What a tackle prop is, and what we thought we had
Sportsbooks post a number for individual defenders before kickoff. Fred Warner, 7.5 tackles. You take the over or the under. That's a tackle prop, and it's one of the most popular bets on an NFL Sunday because it asks about one player instead of who wins the game.
We built a projection for those numbers. It reads how often a defender is actually on the field, how many plays the opposing offense is likely to run, and how often that offense chooses to run the ball rather than throw it. When our projection disagreed with the sportsbook's number by enough, we counted it as a bet.
Then we tested it on two seasons of games that had already been played, using the prices the sportsbooks actually posted at the time. Across 752 bets it returned 26.2 cents of profit for every dollar risked, winning 67.4% of the time.
Those numbers are absurd. A genuinely good sports bettor makes 3 to 5 cents on the dollar and is thrilled about it. Twenty-six would be a career.
They were also wrong, and the thing that exposed them was a single test we ran last instead of first.
The checks a number like that has to pass
A result that good is almost always a mistake, so we keep a list of the ways it is usually a mistake.
The first check is price. Winning often and making money are different things. A bet priced at −150 has to win three times in five just to break even, so a 57% win rate there is a slow, confident loss. We grade everything in money at the price that was posted, never in win percentage, and we check the result separately inside each band of prices so a market can't hide a losing payout structure behind a pretty hit rate. Tackle props made money in every band.
The second check asks whether the pattern gets stronger as the disagreement gets bigger. If our projection is genuinely better than the sportsbook's, our best bets should be the ones where we disagree with it most. A single profitable pocket surrounded by losses is noise. Tackle props climbed at every step, across nine levels of disagreement, from the smallest to the largest.
The third is sample size. Seven hundred fifty-two bets is enough to mean something, and so was every slice we cut it into.
Six checks. All six said yes.
The test that doesn't ask anything about your model
Then we ran the seventh.
Take the under on every tackle line in the sample. Every one. No projection, no disagreement, no thought of any kind — just bet the under, always.
That returned 26.4 cents on the dollar, winning 67.1% of the time.
Our model returned 26.2 cents, winning 67.4%.
The same. Whatever our projection was doing, it was landing in exactly the same place as a person with one idea and no information.
That comparison is called a blind baseline, and the logic behind it is simple. Before you can say your model found something, you have to know what somebody betting that market with no model would have earned. If the two numbers match, your model hasn't found an edge. It has found the side that happened to keep winning, and taken the credit.
It came down to half a tackle
That leaves a stranger question. Why would blindly betting the under win two thirds of the time at all? A sportsbook posting numbers anyone could beat by always picking one side would not stay in business for a weekend.
The answer is that it depends who counts the tackles.
When two defenders bring a runner down together, somebody has to decide how to credit it. The NFL's official scorers make that call one way in the game book — the certified box score the league publishes, and the one sportsbooks settle these bets against. The public research data we graded our test on makes it a different way. Neither is wrong. They are two careful sources measuring a genuinely fuzzy thing and landing in slightly different places.
The gap is about half a tackle per player per game.
On a line of 3.5 or 4.5 tackles, half a tackle is the entire bet. Our projections sat a little below the sportsbook's numbers for nearly every player, all season, because they were built on the source that counts fewer of them. So every time we disagreed enough to place a bet, we were below the line. Below the line means under. We were never finding mispriced players. We were finding the same measuring-stick difference, over and over, and calling it an edge.
Re-grading the identical bets against the official source is the proof. Tackles-plus-assists falls from 6.9 cents of profit per dollar to a 1.8 cent loss. The scaled version falls from 9.1 cents to roughly break even. Only about 6% of the bets flip from a win to a loss or back — but that 6% is where all of the profit lived, concentrated in exactly the bets where the two sources disagreed most.
Why the tackle boards are still up, without a pick on them
The projections are still on the site, because the inputs were never the problem. Playing time, opponent pace and run tendency really do predict tackles, and that part of the work held up fine.
What isn't there is a play. We show our projection next to the sportsbook's number and let you see the difference yourself. We don't tell you to bet it, because we can't grade it honestly yet — our scorekeeper and theirs disagree, and any record we posted would be measuring that disagreement instead of the bet.
Every market we screen now gets the blind baseline automatically, printed beside the model's own result, with anything landing within three points of it flagged on the spot. It used to be a test we remembered to run. Now it is the last line of the report, every time, whether or not anyone wants to see it.
What this doesn't say
It doesn't say tackle props are unbeatable. It says we can't prove anything about them while our data and the sportsbook's data disagree about what a tackle is. Graded against the official source, the same model might find something real. We'd have to run it to know.
It doesn't say the testing was bad. The testing is what caught this. Six checks passed and the seventh didn't, and a test that catches a false positive is a test doing its job. The uncomfortable part is only that we ran it seventh.
And the lesson isn't really about football. Any time your data has a consistent relationship to the data your bets are settled on, this will happen, and it will look like a discovery every time. Your numbers run low, so you always land on the under, and the under wins at whatever rate reality provides. A source that runs high produces the identical illusion on the over.
Two seasons and 752 bets was enough to find the signal and enough to find out what was sitting underneath it. We re-run this screen at the start of every season. If a better source for the official counts becomes available, we'll test it again and publish what it says — including if it says we were wrong to walk away.
Sources
Every number here comes from the graded ledger the boards publish. Research from a statistical model, not betting advice; no outcome is guaranteed. Method: how MatchWiz works.