[BUG] <title> EP on kickoff events is outdated (for 2024 and 2025)

[TheMathNinja]

Is there an existing issue for this?

• I have searched the existing issues

Have you installed the latest development version of the package(s) in question?

• I have installed the latest development version of the package.

If this is a data issue, have you tried clearing your nflverse cache?

I have cleared my nflverse cache and the issue persists.

What version of the package do you have?

[1] ‘5.1.0.9008’

Describe the bug

Kickoff EP is correctly calibrated up through 2023 (such that avg EPA on kickoffs is ~0), but there was no re-calibration for the new kickoff rules in 2024-2025, resulting in average kickoff EPA ~0.252 in those seasons.

Reprex

kickoff_ep_epa_by_season <- nflreadr::load_pbp(2021:2025) %>%
  filter(
    play_type == "kickoff",
    !is.na(ep),
    !is.na(epa)
  ) %>%
  group_by(season) %>%
  summarise(
    n        = n(),
    mean_ep  = round(mean(ep), 4),
    mean_epa = round(mean(epa), 4),
    .groups  = "drop"
  ) %>%
  arrange(desc(season))

> kickoff_ep_epa_by_season
# A tibble: 5 × 4
  season     n mean_ep mean_epa
   <int> <int>   <dbl>    <dbl>
1   2025  2636    1.09   0.253 
2   2024  2629    1.09   0.251 
3   2023  2561    1.09  -0.0026
4   2022  2544    1.09   0.0009
5   2021  2602    1.08  -0.0306

Expected Behavior

I expect EPA to be ~0 for all years (and therefore EP in 2024 and 2025 to be ~1.34). There' aslo an argument for 2024 and 2025 to be calibrated separately due to the touchback rule change between the two years, though it's scales of magnitute smaller than the 2023-24 change.

It's worth considering a total recalibration of EPA considering this also reduces the value of scoring events relative to non-scoring events that result in change of possession (e.g. turnovers). In other words, if we previously considered a TD worth 6 + 0.95 - 1.09 = 5.86 net points, we should adjust it to 6 + 0.95 - 1.34 = 5.61 net points.

nflverse_sitrep

> nflverse_sitrep()
── System Info ────────────────────────────────────────────────────────────────────────────────────
• R version 4.5.2 (2025-10-31 ucrt) • Running under: Windows 11 x64 (build 26100)
── Package Status ─────────────────────────────────────────────────────────────────────────────────
   package  installed  cran        dev behind
1   nfl4th 1.0.4.9007 1.0.4 1.0.4.9007       
2 nflfastR 5.1.0.9008 5.1.0 5.1.0.9008       
3 nflplotR      1.6.0 1.6.0      1.6.0       
4 nflreadr 1.5.0.9001 1.5.0 1.5.0.9001       
5 nflseedR      2.0.2 2.0.2      2.0.2       
6 nflverse 1.0.3.9001 1.0.3 1.0.3.9001       
── Package Options ────────────────────────────────────────────────────────────────────────────────
• No options set for above packages
── Package Dependencies ───────────────────────────────────────────────────────────────────────────
• askpass     (1.2.1)       • httr         (1.4.7)   • scales      (1.4.0)    
• backports   (1.5.0)       • isoband      (0.3.0)   • snakecase   (0.11.1)   
• base64enc   (0.1-3)       • janitor      (2.2.1)   • stringi     (1.8.7)    
• bigD        (0.3.1)       • jquerylib    (0.1.4)   • stringr     (1.6.0)    
• bitops      (1.0-9)       • jsonlite     (2.0.0)   • sys         (3.4.3)    
• bslib       (0.9.0)       • juicyjuice   (0.1.0)   • tibble      (3.3.0)    
• cachem      (1.1.0)       • knitr        (1.51)    • tidyr       (1.3.2)    
• cli         (3.6.5)       • labeling     (0.4.3)   • tidyselect  (1.2.1)    
• commonmark  (2.0.0)       • lifecycle    (1.0.5)   • timechange  (0.3.0)    
• cpp11       (0.5.2)       • listenv      (0.10.0)  • tinytex     (0.58)     
• crayon      (1.5.3)       • litedown     (0.9)     • utf8        (1.2.6)    
• curl        (7.0.0)       • lubridate    (1.9.4)   • V8          (8.0.1)    
• data.table  (1.18.0)      • magick       (2.9.0)   • vctrs       (0.6.5)    
• digest      (0.6.39)      • magrittr     (2.0.4)   • viridisLite (0.4.2)    
• dplyr       (1.1.4)       • markdown     (2.0)     • withr       (3.0.2)    
• evaluate    (1.0.5)       • memoise      (2.0.1)   • xfun        (0.55)     
• farver      (2.1.2)       • mime         (0.13)    • xgboost     (3.1.2.1)  
• fastmap     (1.2.0)       • openssl      (2.3.4)   • xml2        (1.5.1)    
• fastrmodels (2.0.0.9000)  • parallelly   (1.45.0)  • yaml        (2.3.12)   
• fontawesome (0.5.3)       • pillar       (1.11.1)  • codetools   (0.2-20)   
• fs          (1.6.6)       • pkgconfig    (2.0.3)   • compiler    (4.5.2)    
• furrr       (0.3.1)       • progressr    (0.18.0)  • graphics    (4.5.2)    
• future      (1.68.0)      • proto        (1.0.0)   • grDevices   (4.5.2)    
• generics    (0.1.4)       • purrr        (1.0.4)   • grid        (4.5.2)    
• ggpath      (1.1.1)       • R6           (2.6.1)   • lattice     (0.22-7)   
• ggplot2     (4.0.1)       • rappdirs     (0.3.3)   • Matrix      (1.7-4)    
• globals     (0.18.0)      • RColorBrewer (1.1-3)   • methods     (4.5.2)    
• glue        (1.8.0)       • Rcpp         (1.1.0)   • mgcv        (1.9-3)    
• gsubfn      (0.7)         • reactable    (0.4.5)   • nlme        (3.1-168)  
• gt          (1.2.0)       • reactR       (0.6.1)   • parallel    (4.5.2)    
• gtable      (0.3.6)       • rlang        (1.1.6)   • splines     (4.5.2)    
• highr       (0.11)        • rmarkdown    (2.30)    • stats       (4.5.2)    
• hms         (1.1.4)       • rstudioapi   (0.17.1)  • tools       (4.5.2)    
• htmltools   (0.5.9)       • S7           (0.2.1)   • utils       (4.5.2)    
• htmlwidgets (1.6.4)       • sass         (0.4.10)    
───────────────────────────────────────────────────────────────────────────────────────────────────

Screenshots

Additional context

No response

[TheMathNinja]

While I’m not sure if this would totally nail it, I believe if the benchmark for EP on kickoffs moved from “a zero-second return and ball placed at the 25” from 2023 and earlier to “a zero second return and ball placed at the 30” in 2024 and possibly also 2025, that would be a good quick fix that gets us most of the way there.

In 2025 the touchback got pushed from the 30 to 35 but therefore no longer became a useful norm for the kickoff expectation, as it became a minority experience and teams generally took it to the 31.5 yard line. I’m not sure whether using the 30 or 31 would be better given the impact of turnovers. I suppose we also don’t have to hard-code EP on kickoffs like this and could just let the model decide as it does on all the other plays.