How Our NBA Player Projections Work

Most player projections give you one number: "25.3 points tonight". That number says nothing about how likely 18 or 32 would be, or how sure the model is.

We publish a full probability distribution instead: the chance of every possible outcome. We do it for twelve stats and for every player on the projected roster of each team playing tonight, about 19 players a side, not just the starting five.

Twelve stats, each with its own distribution

The four headline stats are points, rebounds, assists and three-pointers made. Behind them come steals, blocks, turnovers, field goals made and attempted, free throws made and attempted, and the NBA Salary Cap fantasy score.

Five of the twelve (points, rebounds, field goals made, field goals attempted and Salary Cap points) passed our accuracy test for a single number, so we show them as an expected value. The other seven are shown as a range from the 10th to the 90th percentile, because for those stats the shape of the distribution is more reliable than its average.

What goes into the models

Each stat has its own model, trained on years of NBA box scores. The inputs are:

• Recent form: the player's averages for the season and for his last five games
• Role: position, and whether he starts or comes off the bench
• Game context: home or away, and how tired each team is from its recent schedule

Reading the PDF chart

The PDF (probability distribution) chart is a bar chart. Each bar is the chance of one exact result, for example exactly 22 points. The tallest bar is the single most likely result. The dashed line marked E[X] is the expected value, the average of all results weighted by their chance.

Look at the width. A narrow chart means the outcome is fairly predictable. A wide one means the player could land almost anywhere across that range.

For a longer walkthrough, see our guide to probability distributions.

Reading the CDF chart

The CDF chart on this page answers one question: what is the chance the player goes over X? The line starts near 100% on the left and falls toward 0% on the right. Where it falls steeply, most of the probability sits in that range.

Example: find 22.5 on the horizontal axis. If the line reads 58%, the model gives a 58% chance of 23 points or more and 42% of 22 or fewer. Hover any point and the tooltip shows both numbers.

The week ahead

Tonight's table answers one question. Fantasy managers usually have another: who is worth adding for the next few days. The page answers it in its own section, and those numbers are made differently.

We run the same models on every game a player has scheduled in the window and add the results. No lineup is published that far ahead, so his role in each game comes from his own recent record: how often he has started and how many minutes he has played.

A sum of expected values is still an exact expected value, so the weekly number is correct as an average. It is published as an average only, with no range and no curve, because the spread of a week of games cannot be drawn from the spread of one game. The per-game column next to it divides the total by the games, which separates a busy schedule from a strong week.

What this ranking does not claim: in a replay of the 2025-26 season with every input fixed at the start of each week, choosing additions by this ranking did no better than choosing them by recent form. An earlier version of the study showed a gain of about 15 Salary Cap points a week, but it used information that only arrived during the week, so we do not quote it as an advantage.

We also do not publish a chance that a player will play. Availability on the page is the starting five the team has published that day, and nothing more.

Why distributions beat single numbers

Two players can both average 23.5 expected points. One lands between 20 and 27 most nights. The other ranges from 12 to 38. Same average, very different players to rely on. Only the distribution shows the difference.