Why Sports Betting Models Get It Wrong

A model that is working perfectly will still be wrong constantly. That's the part most people never make peace with, and it's why so many bettors abandon a sound process after a bad week and adopt a worse one after a lucky one.

Some of a model's misses are just the cost of predicting uncertain things. Others are real, fixable failures. Telling the two apart is most of the skill.

Wrong is not the same as bad

Take 100 bets where the true probability of winning each one is 60%. You should expect around 40 losses. You should also expect that number to wander: the standard deviation on 100 bets at 60% is about five, so finishing with 45 losses is completely ordinary, and 50 isn't shocking.

Nothing is broken in that scenario. A calibrated model's job is to get the probabilities right over many bets, not to win the one you're watching.

Streaks follow the same logic, and they're longer than intuition expects. Over 100 bets at a true 60%, the chance of hitting at least one run of five straight losses is about 46% — close to a coin flip. Stretch it to 250 bets and a run of six or more is just as likely. A bettor who quits the first time that happens is quitting on arithmetic, not on the model.

This is also why short samples are so misleading in both directions. A tipster who goes 8-2 has told you almost nothing. Neither has a model that goes 2-8.

The information that never reaches the model

A model can only use what's in its data when it runs. A player nursing an injury the team hasn't disclosed, a flu going through a locker room, a starter scratched after the projections were built — all of it affects the game, and none of it is visible to the model.

Timing makes this worse than it sounds. Even public news takes time to reach a model. A lineup confirmed ten minutes before tip-off can make a projection generated that morning stale, through no fault of the math.

When history stops applying

Models learn from the past, which quietly assumes the future behaves like it. Usually that's close enough. Sometimes it very much isn't.

When Major League Baseball introduced the pitch clock and bigger bases in 2023, stolen base attempts jumped and game lengths dropped sharply. A model trained on the years before had never seen a league that played like that. The NFL's redesigned kickoff in 2024 changed field position in ways no historical dataset could describe. A new coordinator who throws out the old playbook, a team that turns over half its roster, a matchup between two sides that have never met — each one pushes a model toward guessing.

The start of a season is a milder version of the same problem. There isn't much current data yet, so models lean on last year, and last year's team may not exist anymore. We wrote a separate piece on the early-season problem.

Overfitting: learning the noise

Give a flexible model enough variables and it will find patterns that aren't there. A team that covered as a road underdog in 11 of 13 games during the training window looks like a signal. More likely it's a coin that happened to land heads a lot, and the model has memorized the coin.

Overfit models are dangerous because they look excellent right up until they're used. Their backtests are strong, and their live results aren't. Honest validation on data the model never saw is the main defense, and it reduces the problem without eliminating it.

Averages hide the spread

A projection is usually a center point, and real outcomes scatter around it. A player averaging 22 points with a standard deviation of six will regularly score 14 and regularly score 30. A model that projects 22 can be right about the player and wrong about the night.

This matters more for player props than for game lines, because a single player's output swings more than a team total, which averages out many players' good and bad nights. It's one reason props need wider margins and smaller stakes, which we get into in player props and the variance problem.

The sportsbook has a model too

Books are not passive. They employ analysts, run their own models, and watch where sharp money goes. They move lines as information arrives. Markets aren't perfectly efficient, but they're efficient enough that easy edges don't survive long; once a pattern is widely exploited, the price adjusts and the edge is gone.

That makes historical backtests optimiztic almost by construction. They show what a strategy would have earned against old prices, not against a market that would have reacted to it. Line movement is its own tricky signal, covered in reading line movement.

Models drift

A model can be well built in March and quietly worse by November. Player roles change, a league's scoring environment shifts, a data provider changes how it records a stat. None of that announces itself. The projections keep arriving with the same formatting and the same apparent confidence.

The defense is to keep checking the model against what actually happens — whether calls it rates at 60% are still winning about 60% of the time, broken down by sport and bet type — and to retrain when they aren't, rather than waiting for a bad month to force the question.

What to do about it

You can't remove uncertainty, but you can stop it from wrecking you.

Treat disagreement as information. When independently built models reach different conclusions about the same game, that split tells you the situation is genuinely unclear, and you should bet less or not at all. It's the core idea behind ensemble models, and it's why the PropJuice app shows how strongly our models agree alongside every projection, instead of presenting every call with the same confidence.

Size every bet as though a losing streak is coming, because one is. A stake you can't afford to lose five times in a row is too big, whatever the edge looks like.

Be suspicious of numbers that look too good. An enormous edge on a major market is more often a model error than a gift. Thinking about edge covers what realiztic edges look like.

And keep records by category. If a model's projections for one bet type keep missing in the same direction, that's a pattern worth acting on. If they miss randomly, that's variance, and changing course would be reacting to noise.

For how we approach a league where a lot of this bites, see our NFL predictions page. Every PropJuice plan starts with a 7-day free trial.