Bringing Advanced Stats to the WNBA: How to Fix Plus-Minus

Let’s take regularized adjusted plus-minus (RAPM), a commonly used NBA and NHL lineup-based player impact metric, and adapt for the WNBA.

It’s easy to quantify A’ja Wilson’s impact on the court. But what about the rest of the leauge? | Kiyoshi Mio-Imagn Images

Quick. Try to name the most impactful players from the 2025 WNBA season. How to best quantify that is up to you. 

A'ja Wilson, surely. Alyssa Thomas, Napheesa Collier, Allisha Gray and Breanna Stewart also fit the bill. Now, try some math. Every lineup, every possession and its scoring margin fed as inputs. Solve for which players have the greatest impact on their team’s success (or failure), and out come individual ratings.

Wilson, of course, is still No. 1. But the results get a little weird. Naz Hillmon, last season’s Sixth Player of the Year, ranks second. Leïla Lacan, fifth. And Thomas … seventh? … followed by Veronica Burton, 2025’s Most Improved Player, eighth. The number crunching produces something interesting, at the very least. 

Data compiled by Dan Falkenheim

That framework for evaluating player impact isn’t new: Regularized Adjusted Plus-Minus (RAPM) was first introduced by Joseph Sill at the 2010 MIT Sloan Sports Analytics Conference, and it has lived on as the foundation for all-in-one, lineup-based metrics in the NBA (LEBRON and EPM) and NHL (WAR). It is purpose-built for rating players in ways that traditional box score stats can’t.

For all the advancements in NBA analytics, though, public WNBA plus-minus work remained thinner and less established. That has started to change. Positive Residual developed Estimated Contributions in 2020, and Help the Helper, a WNBA stats app, published wRAPM just last week. But two questions remain: Can a WNBA-tailored RAPM model provide similar predictive power compared to its NBA counterpart, and, if so, how exactly should it be built?

To answer that, I pulled more than 900,000 possessions across all 29 completed WNBA seasons. I tested more than 500 unique configurations to see which frameworks worked best over the league’s entire history. The goal was to find which version of the framework worked best for the WNBA, rather than assuming the NBA template would transfer cleanly. 

The short answer is that RAPM can be adapted for the WNBA, but it needs some tuning and modifications. Before we get to the results, let’s walk through how it works.

Why plus-minus is misleading, and how to fix it

In its basic form, plus-minus calculates the team’s score differential when a player is on the court or the ice. (For example, if a basketball player’s team scored 30 points when they were on the court and the opponent scored 22 points when they were on the court, then that player receives a plus-minus of +8.) By computing the stat in this way, plus-minus attempts to quantify how a player positively or negatively affects their team’s performance. 

But basic plus-minus is a fundamentally flawed statistic. Here’s a real example: On July 27, the Aces beat the Wings 106–80. Wilson was a +25; Las Vegas bench player Aaliyah Nye was a +20. Plus-minus would suggest both players had a similar impact on the game. Plus-minus would be wrong. Wilson had 14 points, 10 rebounds, seven assists, four blocks and two steals, largely against Dallas’s starters. Nye had three points and three rebounds, largely against the Wings’ bench. Plus-minus is blind to the teammates a player shares the floor with and the quality of opponents faced, which are both key factors for assessing a player’s impact. 

So, how can plus-minus be improved? Turn back the clock two decades. Mathematicians Wayne Winston and Jeff Sagarin, in private, developed WINVAL, a plus-minus metric for the NBA that accounted for every player on the floor. While Winston and Sagarin dropped sparing hints for how their statistic was created, Dan Rosenbaum publicly formulated adjusted plus-minus (APM) and published his findings in April 2004.

In simple terms, APM essentially asks and answers the following question: Given who is on the floor, what are the player ratings that best predict what the scoring margin actually was?

APM, like raw plus-minus, is also not perfect. While it does know who is and isn’t on the floor, it may not know how to divvy up credit between two players who often appear on the court together. At the time, it also wasn’t clear just how predictive APM actually was. The stat looked useful but noisy.

Joseph Sill, then an analytics consultant, made another improvement on top of APM in 2010. Two, actually. First, instead of letting player ratings vary wildly, he implemented a technique called regularization that shrinks player ratings toward zero. Remember the bell curve? The idea is that player talent clusters around the middle, so a rating shouldn’t stray too far from the mean unless there is data to prove that it should. Second, he implemented a validation and testing process, which assessed just how well the statistic performed when predicting results it had never seen before. Without validation and testing, it’s hard to trust that a stat like APM is worthwhile. 

His method became known as regularized adjusted plus-minus (RAPM). He showed that RAPM improved predictive accuracy compared to APM, and he won the prize for best non-academic paper at the 2010 Sloan Sports Analytics Conference. 

Adapting RAPM for the WNBA

The NBA and WNBA are not the same. It’s a simple point, but it matters. The WNBA has about a third as many players and plays a fourth as many games. The league has gone through waves of expansion and contraction, takes mid-season Olympic breaks, juggles availability around international commitments and, of course, plays 40-minute games instead of 48. Whatever works for the NBA can't be assumed to work here.

It’s not obvious, though, what modifications are needed. Even on the NBA side, Sill contended with how to make the appropriate modeling choices. “Another challenge regarding APM surrounds various implementation details where choices have to be made,” he wrote in 2010. “Apparently somewhat arbitrarily.”

Consider former Lynx guard Yvonne Anderson. She finished first in net plus-minus per 100 possessions in 2025. She also played only 12 minutes. No one would call her one of the league’s most impactful players. Because RAPM only sees the players on the court and what the scoring margin was, though, it might very well say that she at least belongs in the league’s upper half. A choice has to be made. Either drop low-minutes players from the model (Rosenbaum and Sill’s solution) or pool them into a single shared rating, so that Anderson’s 12 minutes get blended into one generic “low-minute” player.

That’s one choice. There are more. Exactly what should that low-minutes threshold be? (The model needs to balance predictive accuracy while also including enough players.) How many seasons should the model see—one, three, five?—before forecasting future results? (While a player may have a proven track record, including stale data from many seasons ago might blur the reality of how they are playing today.) How hard should those ratings hew to the mean? (Gently, to preserve hierarchy; aggressively, to reduce noise; or somewhere in between.) Should each piece of evidence be a single possession or a stretch of possessions (stints) where the same 10 players stayed on the floor? (It may be valuable to increase the amount of evidence seen by using possessions.) These are the most basic implementation details, and they create hundreds of potential model configurations.

To find out which combinations actually work—the ones that produce the best predictive accuracy—I pulled play-by-play data for every WNBA season from 1997 to 2025. Then, more than 500 different RAPM specifications train on the data, produce results that are validated, train again, and then the most promising models provide predictions that are tested on held-out games.

As it turns out, the best-performing configuration included three years of data, possessions instead of stints, low-minute players pooled instead of dropped and moderate shrinkage. (Lambda equalling 3,500 for those keeping track at home.) Although Sill’s NBA results are not directly comparable because NBA and WNBA games differ in length and the number of possessions, the best WNBA three-year RAPM model achieved a similar level of predictive performance. That is, despite the smaller sample size, the WNBA-tuned RAPM model holds up.

The best one-year specification was about 2.5% worse than the top three-year model. What it sacrifices in predictive accuracy, it may gain in reflecting the quirks of an individual season. It landed on similar implementation choices, except it dropped low-minute players and preferred a lower minutes cutoff.

The chart below shows three-year RAPM ratings for 2023 to 2025. Let’s talk about what those numbers do and do not mean.

Data compiled by Dan Falkenheim

What RAPM means

What RAPM gains in usefulness, it loses in interpretability. Basic and (most) advanced box score stats are easy to understand. Points are points. Turnover percentage is the percentage of a player’s possessions that end in a turnover. Plus-minus is a team’s point differential when a player is on the floor. Simple enough.

RAPM is … a player's estimated effect, shrunk toward zero, on their team’s scoring margin per 100 possessions, adjusted for the teammates and opponents on the floor with them. (What?) Want to complicate it further? There is uncertainty around each player’s estimate, meaning that the one-year 2025 RAPM model was 95% confident that Wilson was at least the 28th-best player. That’s a clear way to see what RAPM isn’t: It isn’t always intuitive. 

Case in point: In the three-year, most accurate RAPM model, Kelsey Plum and Bri Jones are rated as the two most impactful players in the league from 2023 to 2025. Yet, neither has made an All-WNBA first team in that span. RAPM isn’t ground truth for who the best players in the WNBA are. RAPM also can’t be held in the same way that points, rebounds and assists can be seen as inarguable facts. 

None of that means RAPM isn’t valuable. Go back to the first part of the definition, the part that says RAPM is a player’s estimated effect on their team’s scoring margin. Or, in other words, winning. That means that RAPM may capture a different kind of signal that other box score metrics can’t provide.

Which is why a player like Leonie Fiebich might’ve ranked first in one-year RAPM in 2024. She was not the most impactful player that season, and the ranking borders on being silly. But, she did provide something that’s hard to quantify yet was clearly evident. There’s a reason—one that went beyond her modest stat line—why she was the runner-up for Sixth Player of the Year, and why she was a key piece that helped the Liberty win its first WNBA title 

Or take Leïla Lacan. Among guards in 2025, she ranked 23rd in the WNBA’s player impact estimate stat and second in one-year RAPM. Which is the better rating? RAPM overshoots the mark, but it does say more about why Connecticut coach Rachid Meziane feels the way he does about his second-year player:

"I just texted Leila (Lacan) and said we need you," Rachid Meziane jokes about the Sun's current point guard situation.

Lacan is expected to join the team within the next week or two after her overseas season in France concludes.

— Emily Adams (@eaadams6) May 10, 2026

RAPM is only the start. NBA analysts have spawned different variations, ones that include box score statistics, different offensive and defensive ratings and more informative prior information. It’s about time that the same is built for the WNBA.

---