The goal of MGT Madness was to create a machine learning model that reliably predicted the outcome of both March Madness and regular season college basketball games at a level equal or better to ESPN’s industry-standard prediction models.
If you plan to use this repository ensure that you have the following folders stored in the same directory as your .Rproj file. These directories are required to ensure each script runs. Make sure to include each of the .csv and/or .rds data files.
## levelName
## 1 Data
## 2 ¦--clean
## 3 ¦ ¦--ap_poll_clean_tbl.csv
## 4 ¦ ¦--away_team_stats_tbl.csv
## 5 ¦ ¦--date_tbl.csv
## 6 ¦ ¦--game_information_tbl.csv
## 7 ¦ ¦--home_team_stats_tbl.csv
## 8 ¦ ¦--model_data_tbl.csv
## 9 ¦ °--model_data_tbl.zip
## 10 ¦--models
## 11 ¦ ¦--baseline_rolling_logit_model.rds
## 12 ¦ ¦--baseline_season_logit_model.rds
## 13 ¦ ¦--baseline_standard_logit_model.rds
## 14 ¦ ¦--enhanced_rolling_logit_model.rds
## 15 ¦ ¦--enhanced_rolling_xgb_model.rds
## 16 ¦ ¦--enhanced_season_logit_model.rds
## 17 ¦ ¦--enhanced_season_xgb_model.rds
## 18 ¦ ¦--enhanced_standard_logit_model.rds
## 19 ¦ ¦--enhanced_standard_xgb_model.rds
## 20 ¦ ¦--simple_rolling_logit_model.rds
## 21 ¦ ¦--simple_rolling_probit_model.rds
## 22 ¦ ¦--simple_season_logit_model.rds
## 23 ¦ ¦--simple_season_probit_model.rds
## 24 ¦ ¦--simple_standard_logit_model.rds
## 25 ¦ °--simple_standard_probit_model.rds
## 26 ¦--raw
## 27 ¦ ¦--ap_poll_2012_2022_raw_tbl.csv
## 28 ¦ ¦--geocoded_locations_tbl.csv
## 29 ¦ ¦--mbb_attendance_2012_2022_tbl.csv
## 30 ¦ °--mbb_box_score_2012_2022_tbl.csv
## 31 °--sessionInfo.txt
## levelName
## 1 Final Code
## 2 ¦--1. Data Gathering & Cleaning
## 3 ¦ ¦--0. Package Check.R
## 4 ¦ ¦--1. data_gathering_mbb_box.R
## 5 ¦ ¦--2. data_gathering_attendance.R
## 6 ¦ ¦--3. data_gathering_ap_polls.R
## 7 ¦ ¦--4. data_cleanup.R
## 8 ¦ °--5. create_model_data.R
## 9 ¦--2. Model & Visualization
## 10 ¦ ¦--1. model_building.R
## 11 ¦ ¦--2. model_predictions.R
## 12 ¦ °--draw_confusion_matrix.R
## 13 °--Team15FinalReport.Rmd
Within this directory there are 5 R scripts that were finalized to gather and clean each of our data sources. These scripts utilize numerous data sources such as:
- ESPN web API
hoopR::load_mbb_team_box()function- https://www.sports-reference.com/ AP poll data
- Google Maps API to geocode arena information
Below are a few examples of our data gathering and cleaning process.
Most scripts will have packages listed at the top and should install or
load depending on your situation. If has to be installed, you will need
to run the line once more when the install completes in order to load
the package. Refer to the 0. Package Check.R if there are any
concerns.
if (!require('devtools')) install.packages('devtools')
if (!require('tidyverse')) install.packages('tidyverse')
if (!require('tidymodels')) install.packages('tidymodels')
if (!require('tictoc')) install.packages('tictoc')
if (!require('hoopR')) devtools::install_github('sportsdataverse/hoopR')
if (!require('glue')) install.packages('glue')
if (!require('httr')) install.packages('httr')
if (!require('rvest')) install.packages('rvest')
if (!require('here')) install.packages('here')
if (!require('ggmap')) install.packages('ggmap')
if (!require('glmnet')) install.packages('glmnet')
if (!require('vip')) install.packages('vip')
if (!require('caret')) install.packages('caret')
if (!require('xgboost')) install.packages('xgboost')
if (!require('ggcorrplot')) install.packages('ggcorrplot')
if (!require('zoo')) install.packages('zoo')
if (!require('lubridate')) install.packages('lubridate')
if (!require('geosphere')) install.packages('geosphere')
if (!require('ggthemes')) install.packages('ggthemes')
if (!require('forcats')) install.packages('forcats')
if(!require('bookdown')) install.packages('bookdown')
if(!require('kableExtra')) install.packages('kableExtra')
if(!require('doParallel')) install.packages('doParallel')
if(!require('scales')) install.packages('scales')
if(!require('stringr')) install.packages('stringr')
require(parallel)This code will use the hoopR package and tidyverse to quickly collect the 2022 box score data for all teams.
mbb_box_score_2012_2022_tbl <- hoopR::load_mbb_team_box(seasons = 2022)
mbb_box_score_2012_2022_tbl %>% head()## # A tibble: 6 × 41
## team_id team_uid team_…¹ team_…² team_…³ team_…⁴ team_…⁵ team_…⁶ team_…⁷
## <chr> <chr> <chr> <chr> <chr> <chr> <chr> <chr> <chr>
## 1 3101 s:40~l:41~t:3… utah-t… Utah T… Trailb… UTU Utah T… Utah T… 000000
## 2 2250 s:40~l:41~t:2… gonzag… Gonzaga Bulldo… GONZ Gonzag… Gonzaga 002967
## 3 2934 s:40~l:41~t:2… cal-st… Cal St… Roadru… CSUB Cal St… Bakers… 003BAB
## 4 26 s:40~l:41~t:26 ucla-b… UCLA Bruins UCLA UCLA B… UCLA 005C8E
## 5 2305 s:40~l:41~t:2… kansas… Kansas Jayhaw… KU Kansas… Kansas 0022B4
## 6 127 s:40~l:41~t:1… michig… Michig… Sparta… MSU Michig… Michig… 18453B
## # … with 32 more variables: team_logo <chr>, team_alternate_color <chr>,
## # field_goals_made_field_goals_attempted <chr>, field_goal_pct <chr>,
## # three_point_field_goals_made_three_point_field_goals_attempted <chr>,
## # three_point_field_goal_pct <chr>,
## # free_throws_made_free_throws_attempted <chr>, free_throw_pct <chr>,
## # total_rebounds <chr>, offensive_rebounds <chr>, defensive_rebounds <chr>,
## # team_rebounds <chr>, assists <chr>, steals <chr>, blocks <chr>, …
This function was built to scrape the AP poll data from
sports-reference. Both the rvest and tidyverse suite of packages were
used. Additional data cleaning was performed in our data_cleanup.R
script.
get_ap_polls <- function(year){
url <- glue::glue("https://www.sports-reference.com/cbb/seasons/men/{year}-polls.html#ap-polls")
webpage <- read_html(url)
data <- html_table(html_nodes(webpage,"table")[1])[[1]]
names(data) <- c("school","conference",1:ncol(data))
data <- data %>%
mutate(year = year)
return(data)
}
ap_poll_2012_2022_tbl <- lapply(2012:2022, get_ap_polls) %>% bind_rows()
ap_poll_2012_2022_tbl## # A tibble: 544 × 23
## school confe…¹ `1` `2` `3` `4` `5` `6` `7` `8` `9` `10`
## <chr> <chr> <chr> <chr> <chr> <chr> <chr> <chr> <chr> <chr> <chr> <chr>
## 1 "" "" "1" "2" "3" "4" "5" "6" "7" "8" "9" "10"
## 2 "School" "Conf" "Pre" "11/… "11/… "11/… "12/… "12/… "12/… "12/… "1/2" "1/9"
## 3 "Alabama" "SEC" "19" "16" "13" "12" "16" "23" "" "" "" ""
## 4 "Arizona" "Pac-1… "16" "15" "23" "" "" "" "" "" "" ""
## 5 "Baylor" "Big 1… "12" "11" "9" "7" "6" "6" "6" "6" "4" "4"
## 6 "Califor… "Pac-1… "24" "23" "20" "24" "" "" "" "" "" ""
## 7 "Cincinn… "Big E… "21" "20" "" "" "" "" "" "" "" ""
## 8 "Creight… "MVC" "" "" "" "" "19" "25" "23" "21" "" "23"
## 9 "Duke" "ACC" "6" "6" "6" "3" "7" "7" "7" "7" "5" "8"
## 10 "Florida" "SEC" "8" "7" "10" "10" "12" "13" "11" "10" "13" "19"
## # … with 534 more rows, 11 more variables: `11` <chr>, `12` <chr>, `13` <chr>,
## # `14` <chr>, `15` <chr>, `16` <chr>, `17` <chr>, `18` <chr>, `19` <chr>,
## # year <int>, `20` <chr>, and abbreviated variable name ¹conference
This data requires access to the ESPN API. This access is free but will
be throttled depending on use. If the script returns no data or missing
data then a VPN may be required to bypass the ESPN restrictions. This
function can be modified to include box score data similar to the
hoopR::load_mbb_team_box() function. Using the hoopR function saved us
time for this project but pulling directly from the API had many uses as
the data is much more likely to be completely filled in.
# Select First 5 Game IDs from our tibble
game_ids_vec <- mbb_box_score_2012_2022_tbl %>%
pull(game_id) %>%
unique() %>%
.[1:5]
# Initialize ESPN Attendance function
get_attendance_espn_api <- function(game_id){
tryCatch({url = glue::glue("https://site.api.espn.com/apis/site/v2/sports/basketball/mens-college-basketball/summary?event={game_id}")
txt = httr::GET(url) %>% httr::content(as = "text")
game_info <- jsonlite::fromJSON(txt)
temp <- game_info %>%
enframe() %>%
pivot_wider() %>%
select(gameInfo, predictor, odds) %>%
unnest_wider(gameInfo) %>%
unnest_wider(venue) %>%
unnest_wider(predictor) %>%
unnest_wider(homeTeam,names_sep = "_") %>%
unnest_wider(awayTeam,names_sep = "_") %>%
unnest_wider(address) %>%
unnest_wider(odds) %>%
select(-officials, -images, -grass, -id, -header) %>%
mutate(game_id = game_id, type = 1)
temp2 <- game_info %>%
enframe() %>%
pivot_wider() %>%
select(header) %>%
unnest_wider(header)
temp$neutral_site <- temp2$competitions$neutralSite
return(temp)
}, error = function(e) {
message(paste("Error getting data for date", game_id))
temp <- data.frame(game_id = game_id, type = 0)
return(temp)
})
}
tictoc::tic()
mbb_attendance_2012_2022_tbl <- lapply(game_ids_vec, get_attendance_espn_api) %>% bind_rows()
tictoc::toc()## 2.05 sec elapsed
mbb_attendance_2012_2022_tbl## # A tibble: 5 × 16
## fullName city state capac…¹ atten…² homeT…³ homeT…⁴ homeT…⁵ homeT…⁶ awayT…⁷
## <chr> <chr> <chr> <int> <int> <chr> <chr> <chr> <chr> <chr>
## 1 McCarthey… Spok… WA 6000 6000 2250 98.8 1.2 0.0 3101
## 2 Pauley Pa… Los … CA 13800 5618 26 95.0 5.0 0.0 2934
## 3 Madison S… New … NY 19812 0 127 35.5 64.5 0.0 2305
## 4 Finneran … Vill… PA 6501 6501 222 98.3 1.7 0.0 116
## 5 Frank Erw… Aust… TX 16540 14683 251 96.9 3.1 0.0 2277
## # … with 6 more variables: awayTeam_gameProjection <chr>,
## # awayTeam_teamChanceLoss <chr>, awayTeam_teamChanceTie <chr>, game_id <int>,
## # type <dbl>, neutral_site <lgl>, and abbreviated variable names ¹capacity,
## # ²attendance, ³homeTeam_id, ⁴homeTeam_gameProjection,
## # ⁵homeTeam_teamChanceLoss, ⁶homeTeam_teamChanceTie, ⁷awayTeam_id
This directory contains 3 scripts in total:
- model_building.R
- model_predictions.R
- draw_confusion_matrix.R
Model building should be run first in order to create the models, assuming you’ve already create/pulled all relevant data sources. Model predictions will access our stored models and create the relevant R objects for the RMarkdown file. The confusion matrix script is meant to be a helper script that adds a bit of flare to our contingency tables.
We will not cover the models in detail but the general concept here is to read in the data created in the previous steps then perform various modeling techniques ranging from LASSO regression, Logistic Regression, Probit Regression and Decision Trees. Please see the tidymodels documentation to find out more about our approach and see examples. You can find this documentation at https://www.tidymodels.org/
Again, we will not cover the model information in detail as the code is quite long. Once again the concept is similar to Model Building. If you have the models created or pulled from the repository, this script can be ran and our contingency tables created for evaluation.
Below is the session information that was used to create this report. The packages and their versions are listed. R 4.2.3 is required.
sessionInfo()## R version 4.2.3 (2023-03-15 ucrt)
## Platform: x86_64-w64-mingw32/x64 (64-bit)
## Running under: Windows 10 x64 (build 22621)
##
## Matrix products: default
##
## locale:
## [1] LC_COLLATE=English_United States.utf8
## [2] LC_CTYPE=English_United States.utf8
## [3] LC_MONETARY=English_United States.utf8
## [4] LC_NUMERIC=C
## [5] LC_TIME=English_United States.utf8
##
## attached base packages:
## [1] parallel stats graphics grDevices utils datasets methods
## [8] base
##
## other attached packages:
## [1] doParallel_1.0.17 iterators_1.0.14 foreach_1.5.2 kableExtra_1.3.4
## [5] bookdown_0.29 ggthemes_4.2.4 geosphere_1.5-18 lubridate_1.9.0
## [9] timechange_0.1.1 zoo_1.8-10 ggcorrplot_0.1.4 xgboost_1.7.3.1
## [13] caret_6.0-93 lattice_0.20-45 vip_0.3.2 glmnet_4.1-4
## [17] Matrix_1.5-1 ggmap_3.0.1 here_1.0.1 rvest_1.0.3
## [21] httr_1.4.5 glue_1.6.2 hoopR_1.9.1.9000 tictoc_1.1
## [25] yardstick_1.1.0 workflowsets_1.0.0 workflows_1.1.0 tune_1.0.0
## [29] rsample_1.1.0 recipes_1.0.1 parsnip_1.0.1 modeldata_1.0.1
## [33] infer_1.0.3 dials_1.0.0 scales_1.2.1 broom_1.0.1
## [37] tidymodels_1.0.0 devtools_2.4.5 usethis_2.1.6 plyr_1.8.7
## [41] data.tree_1.0.0 forcats_0.5.2 stringr_1.5.0 dplyr_1.0.10
## [45] purrr_1.0.1 readr_2.1.2 tidyr_1.3.0 tibble_3.1.8
## [49] ggplot2_3.3.6 tidyverse_1.3.2
##
## loaded via a namespace (and not attached):
## [1] readxl_1.4.1 backports_1.4.1 systemfonts_1.0.4
## [4] selectr_0.4-2 sp_1.6-0 splines_4.2.3
## [7] listenv_0.9.0 digest_0.6.29 htmltools_0.5.3
## [10] fansi_1.0.3 magrittr_2.0.3 memoise_2.0.1
## [13] googlesheets4_1.0.1 tzdb_0.3.0 remotes_2.4.2
## [16] globals_0.16.2 modelr_0.1.9 gower_1.0.0
## [19] RcppParallel_5.1.5 svglite_2.1.1 hardhat_1.2.0
## [22] prettyunits_1.1.1 jpeg_0.1-10 colorspace_2.0-3
## [25] haven_2.5.1 xfun_0.33 callr_3.7.2
## [28] crayon_1.5.2 jsonlite_1.8.0 progressr_0.13.0
## [31] survival_3.5-3 gtable_0.3.1 gargle_1.2.1
## [34] ipred_0.9-13 webshot_0.5.4 pkgbuild_1.3.1
## [37] shape_1.4.6 future.apply_1.9.1 DBI_1.1.3
## [40] miniUI_0.1.1.1 Rcpp_1.0.9 viridisLite_0.4.1
## [43] xtable_1.8-4 GPfit_1.0-8 stats4_4.2.3
## [46] lava_1.6.10 prodlim_2019.11.13 profvis_0.3.7
## [49] htmlwidgets_1.5.4 ellipsis_0.3.2 urlchecker_1.0.1
## [52] pkgconfig_2.0.3 nnet_7.3-18 dbplyr_2.2.1
## [55] utf8_1.2.2 janitor_2.2.0 reshape2_1.4.4
## [58] tidyselect_1.2.0 rlang_1.0.6 DiceDesign_1.9
## [61] later_1.3.0 munsell_0.5.0 cellranger_1.1.0
## [64] tools_4.2.3 cachem_1.0.6 cli_3.4.0
## [67] generics_0.1.3 evaluate_0.16 fastmap_1.1.0
## [70] yaml_2.3.5 ModelMetrics_1.2.2.2 processx_3.7.0
## [73] knitr_1.40 fs_1.5.2 RgoogleMaps_1.4.5.3
## [76] nlme_3.1-162 future_1.32.0 mime_0.12
## [79] xml2_1.3.3 compiler_4.2.3 rstudioapi_0.14
## [82] curl_4.3.2 png_0.1-7 reprex_2.0.2
## [85] lhs_1.1.5 stringi_1.7.8 ps_1.7.1
## [88] vctrs_0.5.2 pillar_1.8.1 lifecycle_1.0.3
## [91] furrr_0.3.1 bitops_1.0-7 data.table_1.14.2
## [94] httpuv_1.6.6 R6_2.5.1 promises_1.2.0.1
## [97] gridExtra_2.3 parallelly_1.34.0 sessioninfo_1.2.2
## [100] codetools_0.2-19 MASS_7.3-58.2 assertthat_0.2.1
## [103] pkgload_1.3.0 rprojroot_2.0.3 withr_2.5.0
## [106] hms_1.1.2 grid_4.2.3 rpart_4.1.19
## [109] timeDate_4021.104 class_7.3-21 rmarkdown_2.16
## [112] snakecase_0.11.0 googledrive_2.0.0 pROC_1.18.0
## [115] shiny_1.7.2