Skip to contents

Hey everyone, my name is Michael and over the summer I worked on a daily series of plots using ggplot and the cfbfastR package. One of my favorite plots I put together was the fourth down tendency plot for various head coaches. This visualization was inspired by Michael Lopez doing the same thing for NFL coaches. This tutorial is going to walk through how they’re put together. If you haven’t already, you should read the introduction tutorial that Parker made to get used to the data and download the package.

First, we’ll have to install and import the necessary packages

if (!requireNamespace('pak', quietly = TRUE)){
  install.packages('pak')
}
pak::pak(c("tidyverse", "sportyR"))
library(tidyverse)
library(sportyR)
# cfbfastR is deliberately NOT in the pak() list above. pkgdown and
# R CMD check render this vignette against the package being built, and
# installing from CRAN here would overwrite that dev build with the last
# release -- so any function added since it would vanish mid-render.
library(cfbfastR)

We are going to load in data for seasons 2014-2020, it’ll take between 45-90 seconds to run.

tictoc::tic()
pbp <- data.frame()
seasons <- 2014:2020
progressr::with_progress({

  pbp <- cfbfastR::load_cfb_pbp(seasons)
})
tictoc::toc()
## 39.269 sec elapsed

Next, we’ll need to get the coaching information, so we’ll use the cfbd_coaches function:

coaches <- purrr::map_dfr(seasons,function(x){cfbfastR::cfbd_coaches(year = x)})

cfbd_coaches() already returns each coach’s conference, so no extra lookup is needed.

coaches <- coaches |>
  dplyr::mutate(coach = paste(first_name, last_name, sep = " ")) |>
  dplyr::filter(games >= 6) |>
  dplyr::select(coach, school, year, conference)

Since interim coaches are included in the coaches dataframe, we’ll set the cutoff at coaching 6 or more games.

We only have a couple more steps to make our graph. Next, we need to add our coaches to the pbp dataframe using join functions

pbp <- pbp |>
  dplyr::inner_join(coaches, by = c("offense_play" = "school", "year" = "year"))
## Warning in dplyr::inner_join(pbp, coaches, by = c(offense_play = "school", : Detected an unexpected many-to-many relationship between `x` and `y`.
## ℹ Row 160134 of `x` matches multiple rows in `y`.
## ℹ Row 101 of `y` matches multiple rows in `x`.
## ℹ If a many-to-many relationship is expected, set `relationship =
##   "many-to-many"` to silence this warning.

Now we can filter down to only fourth down plays, then we’ll add columns to determine if the play was a punt, FGA, or the team went for it.

down4 <- pbp |>
  dplyr::filter(down == 4) |>
  dplyr::mutate(fga = ifelse(str_detect(play_type, "Field Goal"),
                             1, 0),
                punt = ifelse(play_type == "Punt", 1, 0),
                attempt = ifelse(rush == 1 | pass == 1, 1, 0),
                play = dplyr::case_when(fga == 1 ~ "FG Attempt",
                                 punt == 1 ~ "Punt",
                                 attempt == 1 ~ "Go"))

And now we’ve got all the info we need to make the graph! The code below is listed for current Big 12 head coaches, but you can alter it to show whichever coaches you want.

Rather than a blank panel, we’ll draw each decision on a regulation college field with sportyR’s geom_football("ncaa"). sportyR centres the field on midfield and measures it in yards, so the offense’s own goal line sits at x = -50 and the opponent’s at x = 50: a line of scrimmage yards_to_goal yards from the end zone is at x = 50 - yards_to_goal. The field’s width carries no information here, so the points are jittered across it.

big_12_coaches <- c("Matt Campbell", "Steve Sarkisian", "Lincoln Riley", "Chris Klieman", "Matt Wells",
                    "Neal Brown", "Les Miles", "Mike Gundy", "Gary Patterson", "Dave Aranda")

big_12_down4 <- down4 |>
  dplyr::filter(!is.na(play)) |>
  dplyr::filter(coach %in% big_12_coaches) |>
  dplyr::filter(distance <= 5, distance > 0)

geom_football("ncaa") +
  geom_jitter(data = big_12_down4,
              aes(x = 50 - yards_to_goal, y = 0, color = play),
              width = 0, height = 22, size = 1.5) +
  facet_wrap(. ~ coach, ncol = 4) +
  scale_color_manual(values = c("Go" = "#ffd23f", "FG Attempt" = "#3bceac", "Punt" = "#ee4266")) +
  labs(title = "Big 12 Coaches's Fourth Down Tendencies | CFP Era",
       subtitle = "Fourth and 5 or shorter, at the line of scrimmage. Offense moving left to right.",
       caption = "Data from @cfbfastR | Field from sportyR | Visualization by Michael Egle (@deceptivespeed_)",
       color = "Decision") +
  theme(legend.position = "top")

Ten small college football fields, one per Big 12 head coach, each showing that coach's fourth-and-five-or-shorter plays from 2014 to 2020 as points at the line of scrimmage, coloured by whether the team went for it, kicked a field goal or punted, with the offense moving left to right.

How makeable were those fourth downs?

cfbfastR ships a port of cfb4th’s fourth-down model. calculate_fourth_down() scores the chance the offense gains the yards it needs (fd_conversion_prob) from the down, distance, field position and the pre-game line. The line comes from each game’s spread (the home team’s line, negative when the home team is favoured) and over_under, turned around to the offense’s point of view:

big_12_down4 <- big_12_down4 |>
  dplyr::mutate(
    is_home = pos_team == home,
    posteam_spread = ifelse(is_home, -spread, spread),
    posteam_total = ifelse(is_home, (over_under - spread) / 2, (over_under + spread) / 2)
  ) |>
  cfbfastR::calculate_fourth_down()

On the fourth downs the model rates as better than a coin flip to convert, how often did each coach keep the offense on the field?

big_12_down4 |>
  dplyr::filter(fd_conversion_prob > 0.5) |>
  dplyr::group_by(coach) |>
  dplyr::summarise(plays = dplyr::n(),
                   go_rate = mean(play == "Go"),
                   mean_conversion_prob = mean(fd_conversion_prob)) |>
  dplyr::arrange(dplyr::desc(go_rate))
## # A tibble: 10 × 4
##    coach           plays go_rate mean_conversion_prob
##    <chr>           <int>   <dbl>                <dbl>
##  1 Dave Aranda        16   0.5                  0.655
##  2 Steve Sarkisian    37   0.486                0.659
##  3 Lincoln Riley     114   0.421                0.624
##  4 Chris Klieman      40   0.4                  0.637
##  5 Matt Campbell     215   0.4                  0.628
##  6 Neal Brown        175   0.394                0.625
##  7 Les Miles          96   0.375                0.633
##  8 Matt Wells        239   0.351                0.626
##  9 Gary Patterson    222   0.320                0.647
## 10 Mike Gundy        230   0.317                0.640

Conversion probability is only half the decision: going for it should also be weighed against what a field goal or a punt is worth. cfb4th does that comparison in win probability, and is installed from the SportsDataverse r-universe:

install.packages("cfb4th", repos = c("https://sportsdataverse.r-universe.dev", getOption("repos")))

# go / field goal / punt win probabilities and the recommended call for every
# fourth down of a season. It joins CFBD betting lines, so it needs a
# CFBD_API_KEY; the models download and cache on first use.
fourth_downs <- cfb4th::load_4th_pbp(2020)
fourth_downs |>
  dplyr::select(pos_team, def_pos_team, distance, yards_to_goal, go_boost, go, play_type)

Looks good! Small sample size but can easily be built upon. Hopefully you found this tutorial helpful and can make some more cool CFB related visualizations with ggplot2

Our Authors

Citation

To cite the cfbfastR R package in publications, use:

BibTeX Citation

@misc{cfbfastr,
  author = {Saiem Gilani and Akshay Easwaran and Jared Lee and Eric Hess},
  title = {cfbfastR: Access College Football Play by Play Data},
  url = {https://cfbfastR.sportsdataverse.org/},
  year = {2021}
}