Moneyball Revisited: Sabermetrics past, present, and future

Moneyball Revisited:
Sabermetrics past, present, and future

Ben Baumer
Mount Wachusett Community College
November 28th, 2017
(http://bit.ly/2iZpX39)

What is sabermetrics?

statistics
computing
math
baseball
~~physics~~

A brief history of sabermetrics

Early 20th century

1900s: Henry Chadwick invents the boxscore

Early 20th century: data

Boxscores
- game level
- dozens of rows

Early 20th century: tools

1880: punched cards invented (by IBM)
1935: punched cards used by SSA

Early 20th century: methods

1907: Galton’s paper on wisdom-of-crowds
1908: Gosset’s t-test
1918: Fisher introduces variance
1935: Design of Experiments

Early 20th century: what did we learn?

Batting
- batting average
Pitching:
- earned run average (ERA)
Fielding
- fielding percentage

Branch Rickey era

1947: Allan Roth, Brooklyn Dodgers statistician

Branch Rickey era: data

Tabulated
- recorded by hand (?)
- hundreds of rows (?)
1951: Official Encyclopedia of Baseball
- games played
- batting average (for hitters)
- wins and losses (pitchers)

Branch Rickey era: tools

1940s: Monroe LN-160X
1947: Victor is the world’s largest exclusive manufacturer of adding machines
1950: Turing test

Branch Rickey era: methods

1953: Metropolis’s paper on Monte Carlo methods
1959: Lehmann’s Testing Statistical Hypotheses
1970: Codd’s “A Relational Model of Data for Large Shared Data Banks”

1970: Metropolis-Hastings for MCMC

Roth on math

The figures concerned in baseball statistical work don’t call for integral calculus or even advanced algebra.

Allan Roth

Rosenthal, ed., “The Statistician,” in Baseball Is Their Business, 1952, p. 140

What did we learn?

Batting:
- OBP better than batting average
- Platoon effects
Pitching:
- Pitch counts
Fielding:
- Spray charts

Bill James era

1977–1988: Bill James popularizes “sabermetrics”

Bill James era: data

Compiled:
- seasonal, some batter vs. pitcher
- thousands of rows
1969: “Big Mac”

Bill James era: tools

1973: TI SR-10 pocket calculator
1974: Ingres
1977: Apple II
1981: MS-DOS first released
1986: SQL becomes and ANSI standard

Bill James era: methods

1977: Tukey, Exploratory Data Analysis
1977: EM algorithm
1979: Efron discovers the bootstrap

Bill James era: what did we learn?

Pythagorean Expectation
Win Shares
Batting
- Runs Created
Pitching
- Game Score
Fielding
- Range Factor

Internet era

1990s: rec.sport.baseball connects the next generation of sabermetricians

Internet era: data

Seasonal
- player-team-season
- 100,000 rows seasonal
1995: Sean Lahman publishes first online database

Internet era: tools

1990: Windows 3.0
1995: Excel 7.0 can handle at most 16k rows
1998: iMac G3

Internet era: methods

correlation
regression
Markov chains

Internet era: what did we learn?

There is little if any difference among major-league pitchers in their ability to prevent hits on balls hit in the field of play.

Voros McCracken

McCracken, “Pitching and Defense: How Much Control Do Hurlers Have?,” BaseballProspectus.com, Jan 23, 2001

Moneyball era: data

Play-by-play
- batter vs. pitcher
- several million rows
1996: Retrosheet starts publishing play-by-play data online

Moneyball era: tools

2003: Excel 11.0 can handle at most 64k rows
2005: MySQL powers Google

play,1,0,hendr001,00,X,9/F9S
play,1,0,whitd001,22,BBFFC,K/C
play,1,0,alomr001,32,BBCFBB,W
play,1,0,cartj001,00,X,4/P4
play,1,1,raint001,00,X,43/G34
play,1,1,coraj001,01,CX,4/P4MD
play,1,1,thomf001,30,BBBB,W
play,1,1,ventr001,00,B,WP.1-2
play,1,1,ventr001,11,B.CB,WP.2-3
play,1,1,ventr001,32,B.CB.SFFBS,K

Moneyball era: methods

regression
clustering
$k$-Nearest Neighbor

Moneyball era: what did we learn?

Batting
- Linear Weights
- Value Over Replacement Player
- Projection systems
Pitching
- Leverage
Fielding
- Ultimate Zone Rating

PITCHf/x era

2004–2012: BB works for Mets

PITCHf/x era: data

Pitch-by-pitch
- each pitch
- several million rows per season

PITCHf/x era: tools

2000s: SQL can handle tens of millions of rows on modest hardware
2007: Excel 12.0 can handle at most 1M rows

PITCHf/x era: methods

mixed models
Bayesian methods
generalized linear models

PITCHf/x era: what did we learn?

WAR
Pitching
- Spin rates
- Catcher framing

Statcast era: data

2015: Statcast deployed
Player-tracking
- every 1/15th of a second
- several million rows per game

Statcast era: tools

Hadoop
Apache Spark
Cray supercomputer!

Statcast era: methods

machine learning
- regularized regression
- decision trees
- random forests
- neural networks
- support vector machines

Statcast era: what have we learned?

Batting:
- exit velocity
- launch angle
Fielding:
- route efficiency
- pop time

Summary

data	rows	tool	method	result
boxscores	dozens	paper	arithmetic	rates
tabulated	hundreds	adding machine	arithmetic	formulas
compiled	thousands	calculator	algebra	metrics
seasonal	100 k	spreadsheet	regression	metrics
play-by-play	1 M	SQL (laptop)	regression	fitted models
pitch-by-pitch	10 M	SQL (server)	mixed, Bayesian	statistical models
player tracking	billions	Spark	ML	?

Example: Catcher Framing 1

Jose Molina (good)

Ryan Doumit (bad)

Example: Catcher Framing 2