What we cover

1.At a glance

The rest of this page breaks each of those down, and section 9 lists what is not covered — which is the part most worth reading before you rely on any of it.

2.Leagues

NBA — complete

All 30 teams are covered end to end: schedules, rosters, team box-score totals, player game logs, offensive and defensive ratings, and per-stat form caches, all rebuilt nightly. Every model on the site works for any NBA player against any NBA opponent.

NCAA Men's Division I — partial

The college stack is real and shares the same simulation engine as the NBA path: its own database schema, its own data provider integration, season metadata, team and player tables, schedules and per-game logs, and a game environment that correctly uses 40-minute games rather than 48.

Mixed matchups resolve sensibly: a game counts as a college game only when both sides are college. Any NBA involvement makes it a 48-minute game, and one game length applies to everyone on the floor.

3.The nine models

Eight models predict something directly; the ninth combines them. Each one is trained and loaded per position, so a centre and a guard are not scored by the same weights.

Open any of them here, or read how the model heads work.

4.What Team Matchup produces

Set two teams, edit the lineups and minutes, and one simulation answers all of the following without re-running:

• Spread — the projected margin, cover probabilities against a line you type, and expected value at a price.
• Total — projected combined score, over/under against a line and price, and the break-even probability that price implies.
• Range probabilities — the chance the total, or the margin, lands between two numbers you choose.
• Win probability — blended from three independent estimates rather than one.
• Simulated box score — points allocated across the players you put on the floor, so the team number decomposes into people.
• Playoff mode — a separate game environment for postseason conditions.

5.What the player prop models produce

Every per-stat model returns the same family of outputs:

• The full distribution — a plotted curve over 20,000 simulated outcomes, not a point estimate. Combo Props uses 30,000.
• Median and confidence interval at a level you choose.
• Conditional point estimates — a low and a high estimate reflecting which scoring regime the player is in, rather than one blended number.
• Over/under against a line and American odds on both sides, with break-even probability, probability edge and a strength label.
• Probability between — the chance the stat finishes inside a range.
• Shapley attribution — the projection split into Dime, Poor Defense and Support Skill, reportable at a quantile of your choosing.

Some models add stat-specific analysis: rebounds and 3-point shooting separate easy opportunities from contested ones, and blocks and steals break attribution down further again.

6.Player pool

Current players

Anyone on an NBA roster, projected from nightly-refreshed game logs, recent form and opponent context. Depth charts and bench units are available, so you are not limited to starters.

All-time players

Nine hundred retired players can be dropped into any matchup. Because they have no current-season game logs, they are projected from career production and attribute ratings instead — which is why the mean anchor disables itself when you select one. This is what makes cross-era hypotheticals possible, and it is also why those results should be read as structured speculation rather than as forecasts.

7.Scenario controls

Coverage is not only which players exist — it is how much of the situation you can specify. The right-hand column is roughly how much each control moves the answer.

8.Data freshness

NBA data refreshes nightly at 06:00 US Eastern across thirteen phases — game totals, rosters, team ratings, player stats, per-minute production, hotness, and a dynamic-feature cache per stat. The run is incremental, so it collects the games that actually happened rather than rebuilding the season.

College data is ingested through a separate job against a more heavily rate-limited feed, so its cadence is slower by design.

• A prediction reflects everything up to the last completed refresh.
• Games that finished last night are included after that morning's run.
• Nothing from today — lineup news, a late scratch, a rest decision — is in the model until the next run.
• Phase failures are isolated and logged, so a single upstream problem degrades one slice of data rather than the whole night.

The full phase-by-phase breakdown is in How it works.

9.Not covered yet

An accurate coverage page has to include the holes. These are the current ones.

• Live and in-game. No in-play updating, no live odds, no score-aware re-simulation. Everything is pre-game.
• Injury and news feeds. No injury reports, availability designations, rest notices or beat reporting. You have to apply that context yourself by editing minutes and lineups.
• Full Division I college coverage. The stack works; the carded roster set is still small. See section 2.
• Other leagues. No WNBA, EuroLeague, G League or international competition.
• Other sports. Basketball only.
• Correlated combo props. Combined lines assume independence between stats, so combined ranges are narrower than reality.
• Saved or shareable results. Predictions are computed on demand and are not stored — refreshing the page loses the scenario.
• Published accuracy tracking. Calibration is applied inside the models, but there is no public hit-rate or backtest page yet.
• Automated line shopping. You enter lines and prices yourself; nothing is pulled from a sportsbook.

For how any of this is built, see How it works. For definitions, see the Glossary.