METHODOLOGY

All statistics come from a public, MIT-licensed box-score dataset covering every NBA regular-season game from 2010–11 through 2023–24. This project uses the six most recent complete seasons in that range (2018–19 → 2023–24), combining player box scores (points, rebounds, assists, steals, blocks, turnovers, shooting splits, minutes) with team game logs (wins/losses) to compute each player's team win percentage that season.

To avoid small-sample noise from players who only appeared in a handful of games, the model only ranks players who played in at least ~58% of their team's games (mirroring the NBA's own MVP-eligibility convention) at 24+ minutes per game. This mirrors how award voters filter candidates before they ever look at a stat sheet.

Rather than training a model against historical MVP voting results — which bakes in decades of media narrative and market-size bias — this project builds a fully transparent composite score from six season-relative categories, each expressed as a z-score (how many standard deviations a player is above or below the league average that season):

• Scoring — points per game
• Efficiency — true shooting percentage
• Playmaking — assists per game and assist-to-turnover ratio
• Defense — steals + blocks per game
• Rebounding — total rebounds per game
• Team success — team win percentage

Using z-scores computed within each season means a 27-point scorer in a slow, defensive 2013-14 league and a 27-point scorer in a fast, high-scoring 2023-24 league are judged against their own era, not a raw career-average threshold. The default formula weights scoring and efficiency most heavily (30% and 20%), reflecting how voters have historically valued go-to scorers — but every weight is adjustable on the Formula Builder page.

Public, well-documented MVP voting-share datasets are scarce and inconsistently maintained, and a model trained to imitate past voters mostly ends up re-learning their biases (market size, narrative, name recognition) rather than measuring performance. A transparent weighted formula is more honest about what it is: a specific, statable philosophy about what makes a season valuable, that you can inspect, challenge, and rebuild in seconds.

Beyond settling MVP arguments, the same underlying per-category scores are directly useful for:

• Fantasy roster construction — surface undervalued two-way contributors who score outside the top 10 in points but rank highly on defense + efficiency.
• Scouting narratives — the Compare tool turns a stat sheet into a shareable, presentation-ready shape comparison.
• Front-office "what would it take" modeling — the Formula Builder lets a team explore how a specific archetype (three-and-D wing, pure table-setter) would rank under different value systems.

This model doesn't account for injuries relative to team need, narrative moments (buzzer beaters, playoff seeding races), advanced tracking data (defensive matchup difficulty, shot quality), or media coverage — all real factors in actual MVP voting. It is intentionally a clean, explainable baseline rather than a claim to out-predict human voters.