Using Code to Predict the 2024 March Madness Tournament

ncaa
machinelearning
basketball
Author

Greg Johnson

Published

April 15, 2024

My Advanced Sports Data class at the #cojmcfamily at the University of Nebraska-Lincoln had a bracket challenge contest, and I got 10th place out of 11. Let’s break it down.

Code 
library(tidyverse)
library(tidymodels)
library(hoopR)
library(zoo)
library(gt)
library(bonsai)
set.seed(1234)

games <- load_mbb_team_box(seasons = 2015:2024) |> filter(game_date < as.Date("2024-03-18"))

nond1 <- games |> group_by(team_id, season) |> tally() |> filter(n < 10) |> select(team_id)
nond1 <- pull(nond1)

df <- games |> filter(!team_id %in% nond1 & !opponent_team_id %in% nond1)

teamside <- df |> 
  group_by(team_short_display_name, season) |> 
  arrange(game_date) |> 
  mutate(
    team_possessions = field_goals_attempted - offensive_rebounds + turnovers + (.475 * free_throws_attempted),
    team_points_per_possession = team_score/team_possessions,
    team_defensive_points_per_possession = opponent_team_score/team_possessions,
    team_offensive_efficiency = team_points_per_possession * 100,
    team_defensive_efficiency = team_defensive_points_per_possession * 100,
    team_points_per_possession = team_score/team_possessions,
    team_defensive_points_per_possession = opponent_team_score/team_possessions,
    team_season_offensive_efficiency = lag(cummean(team_offensive_efficiency), n=1),
    team_season_defensive_efficiency = lag(cummean(team_defensive_efficiency), n=1),  
    score_margin = team_score - opponent_team_score,
    absolute_score_margin = abs(score_margin), 
    team_score_margin = team_score - opponent_team_score,
    team_rolling_mean_score_margin = rollmean(lag(team_score_margin, n=1), k=10, align="right", fill=NA),
    team_rolling_mean_offensive_efficiency = rollmean(lag(team_offensive_efficiency, n=1), k=5, align="right", fill=NA),
    team_rolling_mean_defensive_efficiency = rollmean(lag(team_defensive_efficiency, n=1), k=5, align="right", fill=NA),
    team_free_throws_made = free_throws_made,
    team_free_throw_pct = free_throw_pct,
    team_season_free_throw_pct = lag(cummean(team_free_throw_pct), n=1),
    team_season_free_throws_made = lag(cummean(team_free_throws_made), n=1),
    team_offensive_rebounds_per_possession = offensive_rebounds/team_possessions,
    team_steals_per_possession = steals/team_possessions,
    team_steals = steals,
    team_season_steals = lag(cummean(team_steals), n=1),
    team_field_goals_attempted = field_goals_attempted,
    team_free_throws_attempted = free_throws_attempted, 
    team_season_score_margin = lag(cummean(team_score_margin), n=1),
    team_total_rebounds = total_rebounds,
    team_season_total_rebounds = lag(cummean(team_total_rebounds), n=1),
    team_rolling_total_rebounds = rollmean(lag(team_total_rebounds, n=1), k=5, align="right", fill =NA),
    team_fouls = fouls,
    team_rebounds_per_possession = total_rebounds/team_possessions,
    team_three_point_field_goals_made = three_point_field_goals_made,
    team_rolling_three_point_field_goals_made = rollmean(lag(team_three_point_field_goals_made, n=1), k=5, align="right", fill =NA),
    team_three_point_field_goal_pct = three_point_field_goal_pct,
    team_field_goals_made = field_goals_made,
    team_rolling_total_field_goals_made = rollmean(lag(team_field_goals_made, n=1), k=5, align="right", fill =NA),
    team_field_goal_pct = field_goal_pct,
    team_season_field_goal_pct = lag(cummean(team_field_goal_pct), n=1),
    team_rolling_field_goal_pct = rollmean(lag(team_field_goal_pct, n=1), k=5, align="right", fill =NA),
    team_field_goals_made_per_possession = field_goals_made/team_possessions
  ) |> 
  filter(absolute_score_margin <= 40) |>
  ungroup()

opponentside <- teamside |> 
  select(-opponent_team_id) |> 
  rename(
    opponent_team_id = team_id,
    opponent_season_offensive_efficiency = team_season_offensive_efficiency,
    opponent_season_defensive_efficiency = team_season_defensive_efficiency,
    opponent_rolling_mean_offensive_efficiency = team_rolling_mean_offensive_efficiency,
    opponent_rolling_mean_defensive_efficiency = team_rolling_mean_defensive_efficiency,
    opponent_offensive_efficiency = team_offensive_efficiency,
    opponent_defensive_efficiency = team_defensive_efficiency,
    opponent_free_throws_made = team_free_throws_made,
    opponent_season_free_throw_pct = team_season_free_throw_pct,
    opponent_season_free_throws_made = team_season_free_throws_made,
    opponent_offensive_rebounds_per_possession = team_offensive_rebounds_per_possession,
    opponent_steals_per_possession = team_steals_per_possession,
    opponent_season_steals = team_season_steals,
    opponent_field_goals_attempted = team_field_goals_attempted,
    opponent_free_throws_attempted = team_free_throws_attempted,
    opponent_season_score_margin = team_season_score_margin,
    opponent_rolling_mean_score_margin = team_rolling_mean_score_margin,
    opponent_total_rebounds = team_total_rebounds,
    opponent_season_total_rebounds = team_season_total_rebounds,
    opponent_rolling_total_rebounds = team_rolling_total_rebounds,
    opponent_fouls = team_fouls,
    opponent_rebounds_per_possession = team_rebounds_per_possession,
    opponent_three_point_field_goals_made = team_three_point_field_goals_made,
    opponent_rolling_three_point_field_goals_made = team_rolling_three_point_field_goals_made,
    opponent_three_point_field_goal_pct = team_three_point_field_goal_pct,
    opponent_field_goals_made = team_field_goals_made,
    opponent_rolling_total_field_goals_made = team_rolling_total_field_goals_made,
    opponent_field_goal_pct = team_field_goal_pct,
    opponent_season_field_goal_pct = team_season_field_goal_pct,
    opponent_rolling_field_goal_pct = team_rolling_field_goal_pct,
    opponent_field_goals_made_per_possession = team_field_goals_made_per_possession
  ) |> 
  select(
    game_id,
    opponent_team_id,
    opponent_season_offensive_efficiency,
    opponent_season_defensive_efficiency,
    opponent_rolling_mean_offensive_efficiency,
    opponent_rolling_mean_defensive_efficiency,
    opponent_free_throws_made,
    opponent_season_free_throws_made,
    opponent_season_free_throw_pct,
    opponent_offensive_rebounds_per_possession,
    opponent_steals_per_possession,
    opponent_season_steals,
    opponent_field_goals_attempted,
    opponent_free_throws_attempted,
    opponent_season_score_margin,
    opponent_rolling_mean_score_margin,
    opponent_total_rebounds,
    opponent_rolling_total_rebounds,
    opponent_season_total_rebounds,
    opponent_fouls,
    opponent_rebounds_per_possession,
    opponent_three_point_field_goals_made,
    opponent_rolling_three_point_field_goals_made,
    opponent_three_point_field_goal_pct,
    opponent_field_goals_made,
    opponent_rolling_total_field_goals_made,
    opponent_field_goal_pct,
    opponent_season_field_goal_pct,
    opponent_rolling_field_goal_pct,
    opponent_field_goals_made_per_possession
  )

bothsides <- teamside |> inner_join(opponentside)

bothsides <- bothsides |> mutate(
  team_result = as.factor(case_when(
    team_score > opponent_team_score ~ "W",
    opponent_team_score > team_score ~ "L"
)))

levels(bothsides$team_result)

bothsides$team_result <- relevel(bothsides$team_result, ref="W")

levels(bothsides$team_result)

modelgames <- bothsides |> 
  select(
    game_id, 
    game_date, 
    team_short_display_name, 
    opponent_team_short_display_name, 
    season, 
    team_season_offensive_efficiency,
    team_season_defensive_efficiency,
    opponent_season_offensive_efficiency,
    opponent_season_defensive_efficiency,
    team_result, 
    opponent_rolling_mean_offensive_efficiency,
    opponent_rolling_mean_defensive_efficiency,
    team_rolling_mean_offensive_efficiency,
    team_rolling_mean_defensive_efficiency, 
    opponent_season_free_throws_made,
    opponent_season_steals,
    opponent_rolling_mean_score_margin,
    opponent_season_total_rebounds,
    opponent_rolling_total_rebounds,
    opponent_rolling_three_point_field_goals_made,
    opponent_rolling_total_field_goals_made,
    opponent_season_field_goal_pct,
    opponent_rolling_field_goal_pct,
    ) |> na.omit()

game_split <- initial_split(modelgames, prop = .8)
game_train <- training(game_split)
game_test <- testing(game_split)

game_recipe <- 
  recipe(team_result ~ ., data = game_train) |> 
  update_role(game_id, game_date, team_short_display_name, opponent_team_short_display_name, season, new_role = "ID") |>
  step_normalize(all_predictors())

summary(game_recipe)


lightgbm_mod <- 
  boost_tree() |>
  set_engine("lightgbm") |>
  set_mode(mode = "classification")


lightgbm_workflow <- 
  workflow() |> 
  add_model(lightgbm_mod) |> 
  add_recipe(game_recipe)


lightgbm_fit <- 
  lightgbm_workflow |> 
  fit(data = game_train)


lightgbmpredict <- lightgbm_fit |> predict(new_data = game_train) |>
  bind_cols(game_train) 

lightgbmpredict <- lightgbm_fit |> predict(new_data = game_train, type="prob") |>
  bind_cols(lightgbmpredict)


metrics(lightgbmpredict, team_result, .pred_class)

lightgbmtestpredict <- lightgbm_fit |> predict(new_data = game_test) |>
  bind_cols(game_test)

lightgbmtestpredict <- lightgbm_fit |> predict(new_data = game_test, type="prob") |>
  bind_cols(lightgbmtestpredict)

metrics(lightgbmtestpredict, team_result, .pred_class)


teamside <- df |> 
  group_by(team_short_display_name, season) |> 
  arrange(game_date) |> 
  mutate(
    team_possessions = field_goals_attempted - offensive_rebounds + turnovers + (.475 * free_throws_attempted),
    team_points_per_possession = team_score/team_possessions,
    team_defensive_points_per_possession = opponent_team_score/team_possessions,
    team_offensive_efficiency = team_points_per_possession * 100,
    team_defensive_efficiency = team_defensive_points_per_possession * 100,
    team_points_per_possession = team_score/team_possessions,
    team_defensive_points_per_possession = opponent_team_score/team_possessions,
    team_season_offensive_efficiency = cummean(team_offensive_efficiency),
    team_season_defensive_efficiency = cummean(team_defensive_efficiency),  
    score_margin = team_score - opponent_team_score,
    absolute_score_margin = abs(score_margin), 
    team_score_margin = team_score - opponent_team_score,
    team_rolling_mean_score_margin = rollmean(team_score_margin, k=10, align="right", fill=NA),
    team_rolling_mean_offensive_efficiency = rollmean(team_offensive_efficiency, k=5, align="right", fill=NA),
    team_rolling_mean_defensive_efficiency = rollmean(team_defensive_efficiency, k=5, align="right", fill=NA),
    team_free_throws_made = free_throws_made,
    team_free_throw_pct = free_throw_pct,
    team_season_free_throw_pct = cummean(team_free_throw_pct),
    team_season_free_throws_made = cummean(team_free_throws_made),
    team_offensive_rebounds_per_possession = offensive_rebounds/team_possessions,
    team_steals_per_possession = steals/team_possessions,
    team_steals = steals,
    team_season_steals = cummean(team_steals),
    team_field_goals_attempted = field_goals_attempted,
    team_free_throws_attempted = free_throws_attempted, 
    team_season_score_margin = cummean(team_score_margin),
    team_total_rebounds = total_rebounds,
    team_season_total_rebounds = cummean(team_total_rebounds),
    team_rolling_total_rebounds = rollmean(team_total_rebounds, k=5, align="right", fill =NA),
    team_fouls = fouls,
    team_rebounds_per_possession = total_rebounds/team_possessions,
    team_three_point_field_goals_made = three_point_field_goals_made,
    team_rolling_three_point_field_goals_made = rollmean(team_three_point_field_goals_made, k=5, align="right", fill =NA),
    team_three_point_field_goal_pct = three_point_field_goal_pct,
    team_field_goals_made = field_goals_made,
    team_rolling_total_field_goals_made = rollmean(team_field_goals_made, k=5, align="right", fill =NA),
    team_field_goal_pct = field_goal_pct,
    team_season_field_goal_pct = cummean(team_field_goal_pct),
    team_rolling_field_goal_pct = rollmean(team_field_goal_pct, k=5, align="right", fill =NA),
    team_field_goals_made_per_possession = field_goals_made/team_possessions
  ) |> 
  filter(absolute_score_margin <= 40) |>
  ungroup()

opponentside <- teamside |> 
  select(-opponent_team_id) |> 
  rename(
    opponent_team_id = team_id,
    opponent_season_offensive_efficiency = team_season_offensive_efficiency,
    opponent_season_defensive_efficiency = team_season_defensive_efficiency,
    opponent_rolling_mean_offensive_efficiency = team_rolling_mean_offensive_efficiency,
    opponent_rolling_mean_defensive_efficiency = team_rolling_mean_defensive_efficiency,
    opponent_offensive_efficiency = team_offensive_efficiency,
    opponent_defensive_efficiency = team_defensive_efficiency,
    opponent_free_throws_made = team_free_throws_made,
    opponent_season_free_throw_pct = team_season_free_throw_pct,
    opponent_season_free_throws_made = team_season_free_throws_made,
    opponent_offensive_rebounds_per_possession = team_offensive_rebounds_per_possession,
    opponent_steals_per_possession = team_steals_per_possession,
    opponent_season_steals = team_season_steals,
    opponent_field_goals_attempted = team_field_goals_attempted,
    opponent_free_throws_attempted = team_free_throws_attempted,
    opponent_season_score_margin = team_season_score_margin,
    opponent_rolling_mean_score_margin = team_rolling_mean_score_margin,
    opponent_total_rebounds = team_total_rebounds,
    opponent_season_total_rebounds = team_season_total_rebounds,
    opponent_rolling_total_rebounds = team_rolling_total_rebounds,
    opponent_fouls = team_fouls,
    opponent_rebounds_per_possession = team_rebounds_per_possession,
    opponent_three_point_field_goals_made = team_three_point_field_goals_made,
    opponent_rolling_three_point_field_goals_made = team_rolling_three_point_field_goals_made,
    opponent_three_point_field_goal_pct = team_three_point_field_goal_pct,
    opponent_field_goals_made = team_field_goals_made,
    opponent_rolling_total_field_goals_made = team_rolling_total_field_goals_made,
    opponent_field_goal_pct = team_field_goal_pct,
    opponent_season_field_goal_pct = team_season_field_goal_pct,
    opponent_rolling_field_goal_pct = team_rolling_field_goal_pct,
    opponent_field_goals_made_per_possession = team_field_goals_made_per_possession
  ) |> 
  select(
    game_id,
    opponent_team_id,
    opponent_season_offensive_efficiency,
    opponent_season_defensive_efficiency,
    opponent_rolling_mean_offensive_efficiency,
    opponent_rolling_mean_defensive_efficiency,
    opponent_free_throws_made,
    opponent_season_free_throws_made,
    opponent_season_free_throw_pct,
    opponent_offensive_rebounds_per_possession,
    opponent_steals_per_possession,
    opponent_season_steals,
    opponent_field_goals_attempted,
    opponent_free_throws_attempted,
    opponent_season_score_margin,
    opponent_rolling_mean_score_margin,
    opponent_total_rebounds,
    opponent_rolling_total_rebounds,
    opponent_season_total_rebounds,
    opponent_fouls,
    opponent_rebounds_per_possession,
    opponent_three_point_field_goals_made,
    opponent_rolling_three_point_field_goals_made,
    opponent_three_point_field_goal_pct,
    opponent_field_goals_made,
    opponent_rolling_total_field_goals_made,
    opponent_field_goal_pct,
    opponent_season_field_goal_pct,
    opponent_rolling_field_goal_pct,
    opponent_field_goals_made_per_possession
  )

bothsides <- teamside |> inner_join(opponentside)

bothsides <- bothsides |> mutate(
  team_result = as.factor(case_when(
    team_score > opponent_team_score ~ "W",
    opponent_team_score > team_score ~ "L"
)))

levels(bothsides$team_result)

bothsides$team_result <- relevel(bothsides$team_result, ref="W")

levels(bothsides$team_result)

modelgames <- bothsides |> 
  select(
    game_id, 
    game_date, 
    team_short_display_name, 
    opponent_team_short_display_name, 
    season, 
    team_season_offensive_efficiency,
    team_season_defensive_efficiency,
    opponent_season_offensive_efficiency,
    opponent_season_defensive_efficiency,
    team_result, 
    opponent_rolling_mean_offensive_efficiency,
    opponent_rolling_mean_defensive_efficiency,
    team_rolling_mean_offensive_efficiency,
    team_rolling_mean_defensive_efficiency, 
    opponent_season_free_throws_made,
    opponent_season_steals,
    opponent_rolling_mean_score_margin,
    opponent_season_total_rebounds,
    opponent_rolling_total_rebounds,
    opponent_rolling_three_point_field_goals_made,
    opponent_rolling_total_field_goals_made,
    opponent_season_field_goal_pct,
    opponent_rolling_field_goal_pct,
    ) |> na.omit()

eastround1games <- tibble(
  team_short_display_name="UConn",
  opponent_team_short_display_name="Stetson"
) |> add_row(
  team_short_display_name="FAU",
  opponent_team_short_display_name="Northwestern"
) |> add_row(
  team_short_display_name="San Diego St",
  opponent_team_short_display_name="UAB"
) |> add_row(
  team_short_display_name="Auburn",
  opponent_team_short_display_name="Yale"
) |> add_row(
  team_short_display_name="BYU",
  opponent_team_short_display_name="Duquesne"
) |> add_row(
  team_short_display_name="Illinois",
  opponent_team_short_display_name="Morehead St"
) |> add_row(
  team_short_display_name="Washington St",
  opponent_team_short_display_name="Drake"
) |> add_row(
  team_short_display_name="Iowa State",
  opponent_team_short_display_name="S Dakota St"
)

eastround1games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(eastround1games)

eastround1games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(eastround1games) 

eastround1 <- lightgbm_fit |> predict(new_data = eastround1games) |>
  bind_cols(eastround1games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

eastround1 <- lightgbm_fit |> predict(new_data = eastround1games, type="prob") |>
  bind_cols(eastround1) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

eastround1

westround1games <- tibble(
  team_short_display_name="North Carolina",
  opponent_team_short_display_name="Wagner"
) |> add_row(
  team_short_display_name="Mississippi St",
  opponent_team_short_display_name="Michigan St"
) |> add_row(
  team_short_display_name="Saint Mary's",
  opponent_team_short_display_name="Grand Canyon"
) |> add_row(
  team_short_display_name="Alabama",
  opponent_team_short_display_name="Charleston"
) |> add_row(
  team_short_display_name="Clemson",
  opponent_team_short_display_name="New Mexico"
) |> add_row(
  team_short_display_name="Baylor",
  opponent_team_short_display_name="Colgate"
) |> add_row(
  team_short_display_name="Dayton",
  opponent_team_short_display_name="Nevada"
) |> add_row(
  team_short_display_name="Arizona",
  opponent_team_short_display_name="Long Beach St"
)

westround1games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(westround1games)

westround1games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(westround1games) 

westround1 <- lightgbm_fit |> predict(new_data = westround1games) |>
  bind_cols(westround1games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

westround1 <- lightgbm_fit |> predict(new_data = westround1games, type="prob") |>
  bind_cols(westround1) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

westround1

southround1games <- tibble(
  team_short_display_name="Houston",
  opponent_team_short_display_name="Longwood"
) |> add_row(
  team_short_display_name="Nebraska",
  opponent_team_short_display_name="Texas A&M"
) |> add_row(
  team_short_display_name="Wisconsin",
  opponent_team_short_display_name="James Madison"
) |> add_row(
  team_short_display_name="Duke",
  opponent_team_short_display_name="Vermont"
) |> add_row(
  team_short_display_name="Texas Tech",
  opponent_team_short_display_name="NC State"
) |> add_row(
  team_short_display_name="Kentucky",
  opponent_team_short_display_name="Oakland"
) |> add_row(
  team_short_display_name="Florida",
  opponent_team_short_display_name="Colorado"
) |> add_row(
  team_short_display_name="Marquette",
  opponent_team_short_display_name="Western KY"
)

southround1games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(southround1games)

southround1games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(southround1games) 

southround1 <- lightgbm_fit |> predict(new_data = southround1games) |>
  bind_cols(southround1games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

southround1 <- lightgbm_fit |> predict(new_data = southround1games, type="prob") |>
  bind_cols(southround1) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

southround1

midwestround1games <- tibble(
  team_short_display_name="Purdue",
  opponent_team_short_display_name="Grambling"
) |> add_row(
  team_short_display_name="Utah State",
  opponent_team_short_display_name="TCU"
) |> add_row(
  team_short_display_name="Gonzaga",
  opponent_team_short_display_name="McNeese"
) |> add_row(
  team_short_display_name="Kansas",
  opponent_team_short_display_name="Samford"
) |> add_row(
  team_short_display_name="South Carolina",
  opponent_team_short_display_name="Oregon"
) |> add_row(
  team_short_display_name="Creighton",
  opponent_team_short_display_name="Akron"
) |> add_row(
  team_short_display_name="Texas",
  opponent_team_short_display_name="Colorado St"
) |> add_row(
  team_short_display_name="Tennessee",
  opponent_team_short_display_name="Saint Peter's"
)

midwestround1games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(midwestround1games)

midwestround1games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(midwestround1games) 

midwestround1 <- lightgbm_fit |> predict(new_data = midwestround1games) |>
  bind_cols(midwestround1games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

midwestround1 <- lightgbm_fit |> predict(new_data = midwestround1games, type="prob") |>
  bind_cols(midwestround1) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

midwestround1

eastround2games <- tibble(
  team_short_display_name="UConn",
  opponent_team_short_display_name="FAU"
) |> add_row(
  team_short_display_name="San Diego St",
  opponent_team_short_display_name="Auburn"
) |> add_row(
  team_short_display_name="BYU",
  opponent_team_short_display_name="Illinois"
) |> add_row(
  team_short_display_name="Drake",
  opponent_team_short_display_name="Iowa State"
) 

eastround2games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(eastround2games)

eastround2games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(eastround2games) 

eastround2 <- lightgbm_fit |> predict(new_data = eastround2games) |>
  bind_cols(eastround2games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

eastround2 <- lightgbm_fit |> predict(new_data = eastround2games, type="prob") |>
  bind_cols(eastround2) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

eastround2

westround2games <- tibble(
  team_short_display_name="North Carolina",
  opponent_team_short_display_name="Michigan St"
) |> add_row(
  team_short_display_name="Saint Mary's",
  opponent_team_short_display_name="Alabama"
) |> add_row(
  team_short_display_name="New Mexico",
  opponent_team_short_display_name="Baylor"
) |> add_row(
  team_short_display_name="Nevada",
  opponent_team_short_display_name="Arizona"
) 

westround2games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(westround2games)

westround2games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(westround2games) 

westround2 <- lightgbm_fit |> predict(new_data = westround2games) |>
  bind_cols(westround2games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

westround2 <- lightgbm_fit |> predict(new_data = westround2games, type="prob") |>
  bind_cols(westround2) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

westround2

southround2games <- tibble(
  team_short_display_name="Houston",
  opponent_team_short_display_name="Nebraska"
) |> add_row(
  team_short_display_name="James Madison",
  opponent_team_short_display_name="Duke"
) |> add_row(
  team_short_display_name="Texas Tech",
  opponent_team_short_display_name="Kentucky"
) |> add_row(
  team_short_display_name="Colorado",
  opponent_team_short_display_name="Marquette"
)

southround2games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(southround2games)

southround2games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(southround2games) 

southround2 <- lightgbm_fit |> predict(new_data = southround2games) |>
  bind_cols(southround2games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

southround2 <- lightgbm_fit |> predict(new_data = southround2games, type="prob") |>
  bind_cols(southround2) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

southround2

midwestround2games <- tibble(
  team_short_display_name="Purdue",
  opponent_team_short_display_name="TCU"
) |> add_row(
  team_short_display_name="McNeese",
  opponent_team_short_display_name="Samford"
) |> add_row(
  team_short_display_name="South Carolina",
  opponent_team_short_display_name="Creighton"
) |> add_row(
  team_short_display_name="Texas",
  opponent_team_short_display_name="Tennessee"
)

midwestround2games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(midwestround2games)

midwestround2games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(midwestround2games) 

midwestround2 <- lightgbm_fit |> predict(new_data = midwestround2games) |>
  bind_cols(midwestround2games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

midwestround2 <- lightgbm_fit |> predict(new_data = midwestround2games, type="prob") |>
  bind_cols(midwestround2) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

midwestround2

eastround3games <- tibble(
  team_short_display_name="UConn",
  opponent_team_short_display_name="Auburn"
) |> add_row(
  team_short_display_name="BYU",
  opponent_team_short_display_name="Iowa State"
) 

eastround3games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(eastround3games)

eastround3games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(eastround3games) 

eastround3 <- lightgbm_fit |> predict(new_data = eastround3games) |>
  bind_cols(eastround3games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

eastround3 <- lightgbm_fit |> predict(new_data = eastround3games, type="prob") |>
  bind_cols(eastround3) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

eastround3

westround3games <- tibble(
  team_short_display_name="North Carolina",
  opponent_team_short_display_name="Saint Mary's"
) |> add_row(
  team_short_display_name="New Mexico",
  opponent_team_short_display_name="Arizona"
) 

westround3games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(westround3games)

westround3games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(westround3games) 

westround3 <- lightgbm_fit |> predict(new_data = westround3games) |>
  bind_cols(westround3games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

westround3 <- lightgbm_fit |> predict(new_data = westround3games, type="prob") |>
  bind_cols(westround3) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

westround3

southround3games <- tibble(
  team_short_display_name="Houston",
  opponent_team_short_display_name="Duke"
) |> add_row(
  team_short_display_name="Kentucky",
  opponent_team_short_display_name="Colorado"
)

southround3games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(southround3games)

southround3games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(southround3games) 

southround3 <- lightgbm_fit |> predict(new_data = southround3games) |>
  bind_cols(southround3games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

southround3 <- lightgbm_fit |> predict(new_data = southround3games, type="prob") |>
  bind_cols(southround3) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

southround3

midwestround3games <- tibble(
  team_short_display_name="Purdue",
  opponent_team_short_display_name="Samford"
) |> add_row(
  team_short_display_name="Creighton",
  opponent_team_short_display_name="Tennessee"
)

midwestround3games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(midwestround3games)

midwestround3games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(midwestround3games) 

midwestround3 <- lightgbm_fit |> predict(new_data = midwestround3games) |>
  bind_cols(midwestround3games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

midwestround3 <- lightgbm_fit |> predict(new_data = midwestround3games, type="prob") |>
  bind_cols(midwestround3) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

midwestround3

eastround4games <- tibble(
  team_short_display_name="UConn",
  opponent_team_short_display_name="BYU"
) 

eastround4games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(eastround4games)

eastround4games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(eastround4games) 

eastround4 <- lightgbm_fit |> predict(new_data = eastround4games) |>
  bind_cols(eastround4games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

eastround4 <- lightgbm_fit |> predict(new_data = eastround4games, type="prob") |>
  bind_cols(eastround4) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

eastround4

westround4games <- tibble(
  team_short_display_name="Saint Mary's",
  opponent_team_short_display_name="Arizona"
) 

westround4games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(westround4games)

westround4games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(westround4games) 

westround4 <- lightgbm_fit |> predict(new_data = westround4games) |>
  bind_cols(westround4games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

westround4 <- lightgbm_fit |> predict(new_data = westround4games, type="prob") |>
  bind_cols(westround4) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

westround4

southround4games <- tibble(
  team_short_display_name="Houston",
  opponent_team_short_display_name="Colorado"
)

southround4games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(southround4games)

southround4games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(southround4games) 

southround4 <- lightgbm_fit |> predict(new_data = southround4games) |>
  bind_cols(southround4games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

southround4 <- lightgbm_fit |> predict(new_data = southround4games, type="prob") |>
  bind_cols(southround4) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

southround4

midwestround4games <- tibble(
  team_short_display_name="Samford",
  opponent_team_short_display_name="Tennessee"
)

midwestround4games <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(midwestround4games)

midwestround4games <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(midwestround4games) 

midwestround4 <- lightgbm_fit |> predict(new_data = midwestround4games) |>
  bind_cols(midwestround4games) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

midwestround4 <- lightgbm_fit |> predict(new_data = midwestround4games, type="prob") |>
  bind_cols(midwestround4) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

midwestround4

finalfourgames <- tibble(
  team_short_display_name="UConn",
  opponent_team_short_display_name="Arizona"
) |> add_row(
  team_short_display_name="Houston",
  opponent_team_short_display_name="Tennessee"
) 

finalfourgames <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(finalfourgames)

finalfourgames <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(finalfourgames) 

finalfour <- lightgbm_fit |> predict(new_data = finalfourgames) |>
  bind_cols(finalfourgames) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

finalfour <- lightgbm_fit |> predict(new_data = finalfourgames, type="prob") |>
  bind_cols(finalfour) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

finalfour

nationalchampionshipgame <- tibble(
  team_short_display_name="UConn",
  opponent_team_short_display_name="Houston"
)

nationalchampionshipgame <- modelgames |> 
  group_by(team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("opponent")) |> 
  right_join(nationalchampionshipgame)

nationalchampionshipgame <- modelgames |> 
  group_by(opponent_team_short_display_name) |> 
  filter(game_date == max(game_date) & season == 2024) |> 
  ungroup() |> 
  select(-team_result, -starts_with("team"), -game_id, -game_date, -season) |> 
  right_join(nationalchampionshipgame) 

nationalchampionship <- lightgbm_fit |> predict(new_data = nationalchampionshipgame) |>
  bind_cols(nationalchampionshipgame) |> select(.pred_class, team_short_display_name, opponent_team_short_display_name, everything())

nationalchampionship <- lightgbm_fit |> predict(new_data = nationalchampionshipgame, type="prob") |>
  bind_cols(nationalchampionship) |> select(.pred_class, .pred_W, .pred_L, team_short_display_name, opponent_team_short_display_name, everything())

nationalchampionship

Using data from hoopR and lightGBM as a predictor, our class built prediction models that we could customize to what we think gives us the best shot at predicting the most NCAA Tournament games correctly. In my model, I added offensive and defensive efficiency, score margin, steals, free throws made, rebounds and a few more stats that I thought are especially important to helping a team win a basketball game. Some of these stats were cumulative to the whole season, and some were on a rolling basis of a team’s last 5 games.

I had only ever filled out brackets using my judgement, so I was very excited to watch the tournament this year after I built and machine learning bracket.

Let’s take a look at a regional:

Code 
southround1 %>% 
  select(team_short_display_name, .pred_class, .pred_W, opponent_team_short_display_name) %>%
  gt() %>% 
  cols_label(
    team_short_display_name = "Team",
    .pred_class = "Prediction",
    .pred_W = "Win Confidence",
    opponent_team_short_display_name = "Opponent"
  ) %>%
  tab_header(
    title = "South Regional: Round 1",
    subtitle = "I had the James Madison and Colorado upsets correct!"
  ) %>%  
  tab_source_note(
    source_note = md("**By:** Greg Johnson")
  ) %>% 
  tab_style(
    style = cell_text(color = "black", weight = "bold", align = "left"),
    locations = cells_title("title")
  ) %>% 
  tab_style(
    style = cell_text(color = "black", align = "left"),
    locations = cells_title("subtitle")
  ) %>%
  tab_style(
     locations = cells_column_labels(columns = everything()),
     style = list(
       cell_borders(sides = "bottom", weight = px(3)),
       cell_text(weight = "bold", size=12)
     )
   ) %>%
  opt_row_striping() %>% 
  opt_table_lines("none") %>%
    fmt_percent(
    columns = c(.pred_W),
    decimals = 1
  )
South Regional: Round 1
I had the James Madison and Colorado upsets correct!
Team Prediction Win Confidence Opponent
Houston W 81.5% Longwood
Wisconsin L 17.0% James Madison
Kentucky W 75.3% Oakland
Florida L 45.5% Colorado
Texas Tech W 54.6% NC State
Marquette W 50.5% Western KY
Nebraska W 54.5% Texas A&M
Duke W 65.0% Vermont
By: Greg Johnson

In the South Regional, I was surprised to see that James Madison had such a high percentage to beat Wisconsin, but they ended up taking care of business so that was great. I ended up having Colorado going to the Elite 8, so I was pretty hyped when they hit a last second shot to beat Florida.

It is not easy to pick a Cinderella team in the NCAA Tournament, but I should have given a better opportunity for NC State to get on a run. The Wolfpack won 5 games in 5 days in the ACC Conference Tournament, and is a basketball team the summary of its whole season or the summary of its last few games? They were rolling, and nobody could stop them until the Final 4.

Code 
southround2 %>% 
  select(team_short_display_name, .pred_class, .pred_W, opponent_team_short_display_name) %>%
  gt() %>% 
  cols_label(
    team_short_display_name = "Team",
    .pred_class = "Prediction",
    .pred_W = "Win Confidence",
    opponent_team_short_display_name = "Opponent"
  ) %>%
  tab_header(
    title = "South Regional: Round 2",
    subtitle = "My Colorado Elite 8 pick dies by only 4 points"
  ) %>%  
  tab_source_note(
    source_note = md("**By:** Greg Johnson")
  ) %>% 
  tab_style(
    style = cell_text(color = "black", weight = "bold", align = "left"),
    locations = cells_title("title")
  ) %>% 
  tab_style(
    style = cell_text(color = "black", align = "left"),
    locations = cells_title("subtitle")
  ) %>%
  tab_style(
     locations = cells_column_labels(columns = everything()),
     style = list(
       cell_borders(sides = "bottom", weight = px(3)),
       cell_text(weight = "bold", size=12)
     )
   ) %>%
  opt_row_striping() %>% 
  opt_table_lines("none") %>%
    fmt_percent(
    columns = c(.pred_W),
    decimals = 1
  )
South Regional: Round 2
My Colorado Elite 8 pick dies by only 4 points
Team Prediction Win Confidence Opponent
James Madison L 45.4% Duke
Texas Tech L 38.4% Kentucky
Colorado W 52.6% Marquette
Houston W 68.3% Nebraska
By: Greg Johnson

Here, Houston, my predicted tournament champion, barely escapes in overtime against Texas A&M, and my Colorado run comes to a close. Things are not looking good going forward, and sure enough, Houston falls to Duke in the Sweet 16 and NC State makes the Final 4.

On ESPN Bracket Challenge, my bracket was titled, “The B1G sucks at BB”. I am in Nebraska’s pep band, so I have been keeping a close eye on Big Ten Basketball for the last three years. Last year, only 1 out of 8 B1G teams in the tournament made the Sweet 16, so Purdue playing in the championship game, and Illinois making the Elite 8 was a big surprise to me. The most games I had a Big Ten team winning was Illinois at 2. I tried to make my model reflect things I notice and think about at games, such as fighting for rebounds and steals and making free throws. I should have broadened my horizons to watch more conferences to see what stats matter to the rest of the country.

In the end, this can be a very reliable way to pick a solid bracket, but I do not think predicting games solely on stats is going to get anyone a perfect bracket.