How our WNBA player projections work

Everything on the site, explained in plain English: what a locked prediction is, what the ranges mean, how accuracy is measured, and how to read the odds page without fooling yourself.

Locked predictions: the whole point

Anyone can publish predictions. Almost nobody publishes predictions that were provably written down before the game and then scored honestly afterwards. That is the entire product.

Before every WNBA game, the model writes a projection for each player — minutes, points, rebounds, assists — into a ledger with a timestamp. Those numbers are never revised. After the game, the real box score is written next to them and the miss is computed. The Record page shows every one of these, including the bad ones, with the biggest misses sorted first.

The 80% ranges (those little bars)

A projection like 18.6 points is a best guess, not a promise. Basketball is noisy: the same player in the same matchup can score 8 or 30. So every points projection ships with a calibrated 80% range, drawn as a band on a bar.

"Calibrated" means the range is tested against reality: the actual result should land inside the band about 80% of the time — not 99% (bands too wide to be useful) and not 50% (bands lying about certainty). The Record page reports the real coverage rate so you can check us.

On scored games you'll see a dot on the bar: green means the actual landed inside the band, red means it landed outside.

How accuracy is measured

Typical miss (MAE)

The average distance between projection and reality, in the stat's own units. A points MAE of 3.4 means we're off by about three and a half points on an average prediction. Mean absolute error, not a cherry-picked highlight reel.

Baseline

Every model must beat the dumbest sensible alternative: the player's season running average. We publish that comparison as lift. If lift is small, we say it's small.

Within 3

The share of predictions that landed within 3 points of the actual.

Band coverage

How often the actual fell inside the 80% range. Target: 80%.

Player tiers

Accuracy split by projected scoring load. Bench players are easier to predict than primary scorers; the split keeps us from hiding behind easy predictions.

Live projections: our best result

While a game is in progress, the model publishes an updated projected final line for every player, refreshed throughout the game. This is the genuinely strong part of the system: once a player has meaningful floor time, the live projection is measurably more accurate than the locked pre-game number.

Live weight

How much of the live number comes from what's happening on the floor versus the pre-game model. Early in games it stays low on purpose — pure in-game extrapolation is worse than the pre-game model in the first quarter, so the blend earns trust as minutes accumulate.

Delta (Δ)

Live projected final minus the locked pre-game projection. A positive delta means the game is telling us to expect more than we did at tip.

The odds page: edges and the hold

The Odds page puts our projection beside real sportsbook player-points lines. The edge is the difference in points. But the single most important column is the hold.

A row only gets a colored edge and a "covers hold" badge when the disagreement is big enough to matter. Everything else is labeled "inside hold" — informative, not actionable. PRED2U is an information product, not betting advice.

The probability columns

Hold

Turn the over price and the under price into implied probabilities and they add up to more than 100%. The excess is the hold — the book's cut on the market, no matter which side wins.

Model ov

The chance the player goes over the line, read off the locked projection: the projection and its 80% range are treated as a bell curve, and Model ov is how much of that curve sits above the line.

MC ov

The same question answered by simulation instead of a curve: the player's game is replayed thousands of times — minutes first, then shot attempts, then makes — and MC ov is the share of simulated games that clear the line. Scoring is skewed (bad nights bunch together, big nights stretch out), and the simulation captures that shape, so in lab testing it was the better-calibrated probability at every threshold, on two full seasons. It deliberately reads lower than Model ov on stars' high lines. When the two disagree, trust MC ov.

Break-even

Open a row and you'll see the book's break-even: the win rate the over must have just to pay for itself at the posted price. It sits above 50% even on a fair line — that's the vig. An over is only interesting when our probability clears the break-even, not just 50%.

Colors, chips and small print

• Green: above the line / inside the band. Never "good" — a projection isn't good or bad.
• Red: below the line / outside the band.
• Amber: inside the noise floor — a difference too small to mean anything.
• DTD chip: the player is listed Day-To-Day. OUT: ruled out. Projections for questionable players carry wider ranges.
• DNP / Inactive in a game log: did not play (coach's decision or injury) / not in uniform.
• "As of" timestamps on live views: if the time looks stale, the data is stale. We show it so you know.

What we're honest about not being good at

• Game totals and spreads: the sportsbook market is more accurate than our game-level model. We publish that comparison and don't sell it as a strength.
• Beating a naive average: our edge over a simple season running average on points is real but modest, and stated as such on the Record page.
• Minutes: the hardest stat — rotations change with blowouts, foul trouble and coach decisions. Our minutes miss is published like everything else.

You will never see "guaranteed", "lock of the day" or "beats Vegas" here. Performance claims are checked against the ledger, and the ledger is public.

Glossary