How Accurate Are Betting Markets? We Checked 5,108 Closing Lines
8 min read · Last updated 2026-07-12 · By the SharpBetz team
Sportsbooks are often described as “efficient markets” — implying the line is close to the truth every time. That’s not quite what efficiency means, and the gap between those two ideas is exactly where sports betting lives. We pulled 5,108 NCAAB games from our history database with a recorded closing spread and compared that spread to what actually happened.
What we measured
For every game, we took the closing spread (the market’s final, most informed number before tip-off) and compared it to the actual final margin. If the spread said the favorite by 6.5 and the favorite won by 6.5, that’s a perfect call. We measured two things across the full 5,108-game sample: the average absolute error (how far off the number was, ignoring direction) and how often the favorite actually covered.
| Metric | Value | Sample |
|---|---|---|
| Avg. absolute error (points) | 8.9 | n=5,108 games |
| Favorite cover rate | 47.4% | n=5,108 games |
Two numbers, two very different lessons.
Markets are unbiased, not precise
An average absolute error of 8.9 points sounds bad until you separate two different questions: “is the market biased?” and “is the market precise?”
The 47.4% favorite cover rate answers the first question, and it’s close to what you’d expect from an unbiased market: favorites and underdogs split close to evenly against the spread (a small dog-side lean, 52.6% for underdogs, which lines up with public research showing bettors slightly overvalue favorites — the “public bias” sportsbooks are well aware of and price for). If the market were systematically wrong in one direction — favorites covering 60% of the time, say — that would mean the spread was too soft, and beating it would be trivial. It isn’t. The market is doing its job of setting a number that splits action close to 50/50.
The 8.9-point average error answers the second question, and it tells a different story: on any single game, the “right” number and the market’s number can be quite far apart. A market that’s right on average can still be wildly wrong on any individual night — a star player’s off night, a hot shooting stretch, a bad whistle, foul trouble. Unbiased and precise are not the same property, and conflating them is how people talk themselves into overconfident bets.
Why this matters more than it sounds like it should
That 8.9-point noise band is the entire reason parlays feel winnable and mostly aren’t. If a single spread carries roughly 9 points of expected error, stacking multiple uncertain outcomes together compounds that uncertainty fast — the parlay “feels” like a modest ask (just three favorites covering) but is statistically a much bigger claim than it appears, because each leg individually already carries that much game-to-game noise.
It’s also why a single hot streak — 8 wins in 10 picks — proves very little. With this much per-game variance built into the sport itself, a short win streak is exactly what you’d expect from a coin flip on a good run, not necessarily evidence of skill. Our results page shows our full track record over a large enough sample to be meaningful, rather than cherry-picked hot stretches, for exactly this reason.
The real bar: 52.38%, and why it’s brutal
Sportsbooks charge a commission on losing bets, typically -110 odds: bet $110 to win $100. That vig means breaking even against the spread doesn’t require winning 50% of your bets — it requires winning:
110 / (110 + 100) = 52.38%That’s the number every bettor actually has to clear, not 50%. It looks like a small gap — just 2.38 percentage points above a coin flip — but given the 8.9-point noise band above, closing even that gap consistently is genuinely hard. It’s why “I hit 8 of my last 10 picks” from a random tout account on social media is nearly always noise, not skill: an 8.9-point average error means a small number of picks tells you almost nothing, and a service needs hundreds of graded picks before its win rate means anything statistically.
This is also why we built the /results page the way we did — full transparency on every graded pick, not a highlight reel — and why our model documentation is public about what the model does and doesn’t account for. A service that won’t show you its full sample, win and loss, isn’t one you should trust with money.
What this means for your betting
- Don’t mistake “unbiased” for “precise.” The market gets the direction right on average (47.4% favorite cover is close to the expected split) but misses the exact number by 8.9 points on average — plenty of room for a specific game to go either way.
- 52.38% is the real target, not 50%. Any claimed win rate below that, after vig, is a losing service — do the math before you believe a track record.
- Small samples lie. With ~9 points of average error per game, a 10-game win streak is well within the range of luck. Look for hundreds of graded picks, not a screenshot.
- Parlays multiply noise, not just odds. Each added leg stacks another game’s worth of that 8.9-point variance onto your bet.
All figures from the SharpBetz history database: 5,108 NCAAB games with a recorded closing spread. Vig math (52.38% breakeven) is standard -110 pricing, publicly verifiable. Favorite/underdog cover split is drawn from the same 5,108-game sample.