DFS Correlation Deep Dive: The Math That Makes Stacking Work
Correlation is the mathematical structure that makes DFS stacking work. Every DFS sport has known position-pair correlation coefficients that reflect real-world scoring dependencies — QB touchdowns are WR touchdowns; batting- order-adjacent hitters share rally events; NHL linemates share goals and assists. Simulation that respects those coefficients prices stacks accurately; point projections that assume independence systematically underprice them.
Correlation gets treated as intuition in most DFS content — "stack your QB with his receiver because they score together." That's true but incomplete. Correlation is a specific mathematical quantity (Pearson coefficient between -1 and 1) with specific effects on lineup ceiling distributions. Understanding the math is what separates heuristic stacking from optimized stacking.
Correlation coefficients by sport
- NFL: QB-WR (same team) ≈ 0.55; QB-TE (same team) ≈ 0.45; QB-RB (same team) ≈ 0.15; QB-DST (same team) ≈ -0.3; QB-DST (opposing) ≈ +0.2.
- MLB: batting-order-adjacent hitters (positions 1-2, 2-3, etc.) ≈ 0.4-0.5; positions 3-5 (heart of the order) ≈ 0.5-0.6; pitcher vs opposing bats ≈ -0.7 (never roster).
- NHL: same-line skaters ≈ 0.5-0.55; same-team defenseman + top-line skater ≈ 0.35; opposing goalie vs stacked line ≈ -0.6.
- NBA: teammates ≈ 0.15-0.25 (weak); game-total participants (both teams' stars in a pace-up game) ≈ 0.3.
- Soccer: forward + attacking midfielder (same team) ≈ 0.45; full defense + goalkeeper (same team) ≈ 0.5-0.6 for clean-sheet outcomes.
- PGA / MMA / NASCAR: essentially zero between competitors. Field sports don't stack via correlation.
The joint-distribution math
For two players with means μ₁, μ₂ and standard deviations σ₁, σ₂ at correlation ρ, the joint distribution's 90th-percentile combined outcome exceeds the sum of independent 90th percentiles by approximately:
Ceiling boost ≈ (2ρσ₁σ₂)^(1/2) — the covariance term that independent-distribution math loses. At ρ = 0.55 with typical DFS standard deviations, that's a 15-20% boost. At ρ = 0.15 (weak NBA correlation), it's under 5%. This is why NFL and MLB stacks are more rewarding than NBA stacks.
Why independence-based projections fail
A point projection for a QB (23 points) and his WR (16 points) adds to 39. That's the sum of means, which is correct for the joint mean. But the sum-of-independent- 90th-percentiles vs the joint 90th percentile diverge — the joint 90th percentile is higher. Cash-game construction uses means (independence is fine); GPP construction uses ceilings (independence is systematically wrong).
This is why every serious DFS operator ships correlated Monte Carlo simulation. Independence-based tools produce cash-game-competitive lineups but structurally underprice GPP-winning stacks.
Negative correlation and bring-backs
Not all correlation is positive. A QB's fantasy points move INVERSELY with his opposing defense's fantasy points (roughly -0.3 correlation). A pitcher's points move inversely with the opposing team's bats (-0.7 correlation). Understanding negative correlation prevents "auto-fail" stacks — never roster a QB and his opposing defense in the same lineup.
Bring-backs exploit a specific correlation pattern: the stack's universe (a shootout game script) correlates positively with the OPPOSING team's attackers benefiting from the same script. The QB-WR stack's 30-point ceiling universe is (with high probability) a universe where the opposing WR also went off — bring-back captures that correlated upside.
Portfolio-level correlation
A 20-lineup portfolio has correlation structure across lineups, not just within them. Two lineups sharing 6 of 9 players are 66% correlated — a bad slate for one is a bad slate for both. Exposure caps enforce portfolio- correlation limits so no single game script determines more than a bounded fraction of your portfolio's outcome. See exposure caps explained for the caps math.
Related
- DFS stacking strategy — the general theory of stacking correlation drives
- Monte Carlo simulation explained — how the simulator samples the correlated distributions
- NFL DFS strategy — where correlation coefficients are highest
- MLB DFS stacking — batting-order correlation deep-dive