Last active
February 5, 2016 14:50
-
-
Save bayesball/7ec0782f12b9ccf3f745 to your computer and use it in GitHub Desktop.
R functions for Improved Component Predictions of Batting and Pitching Measures
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| ####################################################################################### | |
| # R functions for Paper "Improved Component Predictions of Batting and Pitching Measures" | |
| # Journal of Quantitative Analysis of Sports (2016) | |
| # Jim Albert, albert@bgsu.edu | |
| # functions fit_component_average, plot_avg_results | |
| # fit_component_obp, plot_obp_results | |
| # fit_component_fip, plot_fip_results | |
| # require installation of packages Lahman, dplyr, ggplot2, and LearnBayes | |
| ######################################################################################## | |
| fit_component_average <- function(season){ | |
| # implements component method for estimating group of batting averages for nonpitchers | |
| # for a particular baseball season | |
| # | |
| # input: season | |
| # output: a list with components | |
| # component -- gives values of K and eta for all component fits | |
| # shrinkage -- gives value of K and eta for shrinkage fit | |
| # S -- gives raw stats and component and shrinkage estimates for all players | |
| require(Lahman) | |
| require(dplyr) | |
| require(LearnBayes) | |
| ##--------------------------------------------------------------------- | |
| get.data <- function(season){ | |
| B <- summarize(group_by(filter(Batting, yearID==season), playerID), | |
| AB=sum(AB), SO=sum(SO), H=sum(H), HR=sum(HR)) | |
| B <- filter(B, AB > 0) | |
| NonPitchers <- unique(select(filter(Fielding, | |
| yearID==season, POS!="P"), playerID)) | |
| merge(B, NonPitchers, by="playerID") | |
| } | |
| ## ------------------------------------------------------------------------ | |
| fit.model <- function(data){ | |
| mode <- laplace(betabinexch, c(1, 1), | |
| cbind(data$y, data$n))$mode | |
| eta <- exp(mode[1]) / (1 + exp(mode[1])) | |
| K <- exp(mode[2]) | |
| list(eta=eta, K=K, | |
| d=data.frame(data, est=(data$y + K * eta) / (data$n + K))) | |
| } | |
| ## ------------------------------------------------------------------------ | |
| d <- get.data(season) | |
| ## ------------------------------------------------------------------------ | |
| S.SO <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$SO, n=d$AB)) | |
| S.HR <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$HR, n=d$AB - d$SO)) | |
| S.H <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$H - d$HR, | |
| n=d$AB - d$HR - d$SO)) | |
| S <- merge(S.SO$d, S.HR$d, by="playerID") | |
| S <- merge(S, S.H$d, by="playerID") | |
| names(S) <- c("playerID", "SO", "AB", "SO.Rate", | |
| "HR", "AB.SO", "HR.Rate", | |
| "H.HR", "AB.SO.HR", "H.Rate") | |
| component.fit <- data.frame(eta=c(S.SO$eta, S.HR$eta, S.H$eta), | |
| K=c(S.SO$K, S.HR$K, S.H$K)) | |
| row.names(component.fit) <- c("SO", "HR", "H") | |
| ## ------------------------------------------------------------------------ | |
| S$Est <- with(S, | |
| (1 - SO.Rate) * (HR.Rate + (1 - HR.Rate) * H.Rate)) | |
| ## ------------------------------------------------------------------------ | |
| S2 <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$H, n=d$AB)) | |
| shrinkage.fit <- c(eta=S2$eta, K=S2$K) | |
| S <- merge(S, S2$d, by="playerID") | |
| names(S)[c(11:14)] <- c("Comp.Est", "H", "AB1", "Shrinkage.Est") | |
| S$Season <- season | |
| list(S=S, component=component.fit, shrinkage=shrinkage.fit) | |
| } | |
| plot_avg_results <- function(fitwork){ | |
| # function constructs plot of observed and predicted one.minus.strikeout.rates and his.in.non.so.ab rates | |
| # from output of fit_component_average function | |
| require(ggplot2) | |
| require(dplyr) | |
| S <- fitwork$S | |
| myf <- function(x, p) p / x | |
| AVG1 <- data.frame(P=.25, x=seq(.6, 1, .01), y=myf(seq(.6, 1, .01), .25)) | |
| AVG2 <- data.frame(P=.3, x=seq(.6, 1, .01), y=myf(seq(.6, 1, .01), .3)) | |
| AVG3 <- data.frame(P=.2, x=seq(.6, 1, .01), y=myf(seq(.6, 1, .01), .2)) | |
| AVG4 <- data.frame(P=.35, x=seq(.6, 1, .01), y=myf(seq(.6, 1, .01), .35)) | |
| AVG5 <- data.frame(P=.4, x=seq(.6, 1, .01), y=myf(seq(.6, 1, .01), .4)) | |
| AVG <- rbind(AVG1, AVG2, AVG3, AVG4, AVG5) | |
| AVG <- mutate(AVG, BA=factor(P)) | |
| d1 <- data.frame(Type="Observed BA", | |
| One.Minus.SO.Rate=1 - S$SO / S$AB, | |
| H.Rate=S$H / S$AB.SO, | |
| AB=S$AB) | |
| d2 <- data.frame(Type="Predicted BA", | |
| One.Minus.SO.Rate=1 - S$SO.Rate, | |
| H.Rate=S$HR.Rate + (1 - S$HR.Rate) * S$H.Rate, | |
| AB=S$AB) | |
| dd <- rbind(d1, d2) | |
| dd1 <- filter(dd, AB >= 300) | |
| Xlow <- range(dd1$One.Minus.SO.Rate) | |
| Ylow <- range(dd1$H.Rate) | |
| p <- ggplot(dd1, aes(One.Minus.SO.Rate, H.Rate)) + | |
| geom_point() + | |
| facet_wrap(~ Type, ncol=1) + | |
| geom_line(data=AVG, aes(x, y, color=BA)) + | |
| ylim(Ylow[1], Ylow[2]) + xlim(Xlow[1], Xlow[2]) + | |
| ggtitle(paste("Hitters with at least 300 AB in",S$Season,"Season" )) + | |
| labs(y = "Hits in Non-SO-AB Rate") | |
| print(p) | |
| } | |
| fit_component_obp <- function(season){ | |
| # implement component method for estimating group of on-base percentages for nonpitchers | |
| # for a particular baseball season | |
| # | |
| # input: season | |
| # output: a list with components | |
| # component -- gives values of K and eta for all component fits | |
| # shrinkage -- gives value of K and eta for shrinkage fit | |
| # ST -- gives raw stats and component and shrinkage estimates for all players | |
| require(Lahman) | |
| require(dplyr) | |
| require(LearnBayes) | |
| get.data <- function(season){ | |
| B <- summarize(group_by(filter(Batting, | |
| yearID==season), playerID), | |
| PA=sum(AB + BB + HBP), | |
| AB=sum(AB), BB=sum(BB + HBP), H=sum(H), | |
| SO=sum(SO), HR=sum(HR)) | |
| NonPitchers <- unique(select(filter(Fielding, | |
| yearID==season, POS!="P"), | |
| playerID)) | |
| merge(B, NonPitchers, by="playerID") | |
| } | |
| ## ------------------------------------------------------------------------ | |
| fit.model <- function(data){ | |
| mode <- laplace(betabinexch, c(1, 1), | |
| cbind(data$y, data$n))$mode | |
| eta <- exp(mode[1]) / (1 + exp(mode[1])) | |
| K <- exp(mode[2]) | |
| list(eta=eta, K=K, | |
| d=data.frame(data, est=(data$y + K * eta) / (data$n + K))) | |
| } | |
| ## ------------------------------------------------------------------------ | |
| d <- get.data(season) | |
| ## ------------------------------------------------------------------------ | |
| S.BB <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$BB, n=d$PA)) | |
| S.SO <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$SO, n=d$AB)) | |
| S.HR <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$HR, n=d$AB - d$SO)) | |
| S.H <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$H - d$HR, | |
| n=d$AB - d$HR - d$SO)) | |
| S <- merge(S.BB$d, S.SO$d, by="playerID") | |
| S <- merge(S, S.HR$d, by="playerID") | |
| names(S) <- c("playerID", "BB", "PA", "BB.Rate", | |
| "SO", "AB", "SO.Rate", | |
| "HR", "AB.SO", "HR.Rate") | |
| S <- merge(S, S.H$d, by="playerID") | |
| names(S)[11:13] <- c("HIP", "IP", "HIP.Rate") | |
| S$BAEst <- with(S, | |
| (1 - SO.Rate) * (HR.Rate + (1 - HR.Rate) * HIP.Rate)) | |
| S$Est <- with(S, | |
| BB.Rate + (1 - BB.Rate) * BAEst) | |
| component.fit <- data.frame(eta=c(S.BB$eta, S.SO$eta, | |
| S.HR$eta, S.H$eta), | |
| K=c(S.BB$K, S.SO$K, S.HR$K, S.H$K)) | |
| row.names(component.fit) <- c("BB", "SO", "HR", "H") | |
| ## ------------------------------------------------------------------------ | |
| S2 <- fit.model(data.frame(playerID=d$playerID, | |
| y=(d$H + d$BB), | |
| n=d$PA)) | |
| shrinkage.fit <- c(eta=S2$eta, K=S2$K) | |
| ST <- merge(S, S2$d, by="playerID") | |
| names(ST)[c(15, 18)] <- c("Combo.Est", "Shrinkage.Est") | |
| ST$Season <- season | |
| list(ST=ST, component=component.fit, shrinkage=shrinkage.fit) | |
| } | |
| plot_obp_results <- function(fitwork){ | |
| # function constructs plot of observed and predicted walk and hit rates | |
| # from output of fit_component_obp function | |
| require(ggplot2) | |
| require(dplyr) | |
| ST <- fitwork$ST | |
| myf <- function(x, p) (p - x) / (1 - x) | |
| AVG1 <- data.frame(P=.25, x=seq(.02, .3, .01), y=myf(seq(.02, .3, .01), .25)) | |
| AVG2 <- data.frame(P=.3, x=seq(.02, .3, .01), y=myf(seq(.02, .3, .01), .3)) | |
| AVG3 <- data.frame(P=.35, x=seq(.02, .3, .01), y=myf(seq(.02, .3, .01), .35)) | |
| AVG4 <- data.frame(P=.4, x=seq(.02, .3, .01), y=myf(seq(.02, .3, .01), .4)) | |
| AVG5 <- data.frame(P=.45, x=seq(.02, .3, .01), y=myf(seq(.02, .3, .01), .45)) | |
| AVG <- rbind(AVG1, AVG2, AVG3, AVG4, AVG5) | |
| AVG <- mutate(AVG, OBP=factor(P)) | |
| d1 <- data.frame(Type="Observed OBP", | |
| Walk.Rate=ST$BB / ST$PA, | |
| H.Rate=(ST$HIP + ST$HR) / ST$AB, | |
| PA=ST$PA) | |
| d2 <- data.frame(Type="Predicted OBP", | |
| Walk.Rate=ST$BB.Rate, | |
| H.Rate=ST$BAEst, | |
| PA=ST$PA) | |
| dd <- rbind(d1, d2) | |
| dd1 <- filter(dd, PA >= 300) | |
| Xlim <- range(dd1$Walk.Rate); Ylim <- range(dd1$H.Rate) | |
| p <- ggplot(dd1, aes(Walk.Rate, H.Rate)) + | |
| geom_point() + | |
| facet_wrap(~ Type, ncol=1) + | |
| geom_line(data=AVG, aes(x, y, color=OBP)) + | |
| ylim(Ylim[1], Ylim[2]) + xlim(Xlim[1], Xlim[2]) + | |
| ggtitle(paste("Hitters with at least 300 PA in",ST$Season,"Season")) + | |
| labs(x = "Walk Rate", y = "Hit Rate") | |
| print(p) | |
| } | |
| fit_component_fip <- function(season){ | |
| # implement component method for estimating group of FIPs for pitchers in a particular season | |
| # | |
| # input: season | |
| # output: | |
| # S -- gives raw stats and component and shrinkage estimates for all pitches who face at least 50 batters | |
| require(Lahman) | |
| require(dplyr) | |
| require(LearnBayes) | |
| get.data <- function(season){ | |
| summarize(group_by(filter(Pitching, yearID==season), playerID), | |
| PA=sum(BFP), | |
| AB=sum(BFP - BB - HBP, na.rm=TRUE), | |
| SO=sum(SO, na.rm=TRUE), | |
| BB=sum(BB + HBP, na.rm=TRUE), | |
| HR=sum(HR), | |
| H=sum(H), | |
| IP=sum(IPouts) / 3) | |
| } | |
| fit.model <- function(data){ | |
| mode <- laplace(betabinexch, c(1, 1), | |
| cbind(data$y, data$n))$mode | |
| eta <- exp(mode[1]) / (1 + exp(mode[1])) | |
| K <- exp(mode[2]) | |
| list(eta=eta, K=K, | |
| d=data.frame(data, est=(data$y + K * eta) / (data$n + K))) | |
| } | |
| d <- filter(get.data(season), PA >= 50) | |
| S.BB <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$BB, n=d$PA)) | |
| names(S.BB$d) <- c("playerID", "BB", "PA", "BB.Rate") | |
| S.SO <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$SO, n=d$AB)) | |
| names(S.SO$d) <- c("playerID", "SO", "AB", "SO.Rate") | |
| S.HR <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$HR, n=d$AB - d$SO)) | |
| names(S.HR$d) <- c("playerID", "HR", "AB.minus.SO", "HR.Rate") | |
| S.H <- fit.model(data.frame(playerID=d$playerID, | |
| y=d$H - d$HR, | |
| n=d$AB - d$SO - d$HR)) | |
| names(S.H$d) <- c("playerID", "HIP", "OIP", "HIP.Rate") | |
| S <- merge(S.BB$d, S.SO$d, by="playerID") | |
| S <- merge(S, S.HR$d, by="playerID") | |
| S <- merge(S, S.H$d, by="playerID") | |
| S$Est <- with(S, | |
| (39 * (1 - BB.Rate) * (1 - SO.Rate) * HR.Rate + | |
| 9 * BB.Rate - 6 * (1 - BB.Rate) * SO.Rate) / | |
| ((1 - BB.Rate) * (SO.Rate + | |
| (1 - SO.Rate) * (1 - HR.Rate) * (1 - HIP.Rate)))) | |
| S$Obs <- with(S, | |
| ((13 * HR) + (3 * BB) - (2 * SO))) / d$IP | |
| vary <- 5.1 ^ 2 / d$IP | |
| dd <- cbind(S$Obs, vary) | |
| dd <- dd[complete.cases(dd), ] | |
| fit <- laplace(normnormexch, c(0, 0), dd) | |
| mn <- fit$mode[1] | |
| tau <- exp(fit$mode[2]) | |
| S$shrink <- (S$Obs / vary + mn / tau ^ 2) / | |
| (1 / vary + 1 / tau ^ 2) | |
| S | |
| } | |
| plot_fip_results <- function(S){ | |
| # function constructs plot of observed and predicted FIP measures | |
| # from output of fit_component_fip function | |
| library(ggplot2) | |
| d1 <- data.frame(PA=S$PA, FIP=S$Obs, Type="Observed") | |
| d2 <- data.frame(PA=S$PA, FIP=S$Est, Type="Estimated") | |
| d3 <- data.frame(PA=S$PA, FIP=S$shrink, Type="Shrinkage") | |
| d <- rbind(d1, d2, d3) | |
| p1 <- ggplot(filter(d, Type == "Observed" | Type == "Shrinkage"), | |
| aes(PA, FIP, color=Type)) + geom_point() | |
| p2 <- ggplot(filter(d, Type == "Estimated" | Type == "Shrinkage"), | |
| aes(PA, FIP, color=Type)) + geom_point() | |
| print(p1) | |
| print(p2) | |
| } | |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment