It’s that time of the year again–MARCH MADNESS.

A few of us graduate students decided it would be fun to simulate this year’s NCAA Men’s National Championship tournament using R. When we all met to discuss our method to simuluate tournament performance, we quickly realized that each of us had a different method in mind. These methods differed in not only how we determined probability of win via seed, but also the techniques and packages implemented within R. Below showcases 3 different methods we used, each with different results. The methods are arranged in order of increasing use of historical data.

Matt’s Method


The packages needed for this method:

library(tidyverse); library(repmis); library(RCurl)

Simulation

This method (that Nick actually invented) relies on the seed of each team in the tournament. The idea is to assign a range of numbers to each seed with the ranges of lower seeds (e.g., #1, #2, #3, …) containing greater numbers than the ranges of higher seeds (e.g., …, #14, #15, #16). Just like this:

Seed Lower.Bound Upper.Bound
1 160 320
2 150 310
3 140 300
4 130 290
5 120 280
6 110 270
7 100 260
8 90 250
9 80 240
10 70 230
11 60 220
12 50 210
13 40 200
14 30 190
15 20 180
16 10 170

Then, randomly sample from each team’s range and assign the win to the higher number (for each match-up). This is a crude method that will roughly correspond to seed. In other words, each seed has a higher probability of winning over each subsequent seed; however, it is theoretically possible, although very unlikely, for a 16th seed to win the tournament. A simulation given purely historical data would not be able to account for this because a 16th seed has never made it past the first round prior to the 2018 tournament.

Below is the code I used to run the simulation. I tried to avoid for loops as much as possible and learn the map() functions in the purrr package. The code is not as efficient as I would have liked it to be, but it gets the job done and allows for a variety of plots.

# Sets the regions
regions <- c('South', 'East', 'West', 'Midwest')

# Import the team names and their seeds from 2018
# These data are being pulled from this website's Github repo
link <- 'https://raw.githubusercontent.com/mkmiecik14/mkmiecik14.github.io/master/data/ncaa-2018-teams-updated.csv'
ncaa.2018 <- read_csv(getURL(link)) %>% 
  rename(region = Region, team = Team1, seed = Seed)

# Initializations
iters <- 1000
results <- probTableHolder <- vector('list', length = iters)

# Sets seed for reprodicible results
set.seed(18)

The simulation was then ran 1000 times:

for(i in 1:iters) {

  # Creates whole tournament probability set
  probTableHolder[[i]] <- probTable <- regions %>% 
    map_df(~tibble(seed = c(1:16),
                   lowNum = seq(160, 10, -10),
                   highNum = seq(320, 170, -10)) %>%
             group_by(seed) %>% 
             mutate(result = sample(lowNum:highNum, 1)) %>%
             ungroup()) %>%
    mutate(region = rep(regions, each = 16)) %>%
    inner_join(ncaa.2018, by = c('seed', 'region'))
  
  # Round 1
  r1 <- regions %>%
    map_df(~tibble(match.up = 1:8, t1.seed = 1:8, t2.seed = 16:9) %>%
             mutate(t1.res =  probTable$result[1:8],
                    t2.res = probTable$result[16:9],
                    seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
             mutate(region = rep(regions, each = 8)) %>%
             inner_join(probTable, by = c('seed', 'region'))
    
  # Round 1 results
  r1.res <- r1 %>% select(match.up, region, team, seed, result)
  
  # Round 2
  r2 <- regions %>%
    map_df(~tibble(r2.match.up = 1:4,
                r2.match.up.1 = 1:4,
                r2.match.up.2 = as.integer(seq(8, 5, -1)),
                region = .) %>%
          left_join(r1.res, 
                    by = c('r2.match.up.1' = 'match.up', 
                           'region' = 'region')) %>%
          rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
          left_join(r1.res,
                    by = c('r2.match.up.2' = 'match.up', 
                           'region' = 'region')) %>%
          rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
          mutate(seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
    inner_join(probTable, by = c('region', 'seed'))
  
  # Round 2 results
  r2.res <- r2 %>% select(r2.match.up, region, team, seed, result)
  
  # Round 3
  r3 <- regions %>%
    map_df(~tibble(r3.match.up = 1:2, 
                r3.match.up.1 = 1:2, 
                r3.match.up.2 = 4:3,
                region = .) %>%
          left_join(r2.res, 
                    by = c('r3.match.up.1' = 'r2.match.up', 
                           'region' = 'region')) %>%
          rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
          left_join(r2.res,
                    by = c('r3.match.up.2' = 'r2.match.up', 
                           'region' = 'region')) %>%
          rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
          mutate(seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
    inner_join(probTable,  by = c('region', 'seed'))
  
  # Round 3 results
  r3.res <- r3 %>% select(r3.match.up, region, team, seed, result)
  
  # Round 4
  r4 <- regions %>%
    map_df(~tibble(r4.match.up = 1, 
                   r4.match.up.1 = 1, 
                   r4.match.up.2 = 2,
                   region = .) %>%
             inner_join(r3.res, 
                        by = c('r4.match.up.1' = 'r3.match.up', 
                               'region' = 'region')) %>%
             rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
             inner_join(r3.res, by = c('r4.match.up.2' = 'r3.match.up',
                                       'region' = 'region')) %>%
             rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
             mutate(seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
    inner_join(probTable,  by = c('region', 'seed'))
  
  # Round 4 results
  r4.res <- r4 %>% select(r4.match.up, region, team, seed, result)
  
  # Semifinal
  semi.final <- tibble(t1.region = c('South', 'East'),
                       t2.region = c('West', 'Midwest')) %>%
    left_join(r4.res, by = c('t1.region' = 'region')) %>%
    rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
    left_join(r4.res, by = c('t2.region' = 'region')) %>%
    rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
    mutate(team = if_else(t1.res > t2.res, t1.team, t2.team)) %>%
    inner_join(probTable, by = 'team')
  
  # Semifinal results
  semi.final.res <- semi.final %>% select(region, team, seed, result)
  
  # Final
  final <- tibble(t1.region = semi.final.res$team[1],
                  t2.region = semi.final.res$team[2],
                  t1.team = semi.final.res$team[1],
                  t2.team = semi.final.res$team[2],
                  t1.seed = semi.final.res$team[1],
                  t2.seed = semi.final.res$team[2],
                  t1.res = semi.final.res$result[1],
                  t2.res = semi.final.res$result[2]) %>%
    mutate(team = if_else(t1.res > t2.res, t1.team, t2.team)) %>%
    inner_join(probTable, by = 'team')
  
  final.res <- final %>% select(region, team, seed, result)
  
  results[[i]] <- this.result <- bind_rows(r1.res, r2.res, r3.res, r4.res,
                                           semi.final.res, final.res, 
                                           .id = 'round')
  
  #print(i) # Uncomment to see progress
}

Data Viz

The results of this simulation were saved out and can be imported from this website’s Github data repository like this:

# Loading probability tables from github
probTableLink <- 'https://github.com/mkmiecik14/mkmiecik14.github.io/blob/master/data/probTableHolder-updated.RData?raw=true'
source_data(probTableLink)

# Loading simulation results from github
resultsLink <- 'https://github.com/mkmiecik14/mkmiecik14.github.io/blob/master/data/results-updated.RData?raw=true'
source_data(resultsLink)

First, let’s ensure that the probability table results worked across simulations and regions:

# Converts into a dataframe
probTable.long <- probTableHolder %>% map_df(bind_rows, .id = 'iter')

# Calculates summary stats
probTable.sum <- probTable.long %>% 
  group_by(seed, region) %>%
  summarize(m = mean(result),
            sd = sd(result),
            n = n(),
            sem = sd/sqrt(n))

# Probability table plot
ggplot(probTable.long, aes(factor(seed), result, color = region)) +
  geom_point(position = 'jitter', alpha = 1/3, shape = 1) +
  geom_pointrange(data = probTable.sum, 
                  aes(factor(seed), m, color = region, 
                      ymin = m - sd, ymax = m + sd),
                  color = 'black') +
  geom_line(data = probTable.sum, 
            aes(factor(seed), m, group = 1, color = region), 
            color = 'black', alpha = 1/2) +
  scale_color_brewer(palette = 'Set2', name = 'Region') +
  labs(x = 'Seed', y = 'Randomly Chosen Value', caption = 'SD Error Bars') +
  theme_minimal() +
  facet_wrap(~region)

This above visualization depicts the potential ranges from which each seed can draw an arbitrary number. This drawing occurs at the beginning of each simulation iteration and determines each team’s tournament fate. Higher seeds have an advantage of a higher possible range, but this overlap gradually decreases with increases in seed.

Now let’s examine the results of the simulations. As a check, better seeds should win with more frequency each round than worse seeds:

# Converts to long dataframe
results.long <- results %>% map_df(bind_rows, .id = 'iter')

# Win histogram
ggplot(results.long, aes(as.factor(seed), fill = region)) +
  geom_histogram(stat = 'count') +
  facet_grid(round~region) +
  scale_x_discrete(breaks = seq(1, 16, 3)) +
  labs(x = 'Seed', y = 'Win Frequency', title = 'Number of Wins Per Round') +
  scale_fill_brewer(palette = 'Set2', name = 'Region') +
  theme_minimal()

The above plot confirms this; higher seeds win more often than lower seeds for each round.

Now for fun, let’s see which teams won the tournament most often.

# Round 6 summary information
results.long.sum1 <- results.long %>% 
  filter(round == '6') %>% 
  group_by(team, region) %>%
  summarise(n = n(), seed = unique(seed))

# Winner plot
ggplot(results.long.sum1, aes(reorder(team, -1*n), n, fill = region)) +
  geom_bar(stat = 'identity') +
  scale_fill_brewer(palette = 'Set2', name = 'Region') +
  labs(x = 'Team', y = 'NCAA Tournament Wins', 
       caption = paste('Simulation iterations:', iters)) +
  geom_text(aes(label = seed), nudge_y = 4, color = 'grey') + 
  theme_minimal() +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

There are some clear, but expected, patterns that emerge given a simulation based solely on seed. First, #1 seeds tend to win more often than lower seeds. Second, win frequency tracks well with seed, but in a diminishing step-wise way. The margin is closer between #3 vs #4 seeds, than #2 vs #3 seeds, than #1 vs #2. Perhaps this has to do with the NCAA men’s tournament not re-seeding teams on a round-by-round basis (see this video by Jon Bois).

However, this method did allow for some upsets. As we can see, three #6 seeds won the tournament (TCU, Houston, Miami)!

So, does this analysis inform my bracket this year at all? Sort of, but I’d rather just have R do it for me.

I’ll just have the computer pick a random iteration that occured and I’ll use that to fill out my bracket:

# Random selection from the set of simulation iterations
# note: seed was set earlier in script
myPick <- sample(1:iters, 1)

# Matt's bracket from a random simulation iteration
mattsBracket <- results.long %>% 
  filter(iter == myPick) %>% 
  select(round, region, team, seed) %>%
  rename(winner = team)

So here’s what the computer drew for me (winners of each round):

round region winner seed
1 South Virginia 1
1 South Cincinnati 2
1 South Tennessee 3
1 South Arizona 4
1 South Kentucky 5
1 South Miami 6
1 South Nevada 7
1 South Kansas State 9
1 East Villanova 1
1 East Purdue 2
1 East Texas Tech 3
1 East Wichita State 4
1 East West Virginia 5
1 East Florida 6
1 East Arkansas 7
1 East Alabama 9
1 West Xavier 1
1 West North Carolina 2
1 West Michigan 3
1 West Gonzaga 4
1 West Ohio State 5
1 West Houston 6
1 West Texas A&M 7
1 West Florida State 9
1 Midwest Kansas 1
1 Midwest Duke 2
1 Midwest Michigan State 3
1 Midwest Auburn 4
1 Midwest Clemson 5
1 Midwest TCU 6
1 Midwest Rhode Island 7
1 Midwest NC State 9
2 South Kansas State 9
2 South Cincinnati 2
2 South Tennessee 3
2 South Arizona 4
2 East Villanova 1
2 East Purdue 2
2 East Texas Tech 3
2 East West Virginia 5
2 West Florida State 9
2 West North Carolina 2
2 West Michigan 3
2 West Ohio State 5
2 Midwest Kansas 1
2 Midwest Duke 2
2 Midwest Michigan State 3
2 Midwest Clemson 5
3 South Arizona 4
3 South Tennessee 3
3 East Villanova 1
3 East Texas Tech 3
3 West Ohio State 5
3 West Michigan 3
3 Midwest Kansas 1
3 Midwest Michigan State 3
4 South Tennessee 3
4 East Texas Tech 3
4 West Michigan 3
4 Midwest Michigan State 3
5 South Tennessee 3
5 Midwest Michigan State 3
6 Midwest Michigan State 3


Wish me luck!

Linh & David’s Method


Packages needed for this method:

library(tidyverse); library(repmis); library(RCurl)

Similar to Matt’s method, we also decided to go with a sliding range approach; however, we wanted to include historical win probabilities of first round play to determine the width of each team’s number range, as well as the overlap between seeds.

The Basics

We have the most historical data on the matchups in the first round because they happen four times each year. Therefore, it makes the most sense to use these data to predict each matchup, especially for future rounds where certain matchups are rare. To do this, we created a range of values for each team in such a way that when one number is drawn from that range and compared against the number generated for its first round opponent, the odds of that team’s number being higher than the opponent’s match the historical probability of that team winning. In other words, every team has a range of values that can be chosen from, and this range overlaps with every other team by a very specific amount.

Example 1 - Overlap of 2



If team 1 (favorite) has a range of possible values from 5-10, and team 2 (underdog) has a range of 1-6, randomly choosing a number for each team will result in team 1 winning almost every time. The odds of team 2 winning are equal to the odds of team 1 drawing a 5 and team 2 drawing a 6. This equates to a mathematical “and” problem that can be calculated by multiplying the odds of each event happening. Since each team has 6 numbers to draw from, the odds are 1/6 for each, or 1/36 (2.78%) for both happening at the same time.

Example 2 - Overlap of 4



If we use these same dimensions, but want to increase the odds of the underdog winning, we could increase the amount of numbers that overlap by one number. Now the favorite chooses from 4-10 and the underdog chooses from 1-7. Here, the odds of the underdog winning increase to 12.24% because they can win if they draw a 5, 6, or 7 because it is possible for the favorite to draw a 4, 5, or 6.

The Math

Legend for the following equations:

n = number of overlapping values
M = Maximum value possible
R = Size of range

To make this method work for our purposes, we created an algorithm that defines a range of values for both teams in every first round matchup such that the odds of randomly selecting a value within these ranges matches the historical odds of that team winning. To do this, we determined the formula of the underdog winning (i.e., the number of outcomes in which the underdog would win, divided by the formula for the total number of possible outcomes).

\[ \begin{aligned} \text{Probability of UnderdogWin} = \mathbf{\frac{Underdog Wins}{Total Outcomes}} \end{aligned} \]

\[ \begin{eqnarray} \mathbf{P} = \frac{(\frac{n^2 - n}{2})}{R^2} \Leftrightarrow \frac{(\frac{n^2 - n}{2})}{(\frac{M+n}{2})^2} \end{eqnarray} \]

When this equation is rearranged to solve for n (i.e., the range of numbers that needed to overlap for a given probability), you get a quadatratic equation that can be plugged into the quadratic formula.

\[ \begin{align*} & \mathbf{(P-2)n^2 + (2PM + 2)n + P(M^2)} \text{, where} \\ \\ & \mathbf{a = P - 2} \\ & \mathbf{b = 2PM + 2} \\ & \mathbf{c = P(M^2)} \end{align*} \]

Once n was determined, we needed to set the exact range for each team based on this overlap. To do this, we divided m by 2 and set half of n above and below this mid-point. Teams are assigned a value from this range in round 1, and then carry this value for each round to determine the winner of each matchup. Only in the event of a tie is a new number temporarily generated.

\[ \begin{align} \mathbf{UnderdogHigh} = \frac{M+n}{2} \\ \\ \mathbf{Favored Low} = \frac{M-n+2}{2} \end{align} \]

And that’s it. Plug this into Matt’s code for each round.

The Code

calcRanges <- function(UnderdogWinProb, MaxVal) {
    a <- UnderdogWinProb - 2
    b <- (2 * UnderdogWinProb * MaxVal) + 2
    c <- UnderdogWinProb * (MaxVal^2)
    
    quad_minus <- (-b - sqrt(b^2 - 4 * a * c)) / (2 * a)
    quad_plus <- (-b + sqrt(b^2 - 4 * a * c)) / (2 * a)
    
    if (isTRUE(sign(quad_plus) == 1)) {
        n_overlap <- quad_plus
    } else {
        n_overlap <- quad_minus
    }
    
    ## Round n_overlap to nearest even integer
    n_overlap <- 2 * round(n_overlap/2)
    
    UnderdogHigh <- (MaxVal + n_overlap) / 2
    FavoredLow <- ((MaxVal - n_overlap) / 2) + 1

    cat(paste0("UnderdogHigh Value: ", UnderdogHigh))
    cat(paste0(""),sep='\n')
    cat(paste0("FavoredLow Value: ", FavoredLow))
}

The Simulations

We’ve included two simulations:

  1. Implements historical match-up records of the first-round up to, but not including, the 2018 men’s NCAA tournament
  2. Implements all first-round match-up records that include this year’s (2018) tournament first round performance

Click on the separate tabs below to see each one in turn:

Data Prior to 2018

Here are the low and the high numbers matched with each seed:

lowNum <- c(4673, 3947, 3036, 2714, 1351, 1216, 1080, 76, 1, 1, 1, 1, 1, 1, 1, 1)
highNum <- c(10000, 10000, 10000, 10000, 10000, 10000, 10000, 10000,
             9925, 8921, 8785, 8650, 7287, 6965, 6054, 5328)

That correspond to seed like this:

seed lowNum highNum
1 4673 10000
2 3947 10000
3 3036 10000
4 2714 10000
5 1351 10000
6 1216 10000
7 1080 10000
8 76 10000
9 1 9925
10 1 8921
11 1 8785
12 1 8650
13 1 7287
14 1 6965
15 1 6054
16 1 5328
Simulation

This method’s simulation code is identical to Matt’s, except that we will exchange the arbitrary sliding range with these newly created ranges based on first round historical probabilities:

# Sets the regions
regions <- c('South', 'East', 'West', 'Midwest')

# Import the team names and their seeds from 2018
# These data are being pulled from this website's Github repo
link <- 'https://raw.githubusercontent.com/mkmiecik14/mkmiecik14.github.io/master/data/ncaa-2018-teams-updated.csv'
ncaa.2018 <- read_csv(getURL(link)) %>% 
  rename(region = Region, team = Team1, seed = Seed)

# Initializations
iters <- 1000
results <- probTableHolder <- vector('list', length = iters)

This simulation was also ran 1000 times:

for(i in 1:iters) {

  # Creates whole tournament probability set
  probTableHolder[[i]] <- probTable <- regions %>% 
    map_df(~tibble(seed = c(1:16),
                   lowNum = lowNum,
                   highNum = highNum) %>%
             group_by(seed) %>% 
             mutate(result = sample(lowNum:highNum, 1)) %>%
             ungroup()) %>%
    mutate(region = rep(regions, each = 16)) %>%
    inner_join(ncaa.2018, by = c('seed', 'region'))
  
  # Round 1
  r1 <- regions %>%
    map_df(~tibble(match.up = 1:8, t1.seed = 1:8, t2.seed = 16:9) %>%
             mutate(t1.res =  probTable$result[1:8],
                    t2.res = probTable$result[16:9],
                    seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
             mutate(region = rep(regions, each = 8)) %>%
             inner_join(probTable, by = c('seed', 'region'))
    
  # Round 1 results
  r1.res <- r1 %>% select(match.up, region, team, seed, result)
  
  # Round 2
  r2 <- regions %>%
    map_df(~tibble(r2.match.up = 1:4,
                r2.match.up.1 = 1:4,
                r2.match.up.2 = as.integer(seq(8, 5, -1)),
                region = .) %>%
          left_join(r1.res, 
                    by = c('r2.match.up.1' = 'match.up', 
                           'region' = 'region')) %>%
          rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
          left_join(r1.res,
                    by = c('r2.match.up.2' = 'match.up', 
                           'region' = 'region')) %>%
          rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
          mutate(seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
    inner_join(probTable, by = c('region', 'seed'))
  
  # Round 2 results
  r2.res <- r2 %>% select(r2.match.up, region, team, seed, result)
  
  # Round 3
  r3 <- regions %>%
    map_df(~tibble(r3.match.up = 1:2, 
                r3.match.up.1 = 1:2, 
                r3.match.up.2 = 4:3,
                region = .) %>%
          left_join(r2.res, 
                    by = c('r3.match.up.1' = 'r2.match.up', 
                           'region' = 'region')) %>%
          rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
          left_join(r2.res,
                    by = c('r3.match.up.2' = 'r2.match.up', 
                           'region' = 'region')) %>%
          rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
          mutate(seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
    inner_join(probTable,  by = c('region', 'seed'))
  
  # Round 3 results
  r3.res <- r3 %>% select(r3.match.up, region, team, seed, result)
  
  # Round 4
  r4 <- regions %>%
    map_df(~tibble(r4.match.up = 1, 
                   r4.match.up.1 = 1, 
                   r4.match.up.2 = 2,
                   region = .) %>%
             inner_join(r3.res, 
                        by = c('r4.match.up.1' = 'r3.match.up', 
                               'region' = 'region')) %>%
             rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
             inner_join(r3.res, by = c('r4.match.up.2' = 'r3.match.up',
                                       'region' = 'region')) %>%
             rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
             mutate(seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
    inner_join(probTable,  by = c('region', 'seed'))
  
  # Round 4 results
  r4.res <- r4 %>% select(r4.match.up, region, team, seed, result)
  
  # Semifinal
  semi.final <- tibble(t1.region = c('South', 'East'),
                       t2.region = c('West', 'Midwest')) %>%
    left_join(r4.res, by = c('t1.region' = 'region')) %>%
    rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
    left_join(r4.res, by = c('t2.region' = 'region')) %>%
    rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
    mutate(team = if_else(t1.res > t2.res, t1.team, t2.team)) %>%
    inner_join(probTable, by = 'team')
  
  # Semifinal results
  semi.final.res <- semi.final %>% select(region, team, seed, result)
  
  # Final
  final <- tibble(t1.region = semi.final.res$team[1],
                  t2.region = semi.final.res$team[2],
                  t1.team = semi.final.res$team[1],
                  t2.team = semi.final.res$team[2],
                  t1.seed = semi.final.res$team[1],
                  t2.seed = semi.final.res$team[2],
                  t1.res = semi.final.res$result[1],
                  t2.res = semi.final.res$result[2]) %>%
    mutate(team = if_else(t1.res > t2.res, t1.team, t2.team)) %>%
    inner_join(probTable, by = 'team')
  
  final.res <- final %>% select(region, team, seed, result)
  
  results[[i]] <- this.result <- bind_rows(r1.res, r2.res, r3.res, r4.res,
                                           semi.final.res, final.res, 
                                           .id = 'round')
  
  #print(i) # Uncomment to see progress
}
Data Viz

The results of this simulation were saved out and can be imported from this website’s Github data repository like this:

# Loading probability tables from github
probTableLink2 <- 'https://github.com/mkmiecik14/mkmiecik14.github.io/blob/master/data/probTableHolder-linh-david-before-2018.RData?raw=true'
source_data(probTableLink2)

# Loading simulation results from github
resultsLink2 <- 'https://github.com/mkmiecik14/mkmiecik14.github.io/blob/master/data/results-linh-david-before-2018.RData?raw=true'
source_data(resultsLink2)

First, let’s ensure that the probability table results worked across simulations and regions:

# Converts into a dataframe
probTable.long <- probTableHolder %>% map_df(bind_rows, .id = 'iter')

# Calculates summary stats
probTable.sum <- probTable.long %>% 
  group_by(seed, region) %>%
  summarize(m = mean(result),
            sd = sd(result),
            n = n(),
            sem = sd/sqrt(n))

# Probability table plot
ggplot(probTable.long, aes(factor(seed), result, color = region)) +
  geom_point(position = 'jitter', alpha = 1/3, shape = 1) +
  geom_pointrange(data = probTable.sum, 
                  aes(factor(seed), m, color = region, 
                      ymin = m - sd, ymax = m + sd),
                  color = 'black') +
  geom_line(data = probTable.sum, 
            aes(factor(seed), m, group = 1, color = region), 
            color = 'black', alpha = 1/2) +
  scale_color_brewer(palette = 'Set2', name = 'Region') +
  labs(x = 'Seed', y = 'Randomly Chosen Value', caption = 'SD Error Bars') +
  theme_minimal() +
  facet_wrap(~region)

This above visualization depicts the potential ranges each seed can draw from using first round historical win probabilities. This drawing occurs at the beginning of each simulation iteration and determines each team’s tournament performance.

Now let’s examine the results of the simulations.

# Converts to long dataframe
results.long <- results %>% map_df(bind_rows, .id = 'iter')

# Win histogram
ggplot(results.long, aes(as.factor(seed), fill = region)) +
  geom_histogram(stat = 'count') +
  facet_grid(round~region) +
  scale_x_discrete(breaks = seq(1, 16, 3)) +
  labs(x = 'Seed', y = 'Win Frequency', title = 'Number of Wins Per Round') +
  scale_fill_brewer(palette = 'Set2', name = 'Region') +
  theme_minimal()

Higher seeds tend to win more often than lower seeds for each round. In contrast to Matt’s simulation, there is more variability towards seed #7 and this variability becomes more pronounced with each subsequent round.

Now for fun, let’s see which teams won the tournament most often.

# Round 6 summary information
results.long.sum1 <- results.long %>% 
  filter(round == '6') %>% 
  group_by(team, region) %>%
  summarise(n = n(), seed = unique(seed))

# Winner plot
ggplot(results.long.sum1, aes(reorder(team, -1*n), n, fill = region)) +
  geom_bar(stat = 'identity') +
  scale_fill_brewer(palette = 'Set2', name = 'Region') +
  labs(x = 'Team', y = 'NCAA Tournament Wins', 
       caption = paste('Simulation iterations:', iters)) +
  geom_text(aes(label = seed), nudge_y = 4, color = 'grey') + 
  theme_minimal() +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

This simulation has Villanova (#1) with the most wins, while Purdue (#2) comes in second.

Interestingly, compared to Matt’s simulation, this method allows for more variability for higher seeds to win the tournament:

ggplot(results.long.sum1, aes(factor(seed), n)) +
  geom_bar(stat = 'identity') +
  labs(x = 'Seed', y = 'NCAA Tournament Wins',
       caption = paste('Simulation iterations:', iters)) +
  theme_minimal()

According to this method, #5 and #7 seeds won the tournament fewer times than #6 and #8 seeds. Nevertheless, this method results in several upsets winning the tournament, even as high as #9 seeds.

Data Including 2018

This section was added after observing the first round of the 2018 Men’s NCAA Tournament because, for the first time in history, a #16 seed beat a #1 seed. Therefore, we no longer had to guess this probability because it occured once out of 136 games–a probability of 0.0076.

The ranges for each team were adjusted to include data from the 2018 tournament:

Here are the low and the high numbers matched with each seed:

lowNum <- c(4678, 3966, 3077, 2639, 1444, 1182, 1116, 1, 1, 1, 1, 1, 1, 1, 1, 1)
highNum <- c(10000, 10000, 10000, 10000, 10000, 10000, 10000, 10000, 10000,
             8885, 8819, 8557, 7363, 6924, 6035, 5323)

That correspond to seed like this:

seed lowNum highNum
1 4678 10000
2 3966 10000
3 3077 10000
4 2639 10000
5 1444 10000
6 1182 10000
7 1116 10000
8 1 10000
9 1 10000
10 1 8885
11 1 8819
12 1 8557
13 1 7363
14 1 6924
15 1 6035
16 1 5323
Simulation

This method’s simulation code is idential to Matt’s, except that we will exchange the arbitrary sliding range with these newly created ranges based on first round historical probabilities:

# Sets the regions
regions <- c('South', 'East', 'West', 'Midwest')

# Import the team names and their seeds from 2018
# These data are being pulled from this website's Github repo
link <- 'https://raw.githubusercontent.com/mkmiecik14/mkmiecik14.github.io/master/data/ncaa-2018-teams-updated.csv'
ncaa.2018 <- read_csv(getURL(link)) %>% 
  rename(region = Region, team = Team1, seed = Seed)

# Initializations
iters <- 1000
results <- probTableHolder <- vector('list', length = iters)

This simulation was also ran 1000 times:

for(i in 1:iters) {

  # Creates whole tournament probability set
  probTableHolder[[i]] <- probTable <- regions %>% 
    map_df(~tibble(seed = c(1:16),
                   lowNum = lowNum,
                   highNum = highNum) %>%
             group_by(seed) %>% 
             mutate(result = sample(lowNum:highNum, 1)) %>%
             ungroup()) %>%
    mutate(region = rep(regions, each = 16)) %>%
    inner_join(ncaa.2018, by = c('seed', 'region'))
  
  # Round 1
  r1 <- regions %>%
    map_df(~tibble(match.up = 1:8, t1.seed = 1:8, t2.seed = 16:9) %>%
             mutate(t1.res =  probTable$result[1:8],
                    t2.res = probTable$result[16:9],
                    seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
             mutate(region = rep(regions, each = 8)) %>%
             inner_join(probTable, by = c('seed', 'region'))
    
  # Round 1 results
  r1.res <- r1 %>% select(match.up, region, team, seed, result)
  
  # Round 2
  r2 <- regions %>%
    map_df(~tibble(r2.match.up = 1:4,
                r2.match.up.1 = 1:4,
                r2.match.up.2 = as.integer(seq(8, 5, -1)),
                region = .) %>%
          left_join(r1.res, 
                    by = c('r2.match.up.1' = 'match.up', 
                           'region' = 'region')) %>%
          rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
          left_join(r1.res,
                    by = c('r2.match.up.2' = 'match.up', 
                           'region' = 'region')) %>%
          rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
          mutate(seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
    inner_join(probTable, by = c('region', 'seed'))
  
  # Round 2 results
  r2.res <- r2 %>% select(r2.match.up, region, team, seed, result)
  
  # Round 3
  r3 <- regions %>%
    map_df(~tibble(r3.match.up = 1:2, 
                r3.match.up.1 = 1:2, 
                r3.match.up.2 = 4:3,
                region = .) %>%
          left_join(r2.res, 
                    by = c('r3.match.up.1' = 'r2.match.up', 
                           'region' = 'region')) %>%
          rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
          left_join(r2.res,
                    by = c('r3.match.up.2' = 'r2.match.up', 
                           'region' = 'region')) %>%
          rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
          mutate(seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
    inner_join(probTable,  by = c('region', 'seed'))
  
  # Round 3 results
  r3.res <- r3 %>% select(r3.match.up, region, team, seed, result)
  
  # Round 4
  r4 <- regions %>%
    map_df(~tibble(r4.match.up = 1, 
                   r4.match.up.1 = 1, 
                   r4.match.up.2 = 2,
                   region = .) %>%
             inner_join(r3.res, 
                        by = c('r4.match.up.1' = 'r3.match.up', 
                               'region' = 'region')) %>%
             rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
             inner_join(r3.res, by = c('r4.match.up.2' = 'r3.match.up',
                                       'region' = 'region')) %>%
             rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
             mutate(seed = if_else(t1.res > t2.res, t1.seed, t2.seed))) %>%
    inner_join(probTable,  by = c('region', 'seed'))
  
  # Round 4 results
  r4.res <- r4 %>% select(r4.match.up, region, team, seed, result)
  
  # Semifinal
  semi.final <- tibble(t1.region = c('South', 'East'),
                       t2.region = c('West', 'Midwest')) %>%
    left_join(r4.res, by = c('t1.region' = 'region')) %>%
    rename(t1.team = team, t1.seed = seed, t1.res = result) %>%
    left_join(r4.res, by = c('t2.region' = 'region')) %>%
    rename(t2.team = team, t2.seed = seed, t2.res = result) %>%
    mutate(team = if_else(t1.res > t2.res, t1.team, t2.team)) %>%
    inner_join(probTable, by = 'team')
  
  # Semifinal results
  semi.final.res <- semi.final %>% select(region, team, seed, result)
  
  # Final
  final <- tibble(t1.region = semi.final.res$team[1],
                  t2.region = semi.final.res$team[2],
                  t1.team = semi.final.res$team[1],
                  t2.team = semi.final.res$team[2],
                  t1.seed = semi.final.res$team[1],
                  t2.seed = semi.final.res$team[2],
                  t1.res = semi.final.res$result[1],
                  t2.res = semi.final.res$result[2]) %>%
    mutate(team = if_else(t1.res > t2.res, t1.team, t2.team)) %>%
    inner_join(probTable, by = 'team')
  
  final.res <- final %>% select(region, team, seed, result)
  
  results[[i]] <- this.result <- bind_rows(r1.res, r2.res, r3.res, r4.res,
                                           semi.final.res, final.res, 
                                           .id = 'round')
  
  #print(i) # Uncomment to see progress
}
Data Viz

The results of this simulation were saved out and can be imported from this website’s Github data repository like this:

# Loading probability tables from github
probTableLink2 <- 'https://github.com/mkmiecik14/mkmiecik14.github.io/blob/master/data/probTableHolder-linh-david-including-2018.RData?raw=true'
source_data(probTableLink2)

# Loading simulation results from github
resultsLink2 <- 'https://github.com/mkmiecik14/mkmiecik14.github.io/blob/master/data/results-linh-david-including-2018.RData?raw=true'
source_data(resultsLink2)

First, let’s ensure that the probability table results worked across simulations and regions:

# Converts into a dataframe
probTable.long <- probTableHolder %>% map_df(bind_rows, .id = 'iter')

# Calculates summary stats
probTable.sum <- probTable.long %>% 
  group_by(seed, region) %>%
  summarize(m = mean(result),
            sd = sd(result),
            n = n(),
            sem = sd/sqrt(n))

# Probability table plot
ggplot(probTable.long, aes(factor(seed), result, color = region)) +
  geom_point(position = 'jitter', alpha = 1/3, shape = 1) +
  geom_pointrange(data = probTable.sum, 
                  aes(factor(seed), m, color = region, 
                      ymin = m - sd, ymax = m + sd),
                  color = 'black') +
  geom_line(data = probTable.sum, 
            aes(factor(seed), m, group = 1, color = region), 
            color = 'black', alpha = 1/2) +
  scale_color_brewer(palette = 'Set2', name = 'Region') +
  labs(x = 'Seed', y = 'Randomly Chosen Value', caption = 'SD Error Bars') +
  theme_minimal() +
  facet_wrap(~region)

This above visualization depicts the potential ranges each seed can draw from using first round historical win probabilities. This drawing occurs at the beginning of each simulation iteration and determines each team’s tournament performance.

Now let’s examine the results of the simulations.

# Converts to long dataframe
results.long <- results %>% map_df(bind_rows, .id = 'iter')

# Win histogram
ggplot(results.long, aes(as.factor(seed), fill = region)) +
  geom_histogram(stat = 'count') +
  facet_grid(round~region) +
  scale_x_discrete(breaks = seq(1, 16, 3)) +
  labs(x = 'Seed', y = 'Win Frequency', title = 'Number of Wins Per Round') +
  scale_fill_brewer(palette = 'Set2', name = 'Region') +
  theme_minimal()

Higher seeds tend to win more often than lower seeds for each round.

Now for fun, let’s see which teams won the tournament most often.

# Round 6 summary information
results.long.sum1 <- results.long %>% 
  filter(round == '6') %>% 
  group_by(team, region) %>%
  summarise(n = n(), seed = unique(seed))

# Winner plot
ggplot(results.long.sum1, aes(reorder(team, -1*n), n, fill = region)) +
  geom_bar(stat = 'identity') +
  scale_fill_brewer(palette = 'Set2', name = 'Region') +
  labs(x = 'Team', y = 'NCAA Tournament Wins', 
       caption = paste('Simulation iterations:', iters)) +
  geom_text(aes(label = seed), nudge_y = 4, color = 'grey') + 
  theme_minimal() +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

This simulation that incorporated the results from the first round of the 2018 men’s NCAA tournament had Virginia (#1) winning with Xavier (#1) coming in second in overall tournament wins across all 1000 simulations.

Interestingly, compared to Matt’s simulation, this method allows for more variability for higher seeds to win the tournament:

ggplot(results.long.sum1, aes(factor(seed), n)) +
  geom_bar(stat = 'identity') +
  labs(x = 'Seed', y = 'NCAA Tournament Wins',
       caption = paste('Simulation iterations:', iters)) +
  theme_minimal()

According to this method, overall wins track well with increasing seed. Lower seeds win more than higher seeds. However, upsets are definitely allowed to occur with #11 seeds winning the tournament! This is most likely a consequence of this year’s (2018) tournament containing a variety of upsets, namely a 16th seed over a 1st seed victory.

Nick’s Method


R packages needed for this method:

library(tidyverse); library(RCurl); library(repmis)

For my bracket, I wanted to match the probability of any given seed (seed X) beating any other given seed (seed Y) to the actual proportion of times that seed X has beaten seed Y in real life. For example, 15th seed teams have beaten 2nd seed teams 8 times in 132 games, or 6.1% of the time. Luckily for me, someone has already gone through every single game played and calculated the proportions for every match-up.

So, with the help of my friend Giuseppe Miranda, I created a table containing all of these proportions where rows represent winners and columns represent losers. So, if you want the proportion of times that 1st seed teams have beaten 16th seed teams, look at [row 1, column 16] and you will see that 1st seed teams have beaten 16th seed teams 132 times in 132 games, or 100% of the time (132/132 = 1). Likewise, if you go to [row 8, column 9] you can see that 8th seed teams have beaten 9th seed teams 53.2% (or 0.532) of the time.

These data are available on this website’s Github data repository and can be read into R like this:

# Table of probabilites to reference for each seed vs. seed match-up
propLink <- 'https://raw.githubusercontent.com/mkmiecik14/mkmiecik14.github.io/master/data/SeedMatchupProportions.csv'

ProportionsData <- read.table(text = getURL(propLink), sep = ',', 
                              row.names = 1, header = T)

The historical proportions data looks like this:

Seed Seed1 Seed2 Seed3 Seed4 Seed5 Seed6 Seed7 Seed8 Seed9 Seed10 Seed11 Seed12 Seed13 Seed14 Seed15 Seed16
Seed1 0.500 0.534 0.600 0.699 0.840 0.688 0.857 0.800 0.905 0.857 0.571 1.000 1.000 0.000 0.000 1.0
Seed2 0.466 0.500 0.629 0.444 0.200 0.722 0.710 0.444 0.500 0.607 0.867 1.000 0.000 0.000 0.939 0.0
Seed3 0.400 0.371 0.500 0.625 0.500 0.553 0.588 1.000 1.000 0.692 0.706 0.000 0.000 0.841 1.000 0.0
Seed4 0.301 0.556 0.375 0.500 0.565 0.333 0.333 0.364 0.666 1.000 0.000 0.692 0.809 0.000 0.000 0.0
Seed5 0.160 0.800 0.500 0.435 0.500 1.000 0.000 0.250 0.333 1.000 0.000 0.671 0.800 0.000 0.000 0.0
Seed6 0.313 0.278 0.447 0.667 0.000 0.500 0.625 0.333 0.000 0.600 0.638 0.000 0.000 0.875 0.000 0.0
Seed7 0.143 0.290 0.412 0.667 0.000 0.375 0.500 1.000 0.000 0.609 0.000 0.000 0.000 1.000 0.660 0.0
Seed8 0.200 0.556 0.000 0.636 0.750 0.750 0.500 0.500 0.532 0.000 1.000 0.000 1.000 0.000 0.000 0.0
Seed9 0.095 0.500 0.000 0.333 0.666 0.000 0.000 0.468 0.500 1.000 0.000 0.000 1.000 0.000 0.000 0.0
Seed10 0.143 0.393 0.308 0.000 0.000 0.400 0.391 0.000 0.000 0.500 0.333 0.000 0.000 1.000 1.000 0.0
Seed11 0.429 0.133 0.294 0.000 0.000 0.362 1.000 0.000 0.000 0.667 0.500 0.000 0.000 1.000 0.000 0.0
Seed12 0.000 0.000 0.000 0.308 0.329 0.000 0.000 1.000 0.000 0.000 0.000 0.500 0.727 0.000 0.000 0.0
Seed13 0.000 0.000 0.000 0.191 0.200 0.000 0.000 0.273 0.000 0.000 0.000 0.000 0.500 0.000 0.000 0.0
Seed14 0.000 0.000 0.159 0.000 0.000 0.125 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.500 0.000 0.0
Seed15 0.000 0.061 0.000 0.000 0.000 0.000 0.333 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.500 0.0
Seed16 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.5


Next, we will need a data frame for each round of the tournament, where each row represents each game played in the tournament. Each row will contain the name of the teams that are playing each other (‘Team1’ and ‘Team2’ columns) and the seed numbers of both of those teams (‘Team1Seed’ and ‘Team2Seed’). Later we will loop through these data frames to simulate each round of the tournament and predict the winner of each game.

For the first round, it was easier to create the data in excel beforehand and is available via .csv download from this website’s Github data repository or you can retrieve from R like this:

# We will load a data frame containing each match-up for the round of 64
r1Link <- 'https://raw.githubusercontent.com/mkmiecik14/mkmiecik14.github.io/master/data/NCAA2018Bracket.csv'

Round1Data <- read.table(text = getURL(r1Link), sep = ',', 
                         row.names = 1, header = T)

# We also need to coerce the Team Name columns into character vectors
Round1Data$Team1 <- as.character(Round1Data$Team1)
Round1Data$Team2 <- as.character(Round1Data$Team2)

Round1Data looks like this:

Region Team1 Team2 Team1Seed Team2Seed
1 South Virginia UMBC 1 16
2 South Creighton Kansas State 8 9
3 South Kentucky Davidson 5 12
4 South Arizona Buffalo 4 13
5 South Miami Loyola Chicago 6 11
6 South Tennessee Wright State 3 14
7 South Nevada Texas 7 10
8 South Cincinnati Georgia State 2 15
25 West Xavier Play in 4 1 16
26 West Missouri Florida State 8 9
27 West Ohio State South Dakota State 5 12
28 West Gonzaga UNC Greensboro 4 13
29 West Houston San Diego State 6 11
30 West Michigan Montana 3 14
31 West Texas A&M Providence 7 10
32 West North Carolina Lipscomb 2 15
9 East Villanova Play in 1 1 16
10 East Virginia Tech Alabama 8 9
11 East West Virginia Murray State 5 12
12 East Wichita State Marshall 4 13
13 East Florida Play in 2 6 11
14 East Texas Tech SF Austin 3 14
15 East Arkansas Butler 7 10
16 East Purdue CS Fullerton 2 15
17 Midwest Kansas Penn 1 16
18 Midwest Seton Hall NC State 8 9
19 Midwest Clemson New Mexico State 5 12
20 Midwest Auburn Charleston 4 13
21 Midwest TCU Play in 3 6 11
22 Midwest Michigan State Bucknell 3 14
23 Midwest Rhode Island Oklahoma 7 10
24 Midwest Duke Iona 2 15


But we can make the rest of the data frames for the remaining rounds in R:

Round2Data <- data.frame(game = 1:16, 
                         Region = rep(c('South', 'West', 'East', 'Midwest'), 
                                      each = 4),
                         Team1 = character(length = 16),
                         Team2 = character(length = 16),
                         Team1Seed = integer(length = 16),
                         Team2Seed = integer(length = 16),
                         Winner = character(length = 16),
                         stringsAsFactors = F)

Round3Data <- data.frame(game = 1:8, 
                         Region = rep(c('South', 'West', 'East', 'Midwest'), 
                                      each = 2),
                         Team1 = character(length = 8),
                         Team2 = character(length = 8),
                         Team1Seed = integer(length = 8),
                         Team2Seed = integer(length = 8),
                         Winner = character(length = 8),
                         stringsAsFactors = F)

Round4Data <- data.frame(game = 1:4, 
                         Region = c('South', 'West', 'East', 'Midwest'),
                         Team1 = character(length = 4),
                         Team2 = character(length = 4),
                         Team1Seed = integer(length = 4),
                         Team2Seed = integer(length = 4),
                         Winner = character(length = 4),
                         stringsAsFactors = F)

Round5Data <- data.frame(game = 1:2,
                         Team1 = character(length = 2),
                         Team2 = character(length = 2),
                         Team1Seed = integer(length = 2),
                         Team2Seed = integer(length = 2),
                         Winner = character(length = 2),
                         stringsAsFactors = F)

Round6Data <- data.frame(game = 1,
                         Team1 = character(length = 1),
                         Team2 = character(length = 1),
                         Team1Seed = integer(length = 1),
                         Team2Seed = integer(length = 1),
                         Winner = character(length = 1),
                         stringsAsFactors = F)

Simulation

We are then going to loop through our dataset for round 1 and calculate a score for each game played by randomly selecting a decimal number between 0 and 1. After we assign our score we will then refer to the cell of our ProportionsData, where row = seed of team 1 and column = seed of team 2. This cell should give us the proportion (in decimal form) of times that a team with a seed number equal to the seed of team 1 beat a team with a seed number equal to the seed of team 2. If our score variable is less than or equal to the proportion in this cell then team 1 wins, if not then team 2 wins. In other words, the probability of team 1 beating team 2 should be equivalent to the proportion of times that team 1’s seed beat team 2’s seed in real life.

## Round of 64##
print("Round of 64")
for (match in 1:nrow(Round1Data)){
  #Select 1 random decimal between 0 and 1.
  score <- runif(1,0,1)
  print(score)
  #Compare our 'score' variable to the real-life proportion data.
  #If score is less than the proportion of times the seed of team 1 
  #beat the seeed of team 2 then team 1 wins. Otherwise, team 2 wins.
  if (score <= ProportionsData[Round1Data$Team1Seed[match],
                               Round1Data$Team2Seed[match]]) {
    print(paste(Round1Data$Team1[match], " beats ", 
                Round1Data$Team2[match], sep = ""))
    Round1Data$Winner[match] <- Round1Data$Team1[match]
  }else {
    print(paste(Round1Data$Team2[match], " beats ", 
                Round1Data$Team1[match], sep = ""))
    Round1Data$Winner[match] <- Round1Data$Team2[match]
  }
}

Next, we will use the results from round 1 (we now have a winner column in our Round1Data filled in with the winner of each game) to fill in our data frame for round 2.

# Fill in the data frame for the next round using the results of the 
# previous round.
for (winner in 1:nrow(Round2Data)) {
  team1Winner <- as.character(Round1Data$Winner[winner*2-1])
  
  if (Round1Data$Winner[winner*2-1] %in% Round1Data$Team1){
    team1Seed <- Round1Data$Team1Seed[winner*2-1]
  }else {
    team1Seed <- Round1Data$Team2Seed[winner*2-1]
  }
  
  team2Winner <- Round1Data$Winner[winner*2]
  
  if (Round1Data$Winner[winner*2] %in% Round1Data$Team1){
    team2Seed <- Round1Data$Team1Seed[winner*2]
  }else {
    team2Seed <- Round1Data$Team2Seed[winner*2]
  }
  
  # I discovered that the team 1 needs to be the lower seed 
  # (i.e. higher seed number).
  if(team1Seed > team2Seed){
    Round2Data$Team1[winner] <- team1Winner
    Round2Data$Team1Seed[winner] <- team1Seed
    Round2Data$Team2[winner] <- team2Winner
    Round2Data$Team2Seed[winner] <- team2Seed
  } else {
    Round2Data$Team1[winner] <- team2Winner
    Round2Data$Team1Seed[winner] <- team2Seed
    Round2Data$Team2[winner] <- team1Winner
    Round2Data$Team2Seed[winner] <- team1Seed
  }
  
}

And finally, we’re simply going to repeat those last two steps for every remaining round in the tournament.

## Round of 32 ##
print("Round of 32")
for (match in 1:nrow(Round2Data)){
  score <- runif(1,0,1)
  print(score)
  if (score <= ProportionsData[Round2Data$Team1Seed[match],
                               Round2Data$Team2Seed[match]]) {
    print(paste(Round2Data$Team1[match], " beats ", 
                Round2Data$Team2[match], sep = ""))
    Round2Data$Winner[match] <- Round2Data$Team1[match]
  }else {
    print(paste(Round2Data$Team2[match], " beats ", 
                Round2Data$Team1[match], sep = ""))
    Round2Data$Winner[match] <- Round2Data$Team2[match]
  }
}

# Fill in the data frame for the next round using the results of the 
# previous round
for (winner in 1:nrow(Round3Data)) {
  team1Winner <- as.character(Round2Data$Winner[winner*2-1])
  
  if (Round2Data$Winner[winner*2-1] %in% Round2Data$Team1){
    team1Seed <- Round2Data$Team1Seed[winner*2-1]
  }else {
    team1Seed <- Round2Data$Team2Seed[winner*2-1]
  }
  
  team2Winner <- Round2Data$Winner[winner*2]
  
  if (Round2Data$Winner[winner*2] %in% Round2Data$Team1){
    team2Seed <- Round2Data$Team1Seed[winner*2]
  }else {
    team2Seed <- Round2Data$Team2Seed[winner*2]
  }
  
  # I discovered that the team 1 needs to be the lower seed 
  # (i.e. higher seed number)
  if(team1Seed > team2Seed){
    Round3Data$Team1[winner] <- team1Winner
    Round3Data$Team1Seed[winner] <- team1Seed
    Round3Data$Team2[winner] <- team2Winner
    Round3Data$Team2Seed[winner] <- team2Seed
  } else {
    Round3Data$Team1[winner] <- team2Winner
    Round3Data$Team1Seed[winner] <- team2Seed
    Round3Data$Team2[winner] <- team1Winner
    Round3Data$Team2Seed[winner] <- team1Seed
  }
  
}


## Round of 16 ##
print("Sweet 16")
for (match in 1:nrow(Round3Data)){
  score <- runif(1,0,1)
  print(score)
  if (score <= ProportionsData[Round3Data$Team1Seed[match],
                               Round3Data$Team2Seed[match]]) {
    print(paste(Round3Data$Team1[match], " beats ", 
                Round3Data$Team2[match], sep = ""))
    Round3Data$Winner[match] <- Round3Data$Team1[match]
  }else {
    print(paste(Round3Data$Team2[match], " beats ", 
                Round3Data$Team1[match], sep = ""))
    Round3Data$Winner[match] <- Round3Data$Team2[match]
  }
}

# Fill in the data frame for the next round using the results of the 
# previous round
for (winner in 1:nrow(Round4Data)) {
  team1Winner <- as.character(Round3Data$Winner[winner*2-1])
  
  if (Round3Data$Winner[winner*2-1] %in% Round3Data$Team1){
    team1Seed <- Round3Data$Team1Seed[winner*2-1]
  }else {
    team1Seed <- Round3Data$Team2Seed[winner*2-1]
  }
  
  team2Winner <- Round3Data$Winner[winner*2]
  
  if (Round3Data$Winner[winner*2] %in% Round3Data$Team1){
    team2Seed <- Round3Data$Team1Seed[winner*2]
  }else {
    team2Seed <- Round3Data$Team2Seed[winner*2]
  }
  
  # I discovered that the team 1 needs to be the lower seed 
  # (i.e. higher seed number)
  if(team1Seed > team2Seed){
    Round4Data$Team1[winner] <- team1Winner
    Round4Data$Team1Seed[winner] <- team1Seed
    Round4Data$Team2[winner] <- team2Winner
    Round4Data$Team2Seed[winner] <- team2Seed
  } else {
    Round4Data$Team1[winner] <- team2Winner
    Round4Data$Team1Seed[winner] <- team2Seed
    Round4Data$Team2[winner] <- team1Winner
    Round4Data$Team2Seed[winner] <- team1Seed
  }
  
}


## Round of 8 ##
print("Elite 8")
for (match in 1:nrow(Round4Data)){
  score <- runif(1,0,1)
  print(score)
  if (score <= ProportionsData[Round4Data$Team1Seed[match],
                               Round4Data$Team2Seed[match]]) {
    print(paste(Round4Data$Team1[match], " beats ", 
                Round4Data$Team2[match], sep = ""))
    Round4Data$Winner[match] <- Round4Data$Team1[match]
  }else {
    print(paste(Round4Data$Team2[match], " beats ", 
                Round4Data$Team1[match], sep = ""))
    Round4Data$Winner[match] <- Round4Data$Team2[match]
  }
}

# Fill in the data frame for the next round using the results of the 
# previous round
for (winner in 1:nrow(Round5Data)) {
  team1Winner <- as.character(Round4Data$Winner[winner*2-1])
  
  if (Round4Data$Winner[winner*2-1] %in% Round4Data$Team1){
    team1Seed <- Round4Data$Team1Seed[winner*2-1]
  }else {
    team1Seed <- Round4Data$Team2Seed[winner*2-1]
  }
  
  team2Winner <- Round4Data$Winner[winner*2]
  
  if (Round4Data$Winner[winner*2] %in% Round4Data$Team1){
    team2Seed <- Round4Data$Team1Seed[winner*2]
  }else {
    team2Seed <- Round4Data$Team2Seed[winner*2]
  }
  
  # I discovered that the team 1 needs to be the lower seed 
  # (i.e. higher seed number)
  if(team1Seed > team2Seed){
    Round5Data$Team1[winner] <- team1Winner
    Round5Data$Team1Seed[winner] <- team1Seed
    Round5Data$Team2[winner] <- team2Winner
    Round5Data$Team2Seed[winner] <- team2Seed
  } else {
    Round5Data$Team1[winner] <- team2Winner
    Round5Data$Team1Seed[winner] <- team2Seed
    Round5Data$Team2[winner] <- team1Winner
    Round5Data$Team2Seed[winner] <- team1Seed
  }
  
}


## Semi-Finals ##
print("Final Four")
for (match in 1:nrow(Round5Data)){
  score <- runif(1,0,1)
  print(score)
  if (score <= ProportionsData[Round5Data$Team1Seed[match],
                               Round5Data$Team2Seed[match]]) {
    print(paste(Round5Data$Team1[match], " beats ", 
                Round5Data$Team2[match], sep = ""))
    Round5Data$Winner[match] <- Round5Data$Team1[match]
  }else {
    print(paste(Round5Data$Team2[match], " beats ", 
                Round5Data$Team1[match], sep = ""))
    Round5Data$Winner[match] <- Round5Data$Team2[match]
  }
}

# Fill in the data frame for the next round using the results of the 
# previous round
for (winner in 1:nrow(Round6Data)) {
  team1Winner <- as.character(Round5Data$Winner[winner*2-1])
  
  if (Round5Data$Winner[winner*2-1] %in% Round5Data$Team1){
    team1Seed <- Round5Data$Team1Seed[winner*2-1]
  }else {
    team1Seed <- Round5Data$Team2Seed[winner*2-1]
  }
  
  team2Winner <- Round5Data$Winner[winner*2]
  
  if (Round5Data$Winner[winner*2] %in% Round5Data$Team1){
    team2Seed <- Round5Data$Team1Seed[winner*2]
  }else {
    team2Seed <- Round5Data$Team2Seed[winner*2]
  }
  
  # I discovered that the team 1 needs to be the lower seed 
  # (i.e. higher seed number)
  if(team1Seed > team2Seed){
    Round6Data$Team1[winner] <- team1Winner
    Round6Data$Team1Seed[winner] <- team1Seed
    Round6Data$Team2[winner] <- team2Winner
    Round6Data$Team2Seed[winner] <- team2Seed
  } else {
    Round6Data$Team1[winner] <- team2Winner
    Round6Data$Team1Seed[winner] <- team2Seed
    Round6Data$Team2[winner] <- team1Winner
    Round6Data$Team2Seed[winner] <- team1Seed
  }
  
}


## Finals ##
print("National Final")
score <- runif(1,0,1)
print(score)
if (score <= ProportionsData[Round6Data$Team1Seed[1],
                             Round6Data$Team2Seed[1]]) {
  print(paste(Round6Data$Team1[1],  " beats ", 
              Round6Data$Team2[1], " in the Final Round!", sep = ''))
  Round6Data$Winner[1] <- Round6Data$Team1[1]
}else {
  print(paste(Round6Data$Team2[1],  " beats ", 
              Round6Data$Team1[1], " in the Final Round!", sep = ''))
  Round6Data$Winner[1] <- Round6Data$Team2[1]
}

The above steps were repeated 1000 times to simulate performance.

Data Viz

The results from this simulation were saved into a variable named resultsHolder that can be retrieved from this website’s Github data repository like this:

nickResLink <- 'https://github.com/mkmiecik14/mkmiecik14.github.io/blob/master/data/resultsHolder.RData?raw=true'
source_data(nickResLink)

Let’s see what teams won the tournament most often:

# Converting list to long data frame
resultsLong <- resultsHolder %>% map_df(bind_rows, .id = 'iter')

# Summary information for results
resultsLongSum <- resultsLong %>% 
  filter(round == 6) %>%
  group_by(Winner) %>% 
  summarise(n = n(), seed = unique(WinnerSeed))

# Tournament champion histogram
ggplot(resultsLongSum, aes(reorder(Winner, -1*n), n)) +
  geom_bar(stat = 'identity') +
  labs(x = 'Tournament Champion', 
       y = 'Wins',
       caption = '"Play in #" teams were not available at time of simulation') +
  theme_minimal() +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

And it looks like this simulation predicted Xavier to win most often!

Conclusion


Of course, much more goes into tournament performance than seed placement and historical win probabilities. At the time that we first wrote this, a #16 seed had never beaten a #1 seed in the first round. Models using purely historical data (such as Nick’s) were unable to account for #16 seed UMBC’s historical first round victory over #1 seed Virginia in this year’s NCAA Men’s tournament. Nevertheless, this was a super fun exercise in R and thinking about modeling tournament style performance. Hope you enjoyed this post and any feedback is always welcome!

Acknowledgments


A huge thanks to my friends Nick Ray, David Hoagey, Linh Lazarus, Giuseppe Miranda, and Ekarin Pongpipat for their enthusiasm and help with this blog post. They helped turn a small idea into an awesome collaboration where we got to share thoughts, work off each other’s ideas, and learn more about R!