A “kill” occurs when the defense prevents the offense from scoring on three straight possessions. College programs have tracked kills for well over a decade, so let's do the same for the WNBA. Here are the league's leaders in kills per 100 defensive possessions:

My WPA includes misses, makes, free throws, assists, rebounds, turnovers, steals, blocks, fouls drawn and committed. Credit for assists and rebounds can be ambiguous. My stab at it is detailed below.

The hand-wavy, imprecise explanation is that a semi-markov model estimates the home team's win probability by averaging a team's odds of winning across future game states. tried to go more in-depth with the simplified sheet below. (forgive any errors/typos!)

After testing XGBoost, GAMs, Stern’s Brownian motion model, and other methods, I landed on a semi-markov model. On out of sample games, the no-odds variant (used for WPA) is roughly on par with Inpredictable. The odds-aware variant slightly outperforms ESPN Analytics.

Who has been the WNBA’s most clutch player this year? To answer that, I built and tested a win probability model, then measured how much each player shifts their team’s odds of winning late in close games. Sabrina Ionescu comes out on top. Here’s what goes into the numbers 👇

Here is how each team ranks in net half-court efficiency. ◆ Liberty have fallen from third to eighth since 10 days ago (from +0.04 points per play to +0.01) ◆ The Dream's half-court offense is slowly starting to tick up. (Half-court isn't the full picture, of course.)

We're officially at the All-Star break! With more than half the season in the books, here is how every team stacks up efficiency-wise across different play contexts:

finished reading The Fortunate Fall by Cameron Reed—a queer cyberpunk story with meditations on embodiment. it's a punch in the gut. #booksky

and here is each team's net half-court efficiency. think it's fair to say the Valkyries are one of the league's two best team's right now.

With the WNBA season past its halfway mark (yes, you read that right!), here's how each team ranks in offensive and defensive efficiency in different play contexts: 🏀 Lynx, Aces, Fever and Wings are the top four half-court offenses 🔒 The Valkyries now have the top half-court defense in the league.

Point being: It’s best to read the labels as estimates, useful ones at that. Enough about the methodology, though! With the predicted labels in hand, it’s easy to calculate context-based stats. Here’s the defensive variant of the first table in the original tweet:

About nine plays per game (5% of plays) might be misclassified. In general, the rules-based classifier modestly undercounts transition plays. Differentiating between some half-court and transition plays (like seven-seconds-or-less style plays) can be tough, even for the human eye.

Every effort was made to make the labels accurate and precise. The rules-based classifier, which bundles together 20 if-then style rules, achieved 94.8% overall accuracy, 93.0% overall precision and a 92.2% F1 score, outperforming other ML models like HGBC, SVM, XGBoost and KNN.

I watched 14 WNBA games from 2026 and hand-labeled 2,538 plays as half-court, transition, scramble (putbacks and quick kickouts) or other. Then I built a classifier to estimate those same contexts from play-by-play data alone. Here are the definitions I used to label plays:

Cleaning the Glass splits plays into half-court, transition and putback contexts. Those same easy-to-understand numbers haven't been publicly available in the same way for the WNBA. So, let's fix that. Introducing Cleaning the Glass-style contextual stats for the WNBA!

What are the least and most efficient shooting locations in the WNBA? Here is every 3×3 shooting zone for 2025. Each square is colored based on its points per true shot, which is the average points scored for every shot attempt (makes, misses and shooting fouls) from that spot.

Here's the same plot for 2024 and '25 combined. (Larger sample size, but it risks smoothing different team makeups and coaches. Data for Valkyries is just '25.) These are all just exploratory diagnostics.

Getting very into the weeds here: Home vs. away rim attempt rates aren't a clean counterfactual, but the splits might (?) be a rough signal of venue-level effects on how shot locations get encoded. Most teams are within ~5%. Valkyries and Aces are definitely beyond that.

There's also clearly been a sharp uptick in the number of fouls called per game through the first month of the season. Games are averaging about 12% more fouls in 2026 than they were in 2025.

Another effect of the league's emphasis on freedom of movement: Points per 100 possessions* from both made field goals and free throws are up in 2026. (*there might be small discrepancies based on possession counting method and source.)

Say goodbye to the two-hour WNBA game? On average, it's taken 128.2 minutes to get from the opening tip to the final buzzer in 2026. It's never taken longer to complete a game during the first four weeks of the season. More fouls ➡️ Longer, choppier games

Let's talk about stability. The best WNBA one-year RAPM model had an adjacent season Spearman rank correlation (for individually modeled players) of 0.34. That's ~weak to modest and shows that there's signal left on the table.

The results also indicate that WNBA RAPM tends to favor a lambda value between 3000-4500. The threshold for low-minutes players is calculated as a training set minutes percentile. (30th percentile ~ 11 minutes per game.) Even with regularization, a cutoff helps predictive performance.

I believe (?) this is the most exhaustive public look into how various RAPM design choices perform, so let's dig in The difference between using stints or possessions appears negligible One-year setups preferred dropping low-minutes players; multi-year preferred pooling them

For each window, the best performing configuration was selected based on median RMSE across all seasons. Those are listed in the table below. (The constant-margin baseline median test RMSEs were 13.19, 13.19, and 13.21 for the one-year, three-year, and five-year settings.)

Those configurations then get validated through chronological five-fold cross validation and tested on a held-out sample, which corresponds to the final 30% of the evaluation season. Viz showing how this works for a three-year window:

Can a WNBA-tailored RAPM model provide similar predictive power compared with its NBA counterpart? If so, how should it be built? The answer to the first is yes. Following Joe Sill's methodology, I built and tested a WNBA version of RAPM. Here's what I found.

Here's a visual look at how much of last year's team each franchise brought back: (Liberty and Wings appear lower in the table than they otherwise would have been because of their head coaching changes. Lower/higher continuity score doesn't necessarily signal anything bad/good.)

predictions of integrity? juuuuust missed out on the top eight, but I'm very happy with the score my models acheived. (No gambling or hand-modifying of predictions) congrats to everyone else!

Did Veronica Burton beat Courtney Williams that bad? Burton's win was the fastest Unrivaled 1-on-1 tournament victory for any non-finals game.