Variance-to-Mean Ratio (VMR) Explained: When Poisson Fails | DMP Learn

Understanding Variance-to-Mean Ratio (VMR)

What You'll Learn

The Variance-to-Mean Ratio (VMR) is a diagnostic tool that tells you whether a count stat is well-behaved (Poisson-friendly) or overdispersed (boom-or-bust). It's the single most important check before trusting Poisson probabilities for any discrete count prop. This lesson teaches you how to calculate VMR, interpret it, and use it to avoid mispricing traps.

---

Why VMR Matters

In the Poisson lesson, you learned that Poisson assumes the variance equals the mean. That's a strong assumption, and when it fails, your probability estimates can be significantly off.

Consider two tight ends who both average 0.5 touchdowns per game:

Tight End A scores a TD in about 40% of games, almost always exactly 1. His game log looks like: 0, 1, 0, 0, 1, 0, 1, 0, 0, 1. The variance is close to the mean. Poisson works perfectly.

Tight End B has three 2-TD games and one 3-TD game in a 16-game season, but scores zero in the other 12. His game log looks like: 0, 0, 2, 0, 0, 0, 0, 3, 0, 0, 2, 0, 0, 0, 0, 0. Same average, but the variance is much higher than the mean. Poisson would overestimate his anytime TD probability.

VMR catches this difference.

---

How to Calculate VMR

Step 1: Collect the game log

Pull the player's last 12-16 games for the stat in question. Remove games where the player exited early due to injury (partial games distort the data).

Example — Running back touchdown log (last 14 games):

0, 1, 0, 0, 2, 0, 0, 1, 0, 3, 0, 0, 0, 1

Step 2: Calculate the mean

Mean = Sum of values / Number of games
Mean = (0+1+0+0+2+0+0+1+0+3+0+0+0+1) / 14
Mean = 8 / 14 = 0.571

Step 3: Calculate the variance

Variance measures how spread out the values are. Use the sample variance formula (divide by n-1):

Variance = Σ(xi - mean)² / (n - 1)

For our example:

Variance = [(0-0.571)² + (1-0.571)² + (0-0.571)² + ... + (1-0.571)²] / 13
Variance = [0.326 + 0.184 + 0.326 + 0.326 + 2.041 + 0.326 + 0.326 + 0.184 + 0.326 + 5.898 + 0.326 + 0.326 + 0.326 + 0.184] / 13
Variance = 11.429 / 13 = 0.879

Step 4: Divide

VMR = Variance / Mean = 0.879 / 0.571 = 1.54

Spreadsheet shortcut

In Excel or Google Sheets, if your game log is in cells A1:A14:

=VAR(A1:A14) / AVERAGE(A1:A14)

That's it. One formula.

---

Interpreting VMR

---

What Overdispersion Does to Your Estimates

When VMR > 1.3, Poisson systematically gets two things wrong:

1. It underestimates P(zero). Boom-or-bust players have more zero games than Poisson predicts. Poisson assumes a smooth probability curve; the reality is clumped — lots of zeros and occasional big games.

2. It underestimates P(high counts). Poisson puts too little probability in the right tail (2+ TDs, 3+ home runs). The excess variance comes from these multi-event games.

3. It overestimates the middle. To compensate, Poisson puts too much probability on exactly 1 — the "expected" outcome. But for overdispersed players, exactly 1 happens less often than Poisson predicts.

Practical impact on anytime TD props:

A player with λ = 0.7:

That 4.5-point gap on "anytime TD scorer" is the difference between a +EV bet and a -EV bet at typical market prices.

---

Real-World Overdispersion Patterns

High overdispersion (VMR 1.5+):

• Touchdown-dependent running backs on committees — they either get the goal-line carry or they don't, and when they do they sometimes punch in multiples
• Streaky home run hitters — power output varies enormously with pitcher matchup and ballpark
• Boom-or-bust wide receivers who either get a long TD or barely touch the ball

Low overdispersion (VMR near 1.0):

• Workhorse running backs with consistent red zone usage — steady TD production game to game
• Elite pitchers with stable strikeout rates — their K output varies but in a predictable range
• High-usage NHL forwards with consistent shot generation

The pattern: Overdispersion tends to correlate with role volatility. If a player's opportunity varies wildly game to game (snap count, target share, red zone usage), their counting stats will be overdispersed. If their role is locked in, the count stat will follow Poisson more cleanly.

---

Using Negative Binomial for Overdispersed Players

When VMR > 1.3, the Negative Binomial distribution is the better model. It uses two parameters:

• Mean (μ): Same as Poisson's lambda
• Size parameter (r): Controls the dispersion. Smaller r = more overdispersion. As r → ∞, Negative Binomial converges to Poisson.

Estimating r from your data:

r = mean² / (variance - mean)

Using our running back example (mean = 0.571, variance = 0.879):

r = 0.571² / (0.879 - 0.571) = 0.326 / 0.308 = 1.059

Calculating probabilities:

In Excel:

=NEGBINOM.DIST(k, r, r/(r+mean), TRUE)

For P(0 TDs):

=NEGBINOM.DIST(0, 1.059, 1.059/(1.059+0.571), FALSE) = 54.8%

P(Anytime TD) = 1 - 54.8% = 45.2%

Compare to Poisson's estimate of 50.3% — a significant difference that affects bet selection.

---

When to Skip VMR

VMR is only meaningful for discrete count stats. Don't calculate it for:

• Passing yards, points, rebounds, assists (use CV instead for these continuous stats)
• Combo props (PRA, P+A) — these are sums that approximate Normal
• Props with very small samples (fewer than 10 games makes VMR unreliable)

Also, VMR can fluctuate with sample size. A player with VMR = 1.4 over 14 games might be 1.1 over 30 games. Use the most relevant recent sample, but don't over-interpret borderline cases (VMR 1.2-1.4). In the borderline zone, run both Poisson and Negative Binomial — if they agree on the direction of the bet, you're fine. If they disagree, the edge probably isn't real.

---

How DMP Handles This

DMP's prop analysis evaluates each market's variance characteristics as part of its scoring pipeline. Markets where the underlying stat has clean distributional properties (VMR near 1.0 for count stats, moderate CV for continuous stats) score higher on modelability. This is part of why you'll see certain prop types flagged as +EV on DMP more consistently — they're simply easier to model reliably.

When DMP's fair probability differs significantly from the sportsbook's implied probability, it's partly because the platform accounts for these variance dynamics rather than using a simplified one-size-fits-all calculation.

---

Key Takeaways

• VMR = Variance / Mean, calculated from a player's game log. One spreadsheet formula.
• VMR near 1.0 → Poisson works. VMR > 1.3 → player is overdispersed, switch to Negative Binomial.
• Overdispersion means more zeros and more blowup games than Poisson predicts, with fewer "average" games.
• The practical impact is significant: 4-5 percentage points on anytime scorer props, enough to flip EV direction.
• Role volatility drives overdispersion — inconsistent usage creates boom-or-bust count distributions.
• When VMR is borderline (1.2-1.4), check both distributions. If they disagree on the bet, the edge isn't reliable.

---

Next lesson: How to Evaluate Any Prop Market →