dfs_nba.rmd

---
title: "NBA Daily Fantasy"
output: pdf_document
---

```{r}
library(plyr)
library(TTR)
library(tree)

mv_avg_stats <- c("MIN", "FG", "FGA", "X3P", "X3PA", "FT", "FTA",
           "OR", "DR", "TOT", "A", "PF", "ST", "TO", "BL", "PTS")

mv_avg_colnames <- function(stats, k) {
  paste(stats, "_avg_past_", k, sep = "")
}

mv_avg_window <- 20

# TODO Match up player and team datasets.
nba_player <- read.csv("2016/player.csv")
nba_player$DATE <- as.Date(strptime(nba_player$DATE, f = "%m/%d/%Y"))
nba_team <- read.csv("2016/team.csv")
nba_team$DATE <- as.Date(strptime(nba_team$DATE, f = "%m/%d/%Y"))

mov_avg <- function(dest_df, source_df, stats, k) {
  dest_colnames <- mv_avg_colnames(stats, k)
  n <- min(k, nrow(source_df) - 1)
  for (i in seq(1, length(stats))) {
    s <- stats[i]
    # Weight of current stat is 0 as we only want historical stats.
    wts <- c(rep(1, n), 0)
    dest_df[, dest_colnames[i]] <- WMA(source_df[, s], n = n+1, wts = wts)
  }
  return(dest_df)
}

indiv_train <- ddply(nba_player, .(PLAYER.FULL.NAME), function(df) {
  # Sort by date for easy tabulation of moving averages.
  df <- df[order(df$DATE), ]
  
  # Features
  train_df <- data.frame(venue = df$VENUE..R.H., date = df$DATE)
  train_df$team <- df$OWN.TEAM
  train_df$opp <- df$OPP.TEAM
  train_df$pos <- df$POSITION
  train_df$days_rest <- c(0, diff(df$DATE))
  train_df <- mov_avg(train_df, df, mv_avg_stats, mv_avg_window)
  
  # Response variables
  train_df$resp_pts <- df$PTS
  train_df$resp_3pts <- df$X3P
  train_df$resp_rebounds <- df$TOT
  train_df$resp_assists <- df$A
  train_df$resp_steals <- df$ST
  train_df$resp_blocks <- df$BL
  train_df$resp_turnovers <- df$TO
  double_stats <- c("PTS", "TOT", "A", "ST", "BL")
  train_df$resp_doubles <- rowSums(df[, double_stats] >= 10)
  return(train_df)
})
# NAs created by moving average for past k games (if not enough data)
indiv_train <- na.omit(indiv_train)

# Get aggregate stats of other players.
add_players_stats <- function(curr_train, join_x_only, join_y_only, join_both) {
  join_x <- c(join_x_only, join_both)
  join_y <- c(join_y_only, join_both)
  id_frame <- indiv_train[, c("PLAYER.FULL.NAME", join_x)]
  # Hacky. We remove team so that it doesnt get duplicated in the merged dataframe.
  if (join_y_only == "opp") {
    join_frame <- indiv_train[, -which(names(indiv_train) %in% c("team"))]
  } else {
    join_frame <- indiv_train  
  }
  other_train <- merge(id_frame, join_frame, by.x = join_x, by.y = join_y)
  other_train <- other_train[with(other_train,
                                  PLAYER.FULL.NAME.x != PLAYER.FULL.NAME.y), ]
  
  other_stats <- c("days_rest", 
                   mv_avg_colnames(c("FG", "FGA", "X3P", "X3PA",
                                     "FT", "FTA", "OR", "DR", "TOT", "A",
                                     "PF", "ST", "TO", "BL", "PTS"), mv_avg_window))
  
  other_train <- ddply(other_train,
                       .(PLAYER.FULL.NAME.x, team, date), function(df) {
    apply(df[, other_stats], 2, function(col) {
      weighted.mean(col, w = df[, mv_avg_colnames("MIN", mv_avg_window)])
    })
  })

  all_train <- merge(curr_train, other_train,
                   by.x = c("PLAYER.FULL.NAME", join_x),
                   by.y = c("PLAYER.FULL.NAME.x", join_x))
  return(all_train)
}

all_train <- add_players_stats(indiv_train, "team", "team", "date")
all_train <- add_players_stats(all_train, "team", "opp", "date")

# TODO Verify that join is correct.
team_agg_stats <- c("TEAMS", "DATE", "POSS", "PACE", "OEFF", "DEFF")
team_agg_frame <- nba_team[, team_agg_stats]
all_train <- merge(all_train, team_agg_frame,
                   by.x = c("team", "date"), by.y = c("TEAMS", "DATE"))
all_train <- merge(all_train, team_agg_frame,
                   by.x = c("opp", "date"), by.y = c("TEAMS", "DATE"))

tree.pts <- tree(resp_pts ~ . -
                   team -
                   resp_rebounds - resp_steals - resp_assists - resp_blocks -
                   resp_doubles - resp_3pts - resp_turnovers -
                   PLAYER.FULL.NAME - date,
                 data = all_train, mindev = 0.001)
set.seed(1337)
pts.cv <- cv.tree(tree.pts)
opt.tree <- which(pts.cv$dev == min(pts.cv$dev))
best.leaves <- min(pts.cv$size[opt.tree])
pts.pruned = prune.tree(tree.pts, best=best.leaves)
infos <- summary(pts.pruned)

plot(pts.pruned)
text(pts.pruned)
sqrt(infos$dev / infos$df)

lm.pts <- lm(resp_doubles ~ . -
               team -
               resp_rebounds - resp_steals - resp_assists - resp_blocks -
               resp_pts - resp_3pts - resp_turnovers -
               PLAYER.FULL.NAME - date,
             data = all_train)
summary(lm.pts)$sigma

# Neural networks
```