NBA Methods

Separate the market's estimate from a basketball forecast.

What is available now

We collect NBA odds from The Odds API, compare identical lines across books, remove margin within each book, and show the median fair probability. Missing prices or unpaired outcomes stay unavailable.

Consensus and market watch

Over and under must belong to the same game, player, market, line and book. Spread pairs use opposite signed handicaps. Duplicate quotes never create extra book votes. Best price means the best payout on the exact same outcome and line.

Market Watch compares an offer against the median fair probability from at least three other books at that line. Its price-based return estimate is not a validated win probability or a recommendation. Integer lines can push; the screen compares probabilities conditional on a decision and does not estimate a cash return including pushes.

Historical player baselines

Box scores come from ESPN, starting with the 2023–24 season and updated after each game day. We use up to 30 prior regular-season appearances, with at least 20 required; play-in, playoff and NBA Cup final games are left out, matching the NBA’s official regular-season totals. Games with zero minutes, future dates, duplicate player/game IDs and ambiguous names are excluded. We do not fill missing rookies with a 50% guess.

The mean is the average observed box score. The historical probability is a smoothed hit rate, (wins + 1) ÷ (non-push games + 2). A combination uses that player's points, rebounds and assists from the same game. Equal results on whole-number lines are pushes. Last season's games may be used as historical context; they are not presented as an injury-adjusted prediction for a new rotation.

Select a player’s name for a snapshot: the last ten games against the line, how the hit rate is computed, how his recent games fell around the line and the opponent’s points or stats allowed over its last 20 games. NBA snapshots have no model and no track record yet.

Experimental NBA v1 forecasts

Each game line and player prop can show an NBA v1 probability and projection (its average and median), labeled experimental. The model runs in shadow mode: it is not used for Top Picks or articles until its forecasts have been graded against 2026–27 prices. Player Props and Game Lines can sort by Best value, which blends it with the market the way Today’s Picks does (10% model, 90% the other books’ no-vig price, the rule the shadow grade scores) and shows that blend’s expected value. The model’s own expected value is not shown.

Games: pace and points per possession come from each team’s recent offense and defense, carried over from last season with shrinkage. Home court, rest and back-to-backs then adjust them. One score distribution prices the moneyline, spread and total together, including overtime and pushes on whole-number lines.

Players: projected minutes (recent role, starts, rest, team changes and blowout risk) × per-minute production, adjusted for pace and the opponent’s allowed rates and blended with recent averages. Combinations such as points + rebounds + assists are simulated from the same minutes and correlated outcomes, never added up as if they were independent. Each forecast shows its average and its median, the value with half the model’s chance at or below it. Prop lines sit near the median, and counts are skewed upward, so an average above the line can still favor the Under. Forecasts assume the player plays, because books void props for players who don’t. Players listed out on ESPN’s injury report get no forecast. Other injury statuses are shown but not adjusted for, and teammate absences don’t change v1’s minutes.

Testing: fit on 2023–24 and 2024–25, then tested once on the 2025–26 regular season. v1 beat a home-court-only baseline on moneyline log loss (0.603 vs. 0.688). On every player market it beat our historical hit rate, at lines set from each player’s last ten games. Points forecasts remain somewhat overconfident at the extremes. Early-season forecasts don’t know about offseason roster moves, so large October disagreements with the market are expected and are not evidence of value.

Freshness and season handling

Started games and quotes older than 24 hours are hidden. Failed refreshes are reported as feed errors, not as an empty slate. The last successful snapshot time remains visible. Preseason prices, if supplied, do not enter regular-season history. Opening-night markets can appear weeks before player props; an empty prop board is normal before the season.

Sources

The Odds API: NBA coverage · NBA player box scores · 2026–27 NBA schedule