NBA Match Predictor

Win model: logistic regression

The win model uses ridge-regularized logistic regression: it learns a weight for each input, adds their contributions, and converts the result into a probability. Ridge regularization discourages large weights on the four team-difference inputs to reduce sensitivity to noise; the home-court weight is unpenalized.

Every difference is home minus away. Multiply each input by its learned weight, then add the five contributions to get a combined evidence value called z. Positive evidence favors the home team; negative evidence favors the away team.

The conversion is home win chance = 1 ÷ (1 + exp(−z)), where exp means the exponential function. Zero combined evidence gives a 50% chance, and there is no extra intercept. A weight changes the combined evidence, not a fixed number of percentage points.

Where the inputs come from

Elo in plain English: Elo is a running team-strength rating based on wins, losses and the strength of each opponent. An upset win earns more rating points than an expected win, while an unexpected loss costs more. The predictor uses the gap between the two ratings as one input; rating points measure team strength, while basketball scoring margins are handled separately.

Team-strength ratings start at 1,500 and update after each completed game using the win or loss relative to the rating-based expectation, with an update factor of 20. Rating updates allow 65 rating points for home court, or zero at a neutral venue; this is separate from the learned home-court weight used in predictions. At a new season, half of each rating’s distance from 1,500 is retained, and recent scoring histories reset.

Recent form uses the previous ten completed games in the current season, or the available games when there are fewer. Days since the last completed game are capped at seven, with seven used when no earlier date is known. The back-to-back input records whether each team played on the previous calendar day. In Compare, optional rest assumptions replace those two schedule inputs for that hypothetical only; the original game-history estimate stays visible for reference.

Score model: two linear regressions

What margin points mean: the margin is the home team’s predicted points minus the away team’s predicted points. For example, 112–109 means a +3-point margin for home, while 109–112 means −3. A +2-point input contribution nudges the expected gap two basketball points toward home; −2 nudges it toward away, with other inputs held fixed. If the total stays the same, +2 to the margin means +1 to home’s projected score and −1 to away’s.

Ordinary least-squares regression separately estimates total points and the home team’s scoring margin. The margin uses the same five inputs as the win model, with its own learned weights shown below. A positive margin favors home; a negative margin favors away.

The scoring environment is the average of the two teams’ recent combined game totals: (home points scored + home points conceded + away points scored + away points conceded) ÷ 2, with each value an average over that team’s previous ten current-season games. It measures recent scoring, rather than possessions or pace.

Home points = (total + margin) ÷ 2.
Away points = (total − margin) ÷ 2. A margin contribution moves the two teams’ projected points in opposite directions by half that amount. A total-points contribution is split equally between them.

What the models can see

Scheduled predictions use completed regular-season games from strictly earlier New York game dates. Future hypothetical comparisons can also use today’s already-known finals. The win model needs at least five earlier games per team across seasons; the score model needs at least five earlier games per team in the current season.

During win-model training, each earlier season received half the fitting weight of the next; the score regressions weighted eligible training games equally. Learned weights stay fixed while new results update the inputs. On Today, recent medical Out reports can move the Team model’s win chance. The experimental adjustment adds 0.071 to home’s win evidence for each additional away-team Out compared with home. It uses recent Sleeper news timestamps as a freshness check. The projected score and the Archive’s baseline forecasts use game results.

DARKO preview

The player ratings come from DARKO’s Daily Plus Minus (DPM). DPM estimates how many points a player adds on offense or prevents on defense per 100 possessions, compared with an average player. DARKO methodology

Our preview starts with DARKO’s expected minutes and removes players listed as Out by ESPN. It projects up to ten available players per team and divides 240 minutes among them, up to 42 minutes per player. A replacement’s rating counts in proportion to the minutes they are expected to play. Players with uncertain status are assumed to play.

Expected pace, league scoring and home court turn the two rotations into projected scores. The projected scoring gap then sets the win chance. These are experimental estimates based on projected rotations.

How the models were trained

Training started with 1,230 real regular-season games from 2023–24. For each game, we built inputs from earlier game dates and used the recorded winner and final score as the answers to learn from. The score model used games where both teams already had at least five games that season.

We compared candidate models on later games in 2024–25. We then tested the selected win model on 2025–26 and later ran a separate historical replay for the score model on that season. The live site uses the selected methods refitted on eligible games across all three seasons.