First load the packages and some data. load_4th_pbp() loads nflfastR data and computes 4th down probabilities (depending on your computer, this may take up to a minute or two per season).

library(nfl4th)
library(tidyverse)
library(gt)

#> [15:04:09] WARNING: amalgamation/../src/learner.cc:438:
#>   If you are loading a serialized model (like pickle in Python, RDS in R) generated by
#>   older XGBoost, please export the model by calling Booster.save_model from that version
#>   first, then load it back in current version. See:
#>
#>
#>   for more details about differences between saving model and serializing.
#>
#> [15:04:09] WARNING: amalgamation/../src/learner.cc:438:
#>   If you are loading a serialized model (like pickle in Python, RDS in R) generated by
#>   older XGBoost, please export the model by calling Booster.save_model from that version
#>   first, then load it back in current version. See:
#>
#>
#>   for more details about differences between saving model and serializing.

## Easy mode: using nflfastR data

Here’s what the data obtained using load_4th_pbp() looks like:

data %>%
dplyr::filter(!is.na(go_boost)) %>%
dplyr::select(
posteam, ydstogo, yardline_100, posteam, go_boost, first_down_prob,
wp_fail, wp_succeed, go_wp, fg_make_prob, miss_fg_wp, make_fg_wp,
fg_wp, punt_wp
) %>%
knitr::kable(digits = 2)
posteam ydstogo yardline_100 go_boost first_down_prob wp_fail wp_succeed go_wp fg_make_prob miss_fg_wp make_fg_wp fg_wp punt_wp
SF 3 34 1.40 0.55 0.69 0.78 0.74 0.62 0.69 0.75 0.73 0.72
ARI 10 65 -2.49 0.27 0.18 0.28 0.21 0.00 0.18 0.29 0.18 0.24
ARI 7 72 -0.27 0.38 0.07 0.14 0.10 0.00 0.07 0.15 0.07 0.10
SF 5 64 0.78 0.48 0.84 0.92 0.88 0.00 0.83 0.92 0.83 0.87
SF 3 68 1.71 0.55 0.65 0.81 0.74 0.00 0.62 0.80 0.62 0.72
ARI 9 77 -1.06 0.30 0.16 0.28 0.19 0.00 0.15 0.28 0.15 0.20
SF 1 1 3.22 0.63 0.78 0.89 0.85 0.99 0.76 0.82 0.82 NA
ARI 5 34 0.54 0.45 0.19 0.37 0.27 0.62 0.19 0.31 0.27 0.23
SF 9 36 -0.12 0.32 0.72 0.84 0.76 0.57 0.71 0.80 0.76 0.76
SF 2 6 1.88 0.54 0.77 0.89 0.83 0.98 0.75 0.81 0.81 NA

Or we can add some filters to look up a certain game:

data %>%
dplyr::filter(week == 20, posteam == "GB", down == 4) %>%
dplyr::select(
posteam, ydstogo, yardline_100, posteam, go_boost, first_down_prob,
wp_fail, wp_succeed, go_wp, fg_make_prob, miss_fg_wp, make_fg_wp,
fg_wp, punt_wp
) %>%
knitr::kable(digits = 2)
posteam ydstogo yardline_100 go_boost first_down_prob wp_fail wp_succeed go_wp fg_make_prob miss_fg_wp make_fg_wp fg_wp punt_wp
GB 17 65 -3.74 0.15 0.33 0.47 0.35 0.00 0.31 0.47 0.31 0.38
GB 6 6 -2.16 0.31 0.36 0.57 0.43 0.98 0.33 0.45 0.45 NA
GB 15 86 -2.52 0.19 0.16 0.38 0.20 0.00 0.14 0.34 0.14 0.22
GB 10 76 -0.37 0.32 0.15 0.37 0.22 0.00 0.14 0.32 0.14 0.23
GB 8 8 3.77 0.33 0.04 0.31 0.13 0.98 0.03 0.09 0.09 NA

We see the infamous field goal at the bottom.

## Calculations from user input

The below shows the bare minimum amount of information that has to be fed to nfl4th in order to compute 4th down decision recommendations. The main function on user-input data is add_4th_probs().

The reason teams from a specific game have to be used is that the model depends on factors such as point spread, team totals, and indoor/outdoor and the program automatically looks these up so that users don’t have to provide them.

one_play <- tibble::tibble(

# things to help find the right game (use "reg" or "post" for type)
home_team = "GB",
away_team = "TB",
posteam = "GB",
type = "post",
season = 2020,

# information about the situation
qtr = 4,
quarter_seconds_remaining = 129,
ydstogo = 8,
yardline_100 = 8,
score_differential = -8,

home_opening_kickoff = 0,
posteam_timeouts_remaining = 3,
defteam_timeouts_remaining = 3
)

one_play %>%
dplyr::select(
posteam, ydstogo, yardline_100, posteam, go_boost, first_down_prob,
wp_fail, wp_succeed, go_wp, fg_make_prob, miss_fg_wp, make_fg_wp,
fg_wp, punt_wp
) %>%
knitr::kable(digits = 2)
#> Computing probabilities for 1 plays. . .
posteam ydstogo yardline_100 go_boost first_down_prob wp_fail wp_succeed go_wp fg_make_prob miss_fg_wp make_fg_wp fg_wp punt_wp
GB 8 8 3.77 0.33 0.04 0.31 0.13 0.98 0.03 0.09 0.09 NA

Comparing this and the table above, we see the exact same numbers as expected.

## Make a summary table

Let’s put the play above into a table using the provided function make_table_data(), which makes it easier to interpret the recommendations for a play. This function only works with one play at a time since it makes a table using the results from the play.

one_play %>%
nfl4th::make_table_data() %>%
knitr::kable(digits = 1)
#> Computing probabilities for 1 plays. . .
choice choice_prob success_prob fail_wp success_wp
Go for it 12.7 32.7 3.5 31.5
Field goal attempt 8.9 97.5 3.0 9.0
Punt NA NA NA NA

Looking at the table, the Packers would be expected to have 12.7% win probability if they had gone for it and 8.9% if they kicked a field goal. This difference of 3.8 percentage points is almost exactly the same as PFF’s 3.5 percentage points for the decision.

## Make a summary table for a 2-point decision

nfl4th also contains a function to calculate 2-point decisions. Let’s put in the situation that would have happened if the Packers had scored a touchdown on the 4th & 8. We don’t need a calculator to know that they should have gone for two, but let’s practice by putting in the numbers, assuming that the 4th down play took 6 seconds while resulting in a touchdown.

another_play <- tibble::tibble(

# things to help find the right game (use "reg" or "post")
home_team = "GB",
away_team = "TB",
posteam = "GB",
type = "post",
season = 2020,

# information about the situation
qtr = 4,
quarter_seconds_remaining = 123,
score_differential = -2,

home_opening_kickoff = 0,
posteam_timeouts_remaining = 3,
defteam_timeouts_remaining = 3
)

another_play %>%
nfl4th::make_2pt_table_data() %>%
knitr::kable(digits = 1)
#> Computing probabilities for  1 plays. . .
choice choice_prob success_prob fail_wp success_wp
Go for 2 31.5 58.9 18.4 40.6
Kick XP 26.3 93.3 18.4 26.9

Note that the go for 2 probability here is identical to the win probability associated with a successful 4th down conversion above because the 4th down model assumes that the Packers would go for 2 if they scored.

## Getting 4th down plays from a live game

nflfastR isn’t available for live games and typing all the plays in by hand is annoying. So how does the 4th down bot work? With thanks to the ESPN API, which can be accessed using get_4th_plays().

plays <- get_4th_plays("2020_20_TB_GB") %>%
tail(1)

plays %>%
select(desc, quarter_seconds_remaining)
#> # A tibble: 1 × 2
#>   desc                             quarter_seconds_remaining
#>   <chr>                                                <dbl>
#> 1 "Mason Crosby 26 Yd Field Goal "                       125

plays %>%
#> Computing probabilities for 1 plays. . .