Games: Predictions & Schedule
Win probability
The bar on every matchup card splits 100% between the two teams — e.g. OKC 68% / SAS 32%. This is a probability, not a promise: a 68% favorite still loses roughly one time in three. NBA games swing on hot shooting nights, foul trouble, and late-game bounces; an honest number reflects that.
Model Line (Spread)
“Model line: NYK −4.5” means the model makes the Knicks a 4.5-point favorite — it expects them to win by about that margin. The minus sign is sportsbook convention (negative = favorite), and lines are shown in half-point steps like a real book. A near-even game shows PK(“pick’em”) — no meaningful favorite.
The probability and the line always agree — they are two readings of the same prediction. The line says by how much; the percentage says how often that edge survives the randomness.
Spread Calibration
Zoom out to the whole season and this is how the line above grades over time: does the model’s own spread hold up against what actually happened (not against Vegas). On this one, ~50% is the goal: an honest line should have reality land above and below it equally. Drifting toward 60% or 40% would mean the margins are systematically too timid or too bold. The “line MAE” is the average miss between predicted and actual margin — single-game NBA results are noisy for everyone, including Las Vegas.
Confidence Tiers
- HIGH — one side is given 65% or more.
- MEDIUM — the favorite sits between 55% and 65%.
- NO PICK — the game lands in the 45–55% coin-flip zone. The card says “Too close to call — model declines a pick”rather than pretending to an edge that isn’t there. These games are excluded from the pick record and tracked separately.
Confidence Calibration
Zoom out to the whole season and this is the honesty check on the tiers above: HIGH picks should win more often than MEDIUM, and the games labeled coin-flips should land near 50%. When the bars descend in order, the labels mean what they say.
Prediction Grading
- The pick (Correct / Incorrect) — did the team the model picked win the game outright? Margin doesn’t matter here: a 1-point win and a 30-point win both count as Correct.
- The line (covered / not covered) — did the favorite win by morethan the model’s number? NYK −4.5 winning by 7 covered; winning by 1 is a Correct pick that did not cover. In the prediction log the line turns green or red accordingly.
Predictions lock about 25 minutes before tip-off and are never quietly revised — what you see graded is what the model actually said before the game.
Model Performance Metrics
Overall Accuracy
Straight-up accuracy (who wins outright, margin doesn’t matter) across everygraded game this season, including the LOW-confidence, 45–55% coin-flip games the site declines to publish as an actual pick. It’s the widest, least-filtered view of the model.
Issued Picks
The same straight-up accuracy formula, but narrowed to just the HIGH and MEDIUM confidence games — the picks actually shown on matchup cards. LOW-confidence games are excluded here entirely (the “gated” count next to it is how many got passed on), so this is the honest record on picks the model was willing to stand behind.
Model vs. Vegas
A head-to-head comparison, on the exact same games, of how often the model’s favorite actually won versus how often Vegas’s favorite actually won. Vegas’s side comes from de-viggingthe closing moneyline — stripping the sportsbook’s built-in profit margin out of both teams’ odds so what’s left is a fair implied probability, not the inflated number a bettor would actually see. Only games with a matched closing line on record are counted, so this total runs a little below the season’s full game count.
ATS vs. Closing Spread
Compares the model’s predicted margin to Vegas’s actual closing spread, on the same games. For each game, the model’s edge (its margin minus Vegas’s line) and the real cover (the final margin minus Vegas’s line) are checked to see if they point the same direction. If they do, the model’s number would have beaten the closing line that night; if the final margin lands exactly on Vegas’s line, it’s a push. Shown as a Win–Loss–Push record plus a cover rate. This is different from Spread Calibration(in Model Line above), which checks the model’s line against reality rather than against Vegas.
Brier Score
Measures how well-calibrated the model’s win probabilities are, not just whether it picked the right team. For every graded game, it squares the difference between the predicted win probability (0 to 1) and the actual outcome (1 for a win, 0 for a loss), then averages that across every game. Saying 90% and winning is barely penalized; saying 90% and losing is penalized heavily — so it rewards being both right and appropriately confident, not just right.
Lower is better: 0 is a perfect prediction, 0.25 is what a flat 50/50 guess on every game would score, and it climbs toward 1 the more confidently wrong the model is. Shown side by side with Vegas’s own Brier score (from the de-vigged closing line) on the same games, so you can see who’s actually better calibrated — not just who wins more often.
Upset Call Rate
Isolates the games where the model disagreed with Vegas: it favored the team Vegas didn’t. Every one of those disagreements counts as an “upset call,” and the record shown is how often the model was actually right on just those games. It’s the clearest single number for whether the model is finding real edges against the market or just going against the grain for no reason — a losing record here means the contrarian picks aren’t paying off, a winning one means they are.