fix wpa on end game line, fix pat wp and fantasy id decoding #230
Conversation
|
This doesn't pass Seb's check here but does pass |
|
Note that |
(because of era and roof adjustments)
|
Finally fixed the underlying issue of #183. The actual problem lives here https://github.com/mrcaseb/nflfastR/blob/66d0de88e9093e367cc44a2268191bfb5df334eb/R/helper_add_ep_wp.R#L649 We estimate the probability that the PAT will be made and instead of using the output vector of the bam we only use the first value. Since the fg_model uses model_roof and era as features, the pat_make_probs differ slightly (which is the reason why the overall differences are so small). Line 649 uses the first pat_make_prob of the complete dataset for all games. So if we use two datasets where the first pat_make_prob will differ we will get differing wp all across the complete dataset, which is bad imo. Therefore I deleted the part that uses the first prob and instead keep the complete pat_make_prob vector. @guga31bb please let me know if there was a good reason to do The following code shows that this fixes all our wp difference problems as all the wrong wpa values were only a result of wrong wp values. It also shows the differing pat_make_probs depending on which game is the first game in the dataset. For 2020 games only this results in a maximum difference of library(dplyr, warn.conflicts = FALSE)
library(arsenal)
library(nflfastR)
ids <- c("2020_01_CHI_DET", "2020_15_TB_ATL", "2020_17_TEN_HOU")
# v4.1.0
all_games <- build_nflfastR_pbp(ids)
#> -- Build nflfastR Play-by-Play Data ------------------ nflfastR version 4.1.0 --
#> * 11:12:48 | Start download of 3 games...
sample <- build_nflfastR_pbp(ids[-1])
#> -- Build nflfastR Play-by-Play Data ------------------ nflfastR version 4.1.0 --
#> * 11:13:02 | Start download of 2 games...
c <- comparedf(all_games %>% filter(game_id %in% sample$game_id), sample)
diffs(c) %>% head(20)
#> var.x var.y ..row.names.. values.x values.y row.x row.y
#> 1 wp wp 11 0.697947.... 0.697489.... 11 11
#> 2 wp wp 93 0.134893.... 0.134536.... 93 93
#> 3 wp wp 102 0.943005.... 0.942860.... 102 102
#> 4 wp wp 112 0.132840.... 0.132257.... 112 112
#> 5 wp wp 127 0.285401.... 0.284856.... 127 127
#> 6 wp wp 156 0.666679.... 0.665633.... 156 156
#> 7 wp wp 223 0.679398.... 0.679094.... 223 223
#> 8 wp wp 244 0.801211.... 0.800766.... 244 244
#> 9 wp wp 277 0.905423.... 0.905035.... 277 277
#> 10 wp wp 284 0.168142.... 0.167975.... 284 284
#> 11 wp wp 294 0.930354.... 0.930306.... 294 294
#> 12 wp wp 305 0.354060.... 0.353571.... 305 305
#> 13 wp wp 323 0.695493.... 0.694733.... 323 323
#> 14 wp wp 347 0.682081.... 0.680360.... 347 347
#> 15 def_wp def_wp 11 0.302052.... 0.302510.... 11 11
#> 16 def_wp def_wp 93 0.865106.... 0.865463.... 93 93
#> 17 def_wp def_wp 102 0.056994.... 0.057139.... 102 102
#> 18 def_wp def_wp 112 0.867159.... 0.867742.... 112 112
#> 19 def_wp def_wp 127 0.714598.... 0.715143.... 127 127
#> 20 def_wp def_wp 156 0.333320.... 0.334366.... 156 156
# load fixed code
devtools::load_all(nflfastR_path)
#> i Loading nflfastR
all_games <- build_nflfastR_pbp(ids)
#> -- Build nflfastR Play-by-Play Data ------------- nflfastR version 4.1.0.9001 --
#> * 11:13:07 | Start download of 3 games...
sample <- build_nflfastR_pbp(ids[-1])
#> -- Build nflfastR Play-by-Play Data ------------- nflfastR version 4.1.0.9001 --
#> * 11:13:12 | Start download of 2 games...
c <- comparedf(all_games %>% filter(game_id %in% sample$game_id), sample)
diffs(c) %>% head(20)
#> [1] var.x var.y ..row.names.. values.x values.y
#> [6] row.x row.y
#> <0 Zeilen> (oder row.names mit Länge 0)
# what is going on?
fg_probs <- function(pbp_data) {
make_pat_prob <- as.numeric(
mgcv::predict.bam(
fastrmodels::fg_model,
newdata = pbp_data %>%
make_model_mutations() %>%
prepare_wp_data() %>%
dplyr::mutate(yardline_100 = if_else(.data$season >= 2015, 15, 3)),
type = "response"
)
)
make_pat_prob[1]
}
fg_probs(all_games)
#> [1] 0.9488284
fg_probs(sample)
#> [1] 0.9314286Created on 2021-03-24 by the reprex package (v1.0.0) |
|
I found another small bug but didn't want to modify too much before we talked about it: Example from 2020 Can't we just use I think the 2pt search actually finds the correct plays but it gets overwritten here Or we modify the matches with something like pat_i <- pat_i[!pat_i %in% two_pt_i] |
No good reason, great catch! I must have been doing something dumb like testing on just one play rather than the normal df. |
I think these columns get messed up by plays with penalty which is why we have this additional search.
Yes I think this is a good solution! |
Added this now |
mrcaseb
changed the title
No description provided.