WNBA Prediction Model Explained: Usage Rate, Pace, and 82-Feature Machine Learning

The WNBA is having a moment. League viewership has grown faster than any other major American sport over the past two seasons, expansion brought in the Golden State Valkyries for the 2026 season (with Toronto and Portland on deck), and the Caitlin Clark effect has dragged a generation of new fans into following women's basketball. Sportsbooks have responded by posting more WNBA markets, deeper player prop boards, and live in-game lines that simply did not exist three years ago.

But the modeling infrastructure has not caught up. Sites like Dimers, Covers, and BettingPros publish daily WNBA picks pages — they do not publish methodology pages explaining how those picks are made. ESPN runs a BPI projection table; Her Hoop Stats serves the analytics community. Nobody in the top ten search results for "WNBA prediction model" actually explains how their model works.

This page is that explainer. Everything below describes the model that powers our daily WNBA picks — what it looks at, why those features matter, where it has been wrong, and what we did about it.

Our training data covers 254 WNBA games from 2024 and 315 from 2025 — 569 finished games total. For player-level training, that expanded into roughly 13,000 player-game records once you account for both teams across both seasons. The 2026 season started April 25 and we now have around 30 graded games to layer in for live evaluation.

For comparison, our NBA models train on 11,000+ games across three seasons — almost 20x more game-level data than WNBA. This sample-size disadvantage is the single biggest constraint on what a WNBA model can learn. It forces the architecture toward features that generalize from small samples (usage rate, role) and away from features that only become reliable in large samples (head-to-head matchup tendencies, specific defender effects).

Data sources: API-Sports for game schedules and odds (league_id=13, "NBA W"), official WNBA box scores for player stats, and live boxscores for grading. Player name formatting has historically been inconsistent across seasons — the same player can appear as "Caitlin Clark," "C. Clark," and "Clark, Caitlin" in different feeds — so name normalization is a meaningful step in the pipeline.

The production WNBA props model uses 82 input features. They group into seven categories:

• Player rolling production (24 features): last 3, 5, and 10-game averages for points, rebounds, assists, minutes, FG attempts, and 3PT attempts.
• Usage and role (8): share of expected production, projected minutes, is-likely-starter flag, position-adjusted usage rate, recent minutes trend.
• Team context (12): team pace, offensive rating, defensive rating, recent form, home/away split, days of rest.
• Opponent context (12): opponent defensive rating overall, defensive rating vs guards/forwards/centers, opponent pace, opponent recent form.
• Matchup specifics (10): historical h2h player vs team, days since last meeting, head-to-head pace.
• Schedule context (8): back-to-back flag, road trip length, time-zone change, rest differential.
• Streak and momentum (8): player hot/cold flags, consecutive games above/below recent average, recent shooting variance.

That mirrors the architecture of our MLB props model, which uses an analogous 82-feature set adapted for baseball. The category mix differs — baseball relies more on Statcast biometrics, basketball relies more on usage and minutes — but the modeling philosophy is the same: many small signals stacked carefully into an XGBoost classifier, calibrated against actual outcomes.

The single most important feature in our WNBA props model is player_share_of_expected — the percentage of team production the player generates while on the court, weighted by their projected minutes for the upcoming game. This feature alone outranks every other input including raw recent averages.

The reason is that recent averages are noisy in a 34-game season. A player who scored 18, 22, and 14 in her last three games has a 10-game average that bounces wildly with each new game. Usage rate is far more stable: if A'ja Wilson is using 31% of Las Vegas's possessions when she is on the floor, that number barely moves week to week. Multiply usage by projected minutes and you get a clean estimate of expected output that is robust to the noisy individual-game swings that dominate small-sample basketball.

This is also why our model is structurally different from NBA models. In an 82-game NBA season, recent averages have settled enough by the All-Star break that they are reliable. In a 34-game WNBA season, recent averages are barely past the "hot streak / cold streak" phase by playoffs. Usage rate cuts through that noise.

The second-most-important feature is is_likely_starter — a binary flag predicting whether the player is in the starting five for the upcoming game. The third is projected_minutes. Together, these three top features make up nearly 30% of total feature importance in the model.

Why does "is this person starting?" matter so much? In the WNBA, starter vs bench distinction is sharper than in the NBA. WNBA bench players average 16-22 minutes per game; starters average 28-34. A player who shifts from starter to bench (or back) in a given week sees their expected output change by 40-50%. If the model thinks she is starting and she isn't, that prediction is dead.

Our pipeline pulls confirmed lineups from official feeds when available and falls back to a lineup-prediction sub-model based on recent rotations. The is-likely-starter flag combines those two sources into a single confidence score the main model can consume.

Production V4 is the fourth iteration of the WNBA props model. The earlier versions did not pass our validation gates. Here is what we tried and what failed:

V1 was a straight XGBoost classifier on a baseline 38-feature set ported from our NBA model. It hit 58% on assists at 60%+ confidence — barely above coinflip. The shuffle test showed weak feature signal across the board. Diagnosis: the NBA-derived features assumed an 82-game sample and did not work on the WNBA's smaller sample.

V2 added usage rate as a derived feature, dropped the noisier recent-game features, and rebuilt the time-aware validation split. Hit rate jumped to 61.5% at 60%+. Closer but still below our 64% ship gate.

V3 added projected minutes and the is-likely-starter flag, and switched the calibration from raw probability to rank-aware Platt scaling. Hit rate at 60%+ reached 63.8% on assists. Still 0.2 percentage points below the ship gate.

V4 added the player_share_of_expected feature (usage weighted by projected minutes) and tightened the rolling-window features to 5 and 10 games (dropping noisy 3-game windows). Final hit rate: 64.2% at 55%+ confidence on assists. Passed the gate. Shipped to production on May 14, 2026.

The lesson buried in this iteration history: features matter more than algorithms. We did not change the XGBoost configuration meaningfully across versions. What changed was the input signal. Two of the three top features in V4 did not exist in V1.

We run two separate model families for the WNBA:

Game models predict moneyline (which team wins), spread (winning margin), and totals (combined points). These use team-level features like offensive efficiency, defensive efficiency, pace, rest, and recent form. They are trained on the 569-game historical set and produce one prediction per game per market.

Props models predict individual player over/under outcomes for assists and rebounds. These are the 82-feature models described above. Points props are currently parked because the data is too thin to model reliably — once the 2026 season produces another 50+ player-game days, we will revisit.

Game and props models share a common data-fetch layer (the same boxscores, the same lineup feeds, the same rest calculations) but diverge into separate XGBoost classifiers with separate hyperparameters and separate calibration. This separation makes each model easier to debug — when assists are calibrated and rebounds are over-confident, we can trace exactly which model is wrong without contaminating the other.

Two production filters are applied to V4 output before display:

Assists display confidence cap at 65%. The raw model produces probabilities up to 78% for some assists picks at very high projected minutes. Backtesting showed those high-confidence buckets actually hit at 64-66%. Capping the displayed confidence at 65% honestly represents what the historical data supports. The raw probability is preserved internally for sorting; only the displayed badge is capped.

Rebounds OVER@60+ filtered out. The OVER side of rebounds props at 60%+ confidence backtested to only 45.5% — worse than coinflip. The UNDER side at the same confidence tier hit 69.1%. We display only the UNDER picks at the high tier. This is an unusual filter, but it reflects what the data actually shows: the rebound market is biased OVER, and our model captures that bias only on the UNDER side.

These filters were added based on backtesting after V4 shipped — they are not part of the model itself. They are layered on top of the model output to align production picks with what the validation data can actually defend.

How accurate is your WNBA prediction model? Our V4 props model hits 64.2% on assists at the 55%+ confidence tier and 69.1% on rebounds at 60%+ (UNDER side). Game models passed walk-forward validation on 569 historical games. All picks are graded daily and the running accuracy is published at our performance page.

Can I see picks before subscribing? Some preview picks are available on the public landing page. Full WNBA assist and rebound props live behind the subscription paywall. Game-level moneyline, spread, and totals predictions are also subscriber-only.

When are predictions posted each day? The WNBA pipeline runs at 13:00 UTC and again at 19:00 UTC. The morning run captures predictions before lineups are confirmed; the evening run picks up confirmed lineups and updates the picks accordingly. Predictions for tonight's games are usually finalized 2-4 hours before tip-off.

What about WNBA points props? Points props are currently parked due to a data-limited bias in the model output. We expect to enable them once the 2026 season produces 50+ additional player-game days, likely mid-July.