The NFL prop correlation index

Everyone selling same-game parlays tells you the legs correlate. Almost nobody tells you by how much, on what sample, or where the idea turns out to be wrong. This is our attempt at all three, measured on 22,256 player-weeks across 1,087 games and 4 completed seasons.

The number in every table below is how often both legs landed, divided by how often they would land if the two were unrelated. Above 1.00 means they arrive together more than independence predicts, which is what a same-game parlay hopes for. Below 1.00 means the opposite. A parlay priced as though the legs were independent is then charging you too much.

Two of the most repeated beliefs in this category do not survive it.

First, proof the method is not making it up

A correlation study is only worth reading if it can show its own method returning nothing when nothing is there. So before measuring pairs inside a game, we measured the same pair types across different games in the same week. Those must come out at 1.00. Players in unrelated games cannot affect each other.

All three land within a couple of points of 1.00, and every range covers it. So the method does not manufacture correlation out of nothing. That is what lets you read the rest of this page as signal rather than as an artefact of how we counted.

What actually moves together

Every row was also measured season by season. A pair is only called stable when it kept the same direction in all 4 seasons and its range excludes 1.00. Ranges come from 2,000 bootstrap draws.

The quarterback stack is the whole story

A quarterback going over his passing yards and his WR1 going over his receiving yards happen together 1.352 times as often as independence predicts. Across 1,683 paired games, stable in every season. Pair the passing touchdowns with that receiver scoring instead and it is 1.559, the largest stable number in the study.

The whole receiving tree comes along, in the order you would guess: WR1 1.352, WR1 receptions 1.286, TE1 1.242, WR2 1.229. A quarterback having a big day is a rising tide for everyone he throws to, and it does not matter much which one you pick.

The mirror image is the strongest negative we found: two opposing running backs both going over is 0.855, under 1.00 in all four seasons and falling to 0.795 in the most recent one. One team running the ball takes carries away from the other. That parlay is structurally overpriced when it is priced as though the legs were unrelated.

Three things people believe that we could not find

These are the results we would have left out if the point were selling parlays.

Two receivers on the same team do not cancel each other out

The intuition is airtight: target share is a fixed pie, so WR1 and WR2 must compete. Measured across 1,716 paired games it is 0.997, with the range covering 1.00 exactly. How much a team throws varies enough between games to cancel the competition almost perfectly. Pairing two receivers without their quarterback is priced fairly, and there is no edge in either direction.

Two touchdown scorers on the same team do not help each other

RB1 and WR1 both scoring: 0.994, null in all four seasons. More team touchdowns lifts both players, and the two are also rivals for the same goal-line plays. The effects cancel. The correlated touchdown pair is the quarterback’s passing TD with a receiver’s touchdown, which is 1.522, and not scorer with scorer. Any tool implying that two scorers from one offence are correlated is telling you something this data does not support.

Shootouts do not do what the story says

Quarterback against opposing quarterback is the most marketed same-game story there is. In our data it flips sign between seasons, below 1.00 in two and above in two, which makes it unusable no matter what the pooled figure says.

The related belief fails more cleanly. Conditioning on a high game total makes correlation weaker, not stronger. Every key positive pair has its lower number in high-total games:

The reason is arithmetic rather than football. In a high-total game both legs are more likely on their own, so the shared cause has less left to add on top. Stack a shootout because the individual legs are likelier, not because correlation is stronger there. It is not.

What this cannot tell you

The biggest limit is the line. No archive of historical sportsbook prop lines exists for us to measure against, so every player is judged against his own median for that season instead. That puts both legs near a coin flip by construction, which is a clean place to measure co-movement and is not where a sportsbook sets its number.

• Do not read these as the multiplier at a book’s number. Books price near the mean with their margin on top, and alternate lines sit further out. Co-movement in the tails is usually stronger than at the median for positively linked legs, so the stack numbers here are probably conservative for alternate-line overs.
• These are not live probabilities. The season median is computed over the whole season, including the game being scored, which no one could have known at kickoff. That is fine for measuring how two things move together and useless as a prediction.
• Mechanism is inferred, not observed. Weekly totals carry no drive or situation data, so explanations here reason from final margins and pregame markets rather than from the plays themselves.
• Opposing pairs count each game twice, once from each side. That leaves their sample size honest but their ranges slightly narrower than they should be.

How to use it

• Building a same-game NFL parlay? The quarterback with one of his own pass catchers is the pairing this data supports, and it barely matters which one.
• Two opposing running backs both going over is the pairing to avoid.
• A pairing measured at 1.00 is not a bad bet. It means the parlay price is fair on that count, and the legs have to stand on their own merits.
• Every number here is a median-line measurement on past seasons, not a price. Treat it as the direction and rough size of an effect, not as a multiplier.

Questions

Nothing here is betting advice, and correlation does not make a parlay a good bet. Bet only what you can afford to lose. If it stops being fun, take a break.