PRED2UAnswers › Predicting game winners

Can machine learning predict WNBA game winners?

Yes — to about two games in three, and that is exactly the problem. The best published academic result is 67.48% accuracy on WNBA winners, while the sportsbook closing favorite wins around 70% of the time. A model that cannot beat "always pick the favorite" is a science project, not an edge. This page walks through the published numbers, the baseline they should be compared against, and why we grade our own game model in public but do not sell winner picks.

What the published research shows

A 2025 peer-reviewed study (Alves & Barbosa, Computation 13(10):230) compared thirteen machine learning approaches — logistic regression, random forest, XGBoost, stacking and bagging ensembles, plus LSTM and CNN deep networks — on game-outcome prediction in both leagues. The best of all of them reached 65.50% accuracy on NBA games and 67.48% on WNBA games. Two things in that result are worth keeping. First, the ceiling: more model complexity did not push past the high 60s, which matches every earlier attempt in the literature. Second, the league gap: WNBA outcomes were measurably more predictable than NBA outcomes, consistent with what we see in prop markets, where WNBA lines run 8–12 cents wide against 3–5 for comparable NBA props.

The baseline every accuracy number needs

An accuracy figure means nothing until you name the dumb strategy it has to beat. Three baselines, in ascending order of difficulty: a coin flip (50%), the home team (54% this WNBA season, 176 of 325 games), and the closing moneyline favorite. That last one is the real bar. The average closing favorite carries a vig-removed implied win probability around 70%, and to the extent the market is calibrated — our closing-line-value tracking says it very nearly is — favorites win about that often. Which puts the best published model at or below a strategy that requires no model at all: read the line, pick the favorite.

Our own game model, graded in public

We run a game model too — totals and margins, used as context on game pages — and it is graded on the same locked ledger as everything else. Season to date it has picked the winner in 63 of 95 scored games (66%); the closing favorite went 67–27 (71%) over the same games. Under our own noise rule nothing below 8 slates counts as evidence, so treat both numbers as a sample being accrued, not a claim. The point of publishing them anyway is the habit: every number on this site is scored against reality in public, including the unflattering ones.

Why we sell calibrated projections, not winner picks

Game winners are the most liquid, most efficient market the books run: the closing line already aggregates the same public box-score data an academic model trains on, plus injury news, lineup decisions and sharp money the model has never seen. The places a model can still add something are the less efficient corners — player props with wide lines, and probability calibration, where we publish an 80% band and a 50% core range with every projection and grade the coverage nightly. That claim — when we say 80%, it happens 80% of the time — is checkable in a way "we pick winners" never is.

How to evaluate any AI picks site

Four questions expose most of them. Was the model tested walk-forward — trained only on games before each prediction — or on a random split, which leaks future information and inflates accuracy? Is accuracy compared against the closing favorite, or against a coin flip? Does the win rate clear the vig (about 52.4% at standard -110 pricing), or just 50%? And is there a locked, dated, complete archive of every pick, or only screenshots of the good weeks? A site that fails these is selling a feeling, not a forecast.

All figures are measured against real box scores by the same pipeline that ingests the games. Information only, not betting advice. Methodology guide.