Illustrates the use of a GAM fit to predict home runs and check the predictions on a new dataset

home_run_train_test.R

# load several packages

library(tidyverse)
library(CalledStrike)
library(mgcv)

# read in Statcast data for five seasons

sc <- read_csv("five_seasons_data.csv")

# create month variables

sc %>% mutate(month = substr(game_date, 6, 7)) -> sc

sc %>%
  mutate(month = ifelse(month == "03", "04", month),
         month = ifelse(month == "10", "09", month),
         season = as.character(Season))  %>%
  mutate(Month = as.numeric(month)) -> sc

## write a general function to do fitting and testing stages

fit_predict <- function(data1, data2, the_title = ""){
  fit <- gam(HR ~ s(launch_angle, launch_speed),
             family = binomial,
             data = data1)
  sc_new <- data2
  actual <- sum(sc_new$events == "home_run", na.rm = TRUE)
  prob <- predict(fit, sc_new, type = "response")
  one_prediction <- function(Prob){
    sum(runif(length(prob)) < Prob)
  }
  many_predictions <- replicate(1000, one_prediction(prob))

  ggplot(data = data.frame(Prediction =
                             many_predictions),
         aes(Prediction)) +
    geom_histogram(color = "white",
                   fill = "tan",
                   bins = 15) +
    increasefont() +
    geom_vline(xintercept = actual, size = 2,
               color = "red") +
    ggtitle(the_title) +
    centertitle()
}

# break 2018 data into two random parts

sc2018 <- filter(sc, season == 2018,
                 launch_angle > 10,
                 launch_speed > 80)

N <- nrow(sc2018)
ind <- sample(N, size = N / 2)
data1 <- sc2018[ind, ]
data2 <- sc2018[-ind, ]

fit_predict(data1, data2,
            "Fit on Half of 2018, Predict on Other Half")

# modify second data part
# do random split but change LS and LA in test set
# 0.3 increase in LS, 0.4 increase in LA in 2019

sc2018 <- filter(sc, season == 2018,
                 launch_angle > 10,
                 launch_speed > 80)

N <- nrow(sc2018)
ind <- sample(N, size = N / 2)
data1 <- sc2018[ind, ]
data2 <- sc2018[-ind, ]

data2$launch_angle <- data2$launch_angle + 0.4
data2$launch_speed <- data2$launch_speed + 0.3

fit_predict(data1, data2,
            "Fit on Half of 2018, Predict on Adjusted Other Half")

# summaries for launch angles and launch speeds for 2018 and 2019 seasons

sc2018 <- filter(sc, season == 2018,
                 launch_angle > 10,
                 launch_speed > 80)

sc2019 <- filter(sc, season == 2019,
                 launch_angle > 10,
                 launch_speed > 80)

sc2018 %>%
  summarize(N = n(),
            LS = mean(launch_speed, na.rm = TRUE),
            LS_sd = sd(launch_speed, na.rm = TRUE),
            LA = mean(launch_angle, na.rm = TRUE),
            LA_sd = sd(launch_angle, na.rm = TRUE)) ->
  S18

sc2019 %>%
  summarize(N = n(), LS = mean(launch_speed, na.rm = TRUE),
            LS_sd = sd(launch_speed, na.rm = TRUE),
            LA = mean(launch_angle, na.rm = TRUE),
            LA_sd = sd(launch_angle, na.rm = TRUE)) ->
  S19

S18
S19