R script for Using Statcast to Measure HItters post

statcast_measure.R

# load in tidyverse package and
# load in theme for title

library(tidyverse)
TH <- theme(plot.title = element_text(colour = "blue", 
                   size = 18, 
      hjust = 0.5, vjust = 0.8, angle = 0))

# read in the 2017 statcast data

sc2017 <- read_csv("statcast2017.csv")

# only look at balls in play and define the hit variable

sc2017 %>% filter(type == "X") %>%
  mutate(hit = ifelse(events %in% 
      c("single", "double", "triple", "home_run"), 
          1, 0)) ->  sc2017_ip

# fit gam with binomial distribution (logistic link)

library(mgcv)
fit <- gam(hit ~ s(launch_speed, launch_angle),
           data = sc2017_ip, family = binomial)

# for a particular player, computes
# number of batted balls, number of hits, and number
# of predicted hits from model

one_player <- function(id){
  invlogit <- function(x){exp(x) / (1 + exp(x))}
  sc2017_ip %>% filter(batter == id) %>%
  select(player_name, events, launch_speed,
         launch_angle, hit) -> d
  d$Predict <-  invlogit(predict(fit, d))
  c(length(d$hit), sum(d$hit), sum(d$Predict))
}

# does this work for all batters
# computes Z scores to contrast observed and expected

unique_ids <- unique(sc2017_ip$batter)
S <- sapply(unique_ids, one_player)
S1 <- data.frame(Batter = unique_ids,
                 N = S[1, ],
                 H = S[2, ],
                 Expected = S[3, ])
sc2017_ip %>% group_by(batter) %>% 
  summarize(Player = first(player_name)) -> SC
inner_join(S1, SC,
           by=c("Batter" = "batter")) -> S2
S2$Z = with(S2, (H - Expected) / sqrt(Expected))

# three graphs comparing observed and expected hits

library(ggrepel)
ggplot(S2, aes(Expected, H, label=Player)) +
  geom_point() + geom_smooth(color="red") + TH +
  ggtitle("Observed and Expected Hits for All Players")

ggplot(S2, aes(N, H - Expected, label=Player)) +
  geom_point() +
  geom_hline(yintercept = 0, color="red") + TH +
  ggtitle("Graph of Residuals")

ggplot(S2, aes(N, Z, label=Player)) +
  geom_point() +
  geom_label_repel(data = filter(S2, Z > 2)) +
  geom_label_repel(data = filter(S2, Z < -2)) +
  ylim(-2.5, 4) +
  geom_hline(yintercept = 0, color="red") + TH +
  ggtitle("Graph of Standardized Scores")

# look at a player's hits more carefully

one_player_graph <- function(pname){
  invlogit <- function(x){exp(x) / (1 + exp(x))}
  TH <- theme(plot.title = element_text(colour = "blue", 
                                        size = 18, 
                  hjust = 0.5, vjust = 0.8, angle = 0))
  filter(sc2017_ip, player_name == pname) %>%
     select(launch_angle, launch_speed, hit) -> dg

  dg$P <- invlogit(predict(fit, dg))
  dg$Predict <- ifelse(dg$P > .5, 1, 0)

  df1 <- select(dg, launch_angle, launch_speed)
  df1$Outcome = dg$hit; df1$Type = "Actual"
  df2 <- select(dg, launch_angle, launch_speed)
  df2$Outcome = dg$Predict; df2$Type = "Predicted"
  df <- rbind(df1, df2)
  df$Outcome <- as.factor(df$Outcome)
  print(ggplot(df, aes(launch_angle, launch_speed, 
               color=Outcome)) +
    geom_jitter() + 
    facet_wrap(~ Type, ncol=1) +
    ggtitle(pname) + TH)
  dg
}

# focus on Dee Gordon and Miguel Cabrera

dg <- one_player_graph("Dee Gordon")
mc <- one_player_graph("Miguel Cabrera")

# compute actual and predicted number of hits for
# each player 

summarize(dg, H = sum(hit), P = sum(P))
summarize(mc, H = sum(hit), P = sum(P))