Key Takeaways
- Every player stat outcome follows a probability distribution, a range of values with different likelihoods
- Point predictions ("he'll score 22") ignore the uncertainty that determines whether a prop is worth betting
- Different stats follow different distribution shapes. Points behave differently from threes or blocks
- Understanding the distribution lets you calculate the actual probability of going over or under any line
- Models that output distributions are fundamentally more useful than those that output a single number
Beyond the Average
A player averages 22 points per game. Does that mean he'll score 22 tonight? Obviously not. He might score 14, or 31, or anywhere in between. The average tells you the center of the distribution, but nothing about the spread.
Here's why that matters for betting. Suppose two players both average 22 points:
- Player A scores between 18 and 26 almost every night. He's remarkably consistent.
- Player B swings between 10 and 38. Some nights he explodes, some nights he disappears.
If the line is 22.5, the probability of the Over is very different for these two players, even though their averages are identical. Player A might have a 45% chance of going Over 22.5, while Player B might have a 48% chance. The distribution shapes the bet, not the average.
What Is a Probability Distribution?
A probability distribution describes all possible outcomes of a random variable and how likely each outcome is. For player props, the random variable is a stat (points, rebounds, assists, etc.) and the distribution tells you the probability of every possible value.
Think of it as a histogram. If you could watch a player play 10,000 games under identical conditions, the resulting histogram of their stat totals would approximate their probability distribution.
Common Distribution Shapes for Player Stats
Different statistics tend to follow different distribution families:
Points are roughly normal (bell-shaped) for high-volume scorers, but can be right-skewed for low-usage players who occasionally have breakout games.
Rebounds tend toward a Poisson-like shape for guards (clustered at low values with a right tail) and more normal for centers and forwards.
Threes Made often follow a Poisson distribution, since three-point makes are discrete, relatively rare events. A player who averages 2.5 threes per game might hit 0, 1, 2, 3, 4, or 5+ on any given night with specific probabilities.
Assists are similar to threes in that they're count data, but the distribution depends heavily on role. A primary ball handler has a different assist distribution than a catch-and-shoot wing.
Blocks and Steals are low-count events that cluster near zero. These are typically best modeled with count distributions (Poisson or negative binomial) since most players record 0-3 per game.
Why Distribution Shape Matters for Betting
The shape of the distribution directly determines the probability on each side of a line. Here's a concrete example.
Scenario: Threes Made
A player averages 2.5 three-pointers per game. The sportsbook sets the line at 2.5.
If threes follow a Poisson distribution with mean 2.5, the exact probabilities are:
| Threes Made | Probability | |---|---| | 0 | 8.2% | | 1 | 20.5% | | 2 | 25.7% | | 3 | 21.4% | | 4 | 13.4% | | 5+ | 10.9% |
Over 2.5: 21.4% + 13.4% + 10.9% = 45.6% Under 2.5: 8.2% + 20.5% + 25.7% = 54.3%
The average is 2.5, but the Under is favored. This happens because the Poisson distribution is discrete and slightly right-skewed, so there's more probability mass below the mean than above it when the line sits exactly at the mean.
A bettor using only the average would think this is a coin flip. A bettor using the distribution knows the Under is the right side at standard juice.
The Minutes Connection
For most counting stats, minutes played is the dominant driver of variance. A player who plays 34 minutes has more opportunities to accumulate stats than one who plays 22 minutes due to early foul trouble or a blowout.
This creates a critical modeling insight: the distribution of a player's stat is conditional on their minutes distribution. A full model needs to:
- Project a minutes distribution (accounting for game script, foul risk, rotation depth, rest)
- Project a per-minute production rate
- Combine them to produce the final stat distribution
This two-stage approach naturally captures a major source of variance that single-stage models miss. If a game has high blowout risk, the minutes distribution gets wider, which widens the stat distribution, which changes the Over/Under probabilities.
Correlation Between Stats
Player stats aren't independent. Points, rebounds, and assists all share a common dependency on minutes. If a player plays more minutes, all three stats tend to increase. If they play fewer, all three decrease.
This matters enormously for combination props like PRA (points + rebounds + assists). You can't just add the individual distributions. You need to account for the correlation structure.
Why Naive Addition Fails
Suppose a player's projected distributions are:
- Points: mean 20, std dev 6
- Rebounds: mean 8, std dev 3
- Assists: mean 5, std dev 2.5
If you naively assume independence, PRA has mean 33 and std dev √(36 + 9 + 6.25) ≈ 7.2.
But if the stats are positively correlated (because they all depend on minutes), the true std dev is larger, maybe 8.5-9.0. This wider distribution shifts the probability on both sides of the PRA line.
Models that simulate stats jointly (drawing correlated random variables from a multivariate distribution) handle this correctly. Models that treat each stat independently will misprice combination props.
Practical Implications
Understanding distributions changes how you evaluate props:
Don't anchor on averages. The average is one number from a complex distribution. Two players with the same average can have very different prop probabilities.
Respect the tails. Low-probability, high-impact outcomes (a player's 40-point explosion game or 5-block night) are part of the distribution. They affect the probability calculation even though they're rare.
Watch for distribution shifts. When a key teammate is injured, the distribution doesn't just shift up or down. It can change shape. More usage might increase the mean but also increase variance.
Think in probabilities, not predictions. Instead of asking "will he go over?", ask "what's the probability of the Over, and does that exceed the implied probability from the odds?"
How Propboard Uses Distributions
Propboard's Monte Carlo engine simulates thousands of game outcomes for every player, producing full probability distributions for each prop market. Each prop gets a calibrated probability, not a point estimate, which is compared against sportsbook odds to identify where the market may be off.
This approach naturally handles minutes variance, stat correlations, and the different distribution families that apply to different stat types. Start your free trial to see the full probability breakdown for today's player props.
Related Reading
- What Makes a Good Player Prop Model? for how calibration and feature selection determine model quality
- What Are Player Props? for a primer on the markets where distributions matter most
Frequently Asked Questions
What's the difference between a distribution and an average?
An average is a single number summarizing the center of the data. A distribution is the full picture: every possible outcome and how likely each one is. Two players can have the same average but very different distributions, which means different probabilities on the same prop line.
Why do different stats follow different distributions?
It comes down to the nature of the stat. Points are accumulated gradually across a game and tend toward a bell-shaped (normal) distribution. Threes, steals, and blocks are discrete, relatively rare events that cluster near zero, making them better described by count distributions like Poisson. The right distribution family depends on how the stat is generated.
How does Monte Carlo simulation use distributions?
Monte Carlo simulation draws random samples from estimated distributions (minutes, per-minute rates) thousands of times, producing a simulated histogram of outcomes. Instead of saying "this player will score 22," it says "here's the probability of every possible scoring total." This naturally captures uncertainty and lets you calculate precise Over/Under probabilities.
Do I need to understand distribution math to benefit from distribution-based models?
No. You benefit from distribution-based models by using their outputs (calibrated probabilities and edge estimates) rather than doing the math yourself. Knowing the concepts helps you understand why some bets look good despite average projections being close to the line, but the model does the heavy lifting.
For entertainment purposes only. Must be 21+. Please gamble responsibly.