job is the one people forget. A model that says a strong team will beat a weak one isn't telling you anything; the sportsbook knows that too and has made the strong team a 10 point favorite. What you want is a model that says "this team should be favored by 7, not 10," because that gap is where a bet comes from. A bettor who follows a predictive model is really betting that the model's number is closer to the truth than the line is.

The main types of betting models

You'll see a lot of names for these, but most sports betting models fall into a handful of families.

Rating models

These give every team a number (or two: one for offense and one for defense) and update it after every game. Power ratings and Elo style systems are in this family. To predict a game you compare the two teams' numbers, add something for home field, and get a projected margin. They're simple and easy to check, and they hold up better than you'd expect. Ours is one.

Regression models

A regression model takes a pile of inputs (yards per play, turnovers, pace, shooting, rest days) and finds the weights that best explain past results. The output is a projected margin, a projected total, or a probability. Regression is the workhorse of sports data analysis because you can see exactly which inputs the model leans on.

Simulation models

A simulation plays the game thousands of times with random variation, possession by possession or drive by drive, and counts how often each side wins or covers. The result is a full distribution of outcomes instead of one number, which helps with things like totals and live betting. The catch is that the simulation is only as good as the inputs behind each possession.

Machine learning models

Machine learning models (gradient boosting, neural networks and the like) let an algorithm find the patterns itself in a large dataset. They can pick up interactions a person wouldn't think to test. They can also memorize noise very convincingly, which is the biggest reason they look brilliant in testing and fall apart on real games. We go further into this, and into what "AI picks" usually means in practice, in our guide to AI sports picks.

Market models

Some bettors skip predicting the game and model the betting markets instead: how lines move, which books move first, where the closing line ends up. That's a different job from forecasting game outcomes, and it's mostly beyond what a casual bettor can do with public data.

What goes into a model

The inputs depend on the sport, but the usual list is final scores, efficiency stats (points per possession, yards per play), strength of schedule, home field, rest and travel, injuries, and for outdoor sports, weather. The data science side of this is mostly cleaning: team names that don't match, games played at neutral sites, a season cut short.

More inputs don't make a better model. Every extra input is another chance to fit the past too closely. A model built on final scores alone, adjusted properly for the opponent, is a surprisingly hard baseline to beat. Good feature selection usually means taking inputs out.

From ratings to a projected score

Here's how our model works, step by step, because it's a clean example of a rating model.

Each team gets an offense rating and a defense rating built from its final scores over the last two seasons, with recent games counting more than old ones. Each rating is adjusted for the opponent (scoring 30 on a great defense counts for more than scoring 30 on a bad one) and for home field. A team needs at least 4 games played before the model will rate it.

To project a game, the model sets one team's offense against the other's defense, both ways, and adds the home edge. That gives a projected score for each side. Say it projects the home team to win 27 to 20. That's a 7 point projected margin and 47 projected total points. This is a made-up example, but it's exactly the shape of what the model produces.

Projection vs the line: where a pick comes from

Now compare that projection with the board. If the home team is favored by 3.5, the model thinks they should be favored by 7, so it likes the home team to cover the spread. If the total is 50, the model's 47 says under. The side the model favors is the pick, on both the spread and the total, for every game.

How far apart the two numbers are tells you how confident to be. Our model turns that gap into stars from 1 to 5: 5 is a Top pick, 4 Strong, 3 Solid, 2 Lean and 1 a Slight lean. Once a league has 60 or more picks in a year, 5 stars works out to about the top 10% of picks. A small gap is close to a coin flip; a big one is where a model earns its keep, if it's going to.

Remember what you're up against, too. At standard -110 odds you risk $110 to win $100, so you need to win 52.4% of your bets to break even. A model that's right 51% of the time against the spread is a losing betting system, even though it's "right" more often than not. If -110 and spreads are new to you, our list of common betting terms explains them.

How a model is tested

The basic test is a backtest: run the model on past seasons as if you didn't know the results, and grade its picks. The trap is that you built the model on those same seasons, so of course it looks good. A fair test uses games the model has never seen, grades them against the line that was available at the time, and counts every pick, not just the confident ones.

Ours tunes itself in public. Every week it replays the last 6 weeks with other settings, such as how fast old games fade and how far ratings are pulled toward average, and keeps the settings with the smallest miss. Past picks never change when it does that. Picks update until kickoff, then lock, and they're graded on the line at kickoff. If you want to build and test your own, our walkthrough on how to build a betting algorithm goes through it with a spreadsheet.

Why most sports betting models fail

We've watched a lot of models come and go over the years, and they tend to fail the same few ways.

Overfitting is the big one. A model with enough knobs can explain any past season perfectly and predict nothing. If a model's backtest shows 60% winners, that's usually a sign it memorized the past.

The second is ignoring how good the line already is. Sportsbooks have their own models, and the betting public's money moves the line toward the right number. So your model is up against the best forecast available, minus a few cents of vig.

Then there's the sample. A model that went 30 and 20 last month might be good, or it might have had a lucky month. Variance in sports betting is much bigger than most people think, and it takes hundreds of picks before a record starts to mean much.

And plenty of models are graded on the wrong number. Grading a pick against a line that moved two points after the model saw it, or against the best line anywhere, flatters the record. A pick has to be graded at a price you could actually have bet.

How to judge a model by its record

A record tells you more than the method or the marketing ever will, if it's the right kind of record. These are the questions we'd ask of any model, ours included:

  • Is every pick in the record, losers too, and can past picks be changed afterward?
  • Were the picks posted before the game, and graded on the line at that time?
  • How many picks is it? Under a couple of hundred, treat it as a hint.
  • Do the high confidence picks win more often than the low ones? If 5 star picks don't beat 1 star picks over time, the stars mean nothing.
  • Is it split by spreads and totals, by season and by recent stretch, so you can see where it's strong and where it isn't?

That last point is the one most model sellers skip. Our record is free and shows all of it: overall, 4 stars and up, each star level, spreads and totals, this season, the last 7 and 30 days, and week by week. You can see it on the nfl computer picks page, and the same model runs the nba computer picks, the ncaab computer picks and the college football model picks. The computer picks hub has all four leagues in one place.

Using a model's picks sensibly

Even a good model is a starting point. It knows the scores; it doesn't know that the starting quarterback was ruled out an hour ago unless the line tells it. Use it to find games worth a closer look, check the news, and size every bet the same way. A flat unit and a sensible bankroll plan matter more than any star rating, because they keep a cold streak from wiping you out before the model's edge (if it has one) can show.

On CentSports, Full Access members see the picks before kickoff. Everyone else sees the star ratings ahead of time and the pick itself 10 minutes after kickoff, which is enough to check the record for yourself. And since every bet here is with a free balance, you can follow a model for a whole season and see how it would have gone without spending a cent. No model is a sure thing, ours included, and anyone who promises otherwise is selling something.

CentSports is for players 21 and over. Gambling problem? Call 1-800-GAMBLER.

FAQ

Do sports betting models actually work?
Some do, a little, and most don't. The betting line is already a very good forecast, so a model has to beat it by enough to cover the bookmaker's cut, and only a public, long record can tell you whether one has.
What are the best models for sports betting?
There isn't one best type. Rating models built from scores and efficiency numbers are the most common and the easiest to check; regression, simulation and machine learning models can add detail but are easier to overfit. Judge any of them by its graded record, not by the method.
How do sports bettors make models?
They collect past games, build team ratings or a statistical model from them, project a score or a win probability for each new game, and compare it with the betting line. Then they test it on games the model has never seen before trusting it.
What win rate does a betting model need?
At standard -110 odds, 52.4% just to break even. A model that wins 53% to 55% against the spread over many hundreds of games is doing very well; anyone claiming far more than that over a big sample deserves a hard look.

Build a streak today

15 in a row = $100 · 21 in a row = $1,500 · Free to play.

Join free

More from Computer Picks