Baseball analysis
Research on pitching, hitting, and projections, alongside tools for comparing contracts and building trades. The articles report the tests and their limits; the trade tools use a July 31, 2026 data snapshot.
Featured research October 16, 2026
A weak spot only matters if the pitcher on the mound can reach it. In a retrospective 2024–25 test, the part of opponents’ plans aimed at a specific hitter matched about 46% of what my model said each pitcher could reach, against less than 10% when hitters were shuffled. I can’t show those plans cost hitters anything: that test came back null.
Read the researchOpponents’ plans vs. the Attackability model · 2024–25
Shuffled-label placebo range: 0.027 to 0.098. 45,254 hitter–pitcher groups.
Null (p = .833). The test can’t show the plans cost hitters anything.
Retrospective 2024–25 test, not a pristine holdout.
Says nothing about trade prices.
Research note
Pitch shape reduced forecast error by 1.3% across 284 eligible pitchers. The uncertainty interval includes no improvement, and ShapeCast overestimated K-BB% more than the comparison model did.
Research
Add public Statcast pitch traits to 16 first-half statistics and second-half K-BB% forecasts get 2.6% better (95% interval 0.4% to 4.8%). But the smaller version I committed to in advance failed its own test, and this model’s RMSE was 4.6% higher than ZiPS.
Research note
Adjusting home-run forecasts for the pitches a hitter swings at made them 0.9% worse. Adding a pitcher’s secondary pitches improved fastball forecasts by 0.08%, inside an interval that spans zero. Spray-adjusted xwOBA made hitter forecasts 0.3% worse. A null on half a season of public data doesn’t prove an idea is zero, and I’ve kept all three out of the trade tools.
Research note
A ridge correction cut rest-of-season error for 2025 hitters by 4.4%, though not by enough to rule out noise. When the job was the following season, it was 3.3% worse than ZiPS alone. The trade tool runs mostly on future seasons, so the correction stayed out.
Explainer
At default settings (the offseason view, with July 31, 2026 data), the trade model puts McGonigle at $382.4 million of surplus value and Judge at $18.2 million. Judge is projected to be the better player in 2027, 5.8 fWAR to 5.4. The gap is salary and time: McGonigle is paid $150.0 million over eight seasons, Judge $200.0 million over five. These are model numbers, not asking prices.
Model guide
At default settings (the offseason view, with July 31, 2026 data), Paul Skenes comes out $4.2 million ahead of Julio Rodríguez in the trade builder, 3.9% of the larger side, and their ranges share $34.3 million of possible values. Neither number settles the deal. The ranges are fixed allowances, not confidence intervals, and they ignore a player’s age and experience.
Put players on each side of a trade and compare their estimated surplus value by season. Change projections, contracts, and model settings to see how the result moves.
Compare published ZiPS forecasts and the model’s adjustments. The completed 2026 season contributes no remaining value.
Rank 100 players by estimated surplus value from 2027 onward. Compare their projections, salary commitments, and years of control.
Choose a team and a position to compare projected upgrades and trade packages. A match does not mean the other club is willing to trade.