I built a percentile-normalized performance score for MLB players (wOBA/ISO/FIP/K9/WHIP/BB% weighted composite) — feedback welcome
I've been building a composite performance metric for MLB players — calling it "BAX Score". The goal is a single number that reflects "how good is this player playing right now," built only from real performance stats (no market/demand signal in the mix at all — this matters because the score also drives pricing in a fantasy stock market I'm building, and I didn't want volume to be able to move it).
Methodology, roughly:
- Pull wOBA, ISO, FIP, K/9, WHIP, BB% (hitters and pitchers get separate axis weightings)
- Percentile-normalize each stat against the current league sample
- Apply sample-size shrinkage so a guy with a hot two weeks doesn't swing the score around
- Weighted composite across the normalized axes → single score, updated daily after games finalize
Pulled today's top-of-leaderboard hitters and plotted wOBA (the single most correlated input) against the full composite score to see how much the extra axes actually move things:
A few things stand out:
- Caminero, Harper, and Alvarez score noticeably ABOVE where wOBA alone would put them — power/ISO and BB% are dragging the composite up past what a wOBA-only ranking would suggest
- Kurtz, Soto, and Ohtani score BELOW their wOBA-implied line — despite great wOBA, other axes (BB%/K% mix, sample shrinkage) pull them down relative to the trend
Basically: if I'd just ranked by wOBA, the order would look meaningfully different from the composite. Curious whether this sub thinks that's a feature or a red flag — is baking in ISO/BB%/K% on top of wOBA adding signal, or just re-weighting the same underlying skill twice?
Also still going back and forth on:
- How much weight FIP vs xFIP should get for pitchers with small sample IP
- Whether BB% deserves its own axis or should just feed into wOBA indirectly
Happy to share more of the pipeline if useful and your feedback :)
(Side project — baxeball.com — mostly posting here because I want the methodology stress-tested, not to plug it.)