Models built from pitch-level Statcast data, not just box scores.

Most pitch-quality metrics grade a pitch by its physics — velocity, spin, movement — and stop there. BDI instead asks how uncomfortable the pitch actually made the batter, measured two ways: how much it threw off their swing mechanics, and how often it fooled them into chasing a pitch outside the zone they normally wouldn't.

• The swing-mechanics half measures deviation from a batter's own normal swing against that pitch type — bat speed, bat path, swing plane — so a slower, more off-balance swing scores as more disruption. The chase half measures how often a pitch got that same batter to swing at something outside the zone they'd normally take.
• Both are measured against each batter's own pitch-type-specific baseline, not a league average, so a curveball isn't automatically flagged as "disruptive" just for being a curveball — the baseline already expects it to look different from a fastball.
• The two components combine into one score, then get checked against a batter outcome (xwOBA) the model never saw while it was being built — a way of confirming the "discomfort" it's measuring actually shows up in real results, not just on paper.
• On the 2025 leaderboard, Sonny Gray's sweeper and Fernando Cruz's splitter both rank near the top for opposite reasons — Gray's works almost entirely by getting batters to chase, Cruz's works by wrecking their swing mechanics when they do connect. Same ranking, two different kinds of "nasty."

A decomposition of pitcher xwOBA into the underlying skills that produce it — strikeout-minus-walk rate, weak-contact rate, and barrel rate — to separate repeatable skill from outcome noise, plus a companion metric for how much of a season was luck.

• Regressed weak-contact rate, K%−BB%, and barrel rate against xwOBA across the 2025 season's qualified starters, explaining 93.6% of the variance with every component significant beyond the 1% level.
• Rebuilt the formula fresh for each season back to 2015 and tested it as a year-over-year predictor against a pitcher's own lagged xwOBA — the standard "did last year predict this year" benchmark — and it won in five of eight seasons tested.
• Introduced a companion metric, Outcome Over Expected (OOE) — the gap between a pitcher's actual xwOBA and their SkwOBA — as a proxy for how much of a season was luck rather than skill; confirmed it carries next to no predictive power of its own (R² = 0.001), exactly as a luck proxy should.

A recommendation system for pitchers' repertoires — which pitches to throw more, which to drop, which new pitch to add — built on a historical-precedent engine that finds real pitchers who made the same change and predicts outcomes from what actually happened to them.

• Combines four independently validated components — outcome-based effectiveness, arsenal movement coverage, locational command, and velocity — into one 0–100 score, with weights learned by Ridge regression against real run value and refit fresh every season from 2020–2025.
• Validated the recommendation engine with true walk-forward testing — trained only on events through a cutoff year, tested only on years it never saw — correctly ranking which of two candidate changes actually worked better 64–71% of the time for a single change, and 78.9% of the time for a full pitch swap.
• Built a dedicated swap model after finding that treating a pitch swap as two separate add/drop predictions double-counted the same mean-reversion effect and overstated the real impact.
• Packaged into a 7-page interactive Streamlit dashboard — leaderboards, a per-pitcher breakdown with ranked recommendations and named historical comps, similarity maps, and side-by-side pitcher and pitch comparisons.