NEWS

nfl4th 1.0.7.9001

Field goal make probability is now corrected for selection bias in long attempts. Only teams confident in their kicker try them, so the raw model curve was too generous past about 52 yards; probability from 55 to 61 yards drops by 5 to 10 percentage points. This replaces the linear decay that was anchored on the predicted 58 yard probability. (#63)
The field goal model's era cutoff moves from 2020 to 2023, and the definition is now shared between the prediction and model-fitting code so the two cannot drift. (#63)
From 2024 on, kickoffs after scores and to start the second half put the receiving team at its own 30 instead of its own 25, to match the new kickoff rules. (#61)
Fix the punt model's return touchdown and blocked punt rates, which were swapped. (#60)

The internal cache is now aware of changed model formats and forces a model update for compatibility with xgboost (>= v3). (#53)

Thank you to @isaactpetersen, and @mrcaseb for their questions, feedback, and contributions towards this release.

Stop letting roof break things
New location for pre-computed numbers
Re-work field goal model. Takes into account improvements in accuracy (especially on long kicks) and gives probability > 0 on kicks of over 70 yards
Update models and internals to be compatible with newer xgboost (>= v3) versions. (#41)
nfl4th now requires R 4.1 to allow the package to use R's native pipe |> operator. This follows the Tidyverse R version support rules. (#43)

Create package cache directory with tools::R_user_dir() because CRAN doesn't like rappdirs::user_cache_dir()

Re-export xgboost models to get rid of annoying warning message
Tweak how close to end of game it will calculate probabilities
Required data isn't loaded directly with the package. Instead the download is triggered when necessary and cached in a local package cache. The cache can be cleared with the new function nfl4th_clear_cache(). The package resets the cache partly when loaded. This can be prevented with options(nfl4th.keep_games = TRUE).
Update win probability model. Instead of only using nflfastR, it stacks nflfastR with another model
Fixup cache paths.
Remove tidyverse from Suggests

Fix for aborted plays on punts being called going for it
Re-categorized some plays as unknown (i.e., NA) go: False Start or defensive encroachment along with being lined up to go for it (run formation or pass formation)
Added fast argument to load_4th_pbp() which allows for loading pre-computed go_boost rather than needing to calculate it

Release as package
Fixes with touchdowns. Instead of granting 7 points, assumes teams choose best option between PAT or 2pt and give pre-conversion attempt WP accordingly
Fix for punt WP at the end of first half being too high
Add logic for TD/FG decision on the last play of the first half
Add possibility for muffed punt
Reduce field goal chances on very long field goals