Steamer Projections
The Basics of Projection SystemsForecasting the upcoming season is essentially the same as determining current ability.Most projection systems are modifications on the same simple system (Marcel “the monkey):
Weighs stats from more recent seasons more heavily
Regress to the mean
Why regress to the mean?Results = Ability + Luck
Two Examples of Marcel in Action
18.3%23.0%
SteamerAlong with most “fancier” systems:
Uses adjusted minor league statistics in addition to MLB stats.
Adjusts for home ballparks, league, starting v. relieving
What makes Steamer distinct: We use a different system for each component (K%,
BB%, HR%…) We regress to a different “prior” for each player
Projecting Joaquin Benoit’s K% in 2011:4 possible forecasts
28.0%
26.1% 23.7%
24.9%
Actual K%: 26.1%
K/PA for All Pitchers: 1993-2011
HR/PA for All Pitchers: 1993-2011
Regression is Bayes
Distribution of MLB talent
ProjectionLikelihood of player statisticsGiven different levels of talent
K% v. FBV for Starters
K% v. FBV for Relievers
Matt Thornton 2012
24.0% 27.2%
Marcel error v. Fastball Velocity
More regression = Stronger Relationship
It might be working…
Where to go from here?For Pitchers:
Develop a better measure of stuff than fastball velcoity Jeremy Greenhouse: StuffRV based on velocity and
movement Josh Kalk/Brooksbaseball: Similarity Scores based on
pitchf/x
For Hitters: Can something similar be done with hitf/x? Trackman?
Speed off the bat Trajectory