The Correlated Parlay Math: Correlation, Expected Value, and When the Book Gets Pricing Wrong

At FanDuel's 2025 investor day, the company disclosed that parlays now account for 70 percent of its NFL and NBA betting handle. FanDuel introduced same-game parlays in 2019. In six years, the product went from a novelty to the majority of the book's volume on its two biggest sports.

New Jersey's Division of Gaming Enforcement data shows why the book wants this migration: the state's parlay hold runs 19 to 22 percent, compared to 5 to 6 percent for straight bets. Every dollar that moves from spreads and moneylines to SGPs is worth three to four times as much to the house.

The engine behind that gap is correlation math. Standard parlays assume each leg is independent of every other. When legs come from the same game, that assumption fails. The question is not whether correlation changes the price. The book accounts for the correlation and adjusts the payout accordingly. The question is whether the book's correlation model is accurate, and whether the gap between the book's model and reality ever runs in your favor.

The Independence Assumption

Standard parlay math is straightforward. For two independent events with probabilities p_A and p_B, the joint probability is p_A times p_B. Two -110 bets each carry an implied probability of 52.38 percent after removing the vig from a balanced -110/-110 market. The joint probability is 0.5238 × 0.5238 = 27.44 percent. The fair no-vig price for that parlay is 1 ÷ 0.2744 = +265.

Books pay roughly +260 on two-leg -110/-110 parlays. The house edge is small because the book is capturing only the vig already baked into each -110 price. The independence assumption is reasonable for cross-game bets, where a Chiefs win has no bearing on whether the Dodgers cover their spread.

Same-game legs are not independent. A quarterback throwing for 310 yards and his top receiver going over 90 receiving yards move together. A team winning outright and the game going over a total are linked. The moment you combine these outcomes, multiplication no longer gives the right answer.

The Joint Probability Formula

For two correlated binary outcomes, the correct joint probability is:

Here ρ (rho) is the Pearson correlation coefficient between outcomes A and B, ranging from −1 to +1. When ρ equals zero, the formula collapses back to standard multiplication. When ρ is positive, the joint probability rises above the product of the individual probabilities. When ρ is negative, the joint probability falls below the product.

The standard deviation of a binary outcome with probability p is √[p × (1 − p)]. For a 50-percent outcome, that is 0.50. For a 60-percent outcome, that is √(0.60 × 0.40) = 0.4899.

The table below shows how rho changes the fair price of a two-leg SGP built from two even-money legs (p = 50% each).

A 0.30 correlation between two 50-percent legs cuts the fair parlay odds from +300 to +208, a 31 percent reduction in payout. A 0.50 correlation cuts them to +167, a 44 percent reduction. The book is not paying +260 on your QB-over and receiver-over combo. It is paying +140, because the book sees those legs as 40-plus-percent correlated.

The correlation premium is not a scam. It is a correction for a real mathematical fact: correlated events hit together more often than multiplication predicts, so the book faces higher risk and reduces the payout to compensate. The question is whether the book's estimate of rho is accurate.

The Gaussian Copula

The Pearson correlation formula above works as a first approximation for Bernoulli outcomes when both probabilities are near 50 percent. Sportsbooks use a more rigorous method: the Gaussian copula, originally developed by David X. Li in a 2000 paper for credit-risk modeling at J.P. Morgan. Li's paper ("On Default Correlation: A Copula Function Approach," Journal of Fixed Income, Vol. 9, No. 4) showed how to combine individual event probabilities into a joint model using a multivariate normal distribution structure.

US Patent 12080130, issued to SidePrize LLC in 2024, describes the direct application to sports betting. The patent covers a platform for same-game prop bets using a multivariate normal cumulative distribution function with a pairwise correlation matrix to price correlated legs. Each pair of proposition outcomes receives a correlation value between −1 and +1. The pricing engine integrates over the joint distribution to produce the combined probability.

The practical difference between the linear formula above and the full Gaussian copula matters at the extremes: when both legs are strong favorites (implied probabilities above 70 percent) or when correlations are high (above 0.60). For the typical two- or three-leg NFL SGP built from props near the −110 to −130 range, the gap between the two methods is small enough that the linear formula gives an accurate approximation.

What matters practically is this: every sportsbook runs a proprietary version of this model with its own correlation estimates. Those estimates differ. DraftKings and FanDuel do not share correlation matrices. That disagreement is the only structural source of cross-book price differences on SGPs, and the reason shopping matters more on SGPs than on any other bet type.

Worked Example: The QB-Stack

Take a two-leg SGP: a team wins the moneyline (no-vig probability 60 percent) and the quarterback goes over 275 passing yards (no-vig probability 50 percent).

Both books are running the same formula on the same two legs. Book A estimates the correlation at 0.20. Book B estimates it at 0.40. The result is a 32-point difference in payout (+158 vs. +126) for the same SGP. If the true historical correlation between this team's wins and this QB's passing yards is closer to 0.30, Book A is underestimating the correlation and posting a relatively better price.

Now consider what happens if Book A underestimates correlation significantly. Suppose true ρ is 0.40 and Book A prices as if ρ = 0.20, offering +158 on the SGP:

A genuine positive expected value. A 20-point rho mismatch on a two-leg SGP, from a book that consistently underestimates correlation on this class of combination, produces a +2.7 percent edge. This is not common. Books update their correlation matrices continuously. The theoretical condition exists, and cross-book price divergence is the signal that it is happening somewhere on any given week's board.

The Hold Rate Data

New Jersey's Division of Gaming Enforcement publishes monthly data broken down by bet type. In July 2024, the combined parlay hold across New Jersey sportsbooks was 22.3 percent on $182.1 million in parlay handle. Year-to-date through mid-2024, the state's parlay hold ran 19.5 percent. Single-event football bets held 6.3 percent over the same period. Baseball moneylines and run lines held 5.6 percent.

The gap between parlay and straight-bet hold is not explained by vig compounding alone. Two independent -110 legs produce a house edge of about 4.5 percent. SGPs produce 15 to 22 percent. The difference is the correlation premium, the additional margin the book charges for modeling and bearing the risk of correlated outcomes.

Niusha Moshrefi's 2026 preprint at Princeton (arXiv 2607.14430, "Prices, Probabilities, and Parlays: Systematic Bias in Sports Prediction Markets") analyzed 23 million moneyline trades on Kalshi and found that cross-game parlay overpricing grows geometrically in leg count. Each additional leg adds a compounding markup beyond what the individual leg prices imply. Moshrefi concluded that the overpricing is not inherited from calibration errors on individual legs but imposed as a separate market-level fee at the parlay-pricing stage.

DraftKings' 2025 annual SEC filing reported that parlay handle mix increased 4.3 percent year-over-year, and the company cited parlay growth as the primary driver of structural hold expansion to 10.9 percent, up from 9.7 percent in the prior year. The hold rate improvement translated directly into sportsbook net revenue margin of 8.7 percent, a company record.

Where the Book Underprices Correlation

Sportsbooks fit correlation estimates on historical data. Sample sizes are limited. NFL teams play 17 regular-season games, producing roughly 340 player-game pairings per season. Estimating the correlation between a specific quarterback's passing yards and a specific receiver's reception yards from two seasons of data means working with fewer than 34 observations. The standard error on that estimate is large.

The same SGP construction at FanDuel and DraftKings regularly differs by 20 to 40 percent in payout. That gap reflects genuine disagreement between two separate correlation models built on the same underlying data. One book is wrong. The bettor who shops both prices is always at the book with the lower correlation estimate for that particular combination, which means the book that believes the legs are less connected.

The specific SGP classes most likely to show cross-book divergence:

• Props combining a player with their team's situational outcome (e.g., receiver yards over combined with team winning by fewer than 10 points, where the game-script interaction is less obvious than a pure win/loss correlation)
• Two players from opposite teams whose outputs are inversely linked through defensive matchup (e.g., a pass-rushing edge defender's sacks over and the opposing quarterback's completion percentage over, a combination with negative correlation that books model inconsistently)
• Team total props combined with individual player props from the second half only, where the game-script correlation differs meaningfully from the full-game version

The core principle: books model the obvious positive correlations (QB yards + receiver yards + game total over) accurately because those combinations are heavily bet and the correlation signal is easy to estimate. The subtle combinations, those with less volume and less obvious directional correlation, are where the model diverges across books.

Negative Correlation: The Overlooked Side

Most SGP analysis focuses on positive correlation because the popular stacks move in the same direction. Negatively correlated outcomes are less common but worth understanding because the math runs in reverse.

Example: a team wins by a blowout (60 percent probability) and the opposing quarterback goes over 250 passing yards (50 percent probability). A dominant defensive performance and large opposing QB volume are negatively correlated. When a team routes an opponent, the opposing quarterback often gets benched or faces a garbage-time situation that does not produce the clean volume the prop requires.

If a book applies a positive or zero correlation adjustment to this combination (treating the blowout win and opposing QB's volume as unrelated), the book underestimates the joint difficulty. The SGP is harder to hit than the book's model says. This does not produce positive expected value. The book still applies vig. The bettor is paying for a bet that is less likely to win than both parties think. Avoid these combinations.

Conversely, a book that correctly identifies the negative correlation and prices the SGP at +270 or +290 (instead of the independence-implied +233) is being transparent about the difficulty. The bet is still negative expected value after vig, but the stated odds more accurately reflect the true probability.

The Practical Decisions

The math produces a clear framework:

• Identify the correlation structure before placing. The most popular positive stacks (QB + receiver + game total over from the same team) are the most accurately priced by every major book. The correlation premium on these combinations is already at or above the true rho. The book does not miss these.
• Search for less obvious combinations. Situational correlations (game script, weather impact on both sides, defensive matchup-specific props) are modeled with more estimation error because the historical sample for each specific combination is smaller.
• Shop the same SGP across two or three books before placing. Cross-book payout differences of 20 to 40 percent exist for identical constructions. Choosing the highest payout does not change the house edge to zero, but it materially reduces the size of the edge the book takes.
• Keep leg count low. Each additional leg compounds the correlation premium. A two-leg SGP with a 15 percent hold becomes a four-leg SGP with a 30 to 40 percent hold. At four or more legs, estimation error in the book's correlation model is no longer large enough to offset the geometric compounding of the vig layer.
• Avoid negatively correlated combinations. The joint probability is lower than standard multiplication predicts, and the book does not always price this correctly.

The Gaussian copula pricing model exists because the independence assumption is wrong for same-game bets. Every book knows this. The correlation premium is the book's answer to the risk that correlated outcomes hit together more often than multiplication predicts. The cases where the book's model is materially too low do exist. The cross-book price divergence of 20 to 40 percent proves at least one book is wrong on any given combination, but they are rare and the underlying expected value is still negative on most SGP constructions.

The math does not produce a path to positive expected value on SGPs as a category. What it produces is a ranking: some constructions drain bettors faster than others, and some books drain faster than others. The formula tells you which combinations sit at the low end of that range, and which book is offering them at the lowest correlation premium on any given week.