Elo Rating for NCAA 2018 March Madness Predictions¶

Importing necessary libraries¶

In [1]:

import numpy as np
import pandas as pd
from sklearn.metrics import log_loss

Setting K value for elo formula and setting the home advantage increase on elo score¶

In [2]:

K = 20.
HOME_ADVANTAGE = 100.

Loading game results and filtering data to only grab games after 2001¶

In [3]:

compact = pd.read_csv("./Documents/R Scripts/March Madness Submission Data/RegularSeasonCompactResults.csv")
compact = compact.loc[compact['Season'] >= 2002]

Creating dictionary for elo ratings per team¶

In [4]:

team_ids = set(compact.WTeamID).union(set(compact.LTeamID))
elo_dict = dict(zip(list(team_ids), [1500] * len(team_ids)))

Calculating point margin per game¶

In [5]:

compact['spread'] = compact.WScore - compact.LScore
compact['w_elo'] = None
compact['l_elo'] = None

Defining elo formulas¶

In [6]:

def elo_pred(elo1, elo2):
    return(1. / (10. ** (-(elo1 - elo2) / 400.) + 1.))

def expected_margin(elo_diff):
    return((7.5 + 0.006 * elo_diff))

def elo_update(w_elo, l_elo, margin):
    elo_diff = w_elo - l_elo
    pred = elo_pred(w_elo, l_elo)
    mult = ((margin + 3.) ** 0.8) / expected_margin(elo_diff)
    update = K * mult * (1 - pred)
    return(pred, update)

Running elo functions for all regular season games and updating elo¶

In [7]:

preds = []

for i in range(69493, (69493 + compact.shape[0])):
    
    # Get key data from current row
    w = compact.at[i, 'WTeamID']
    l = compact.at[i, 'LTeamID']
    margin = compact.at[i, 'spread']
    wloc = compact.at[i, 'WLoc']
    
    # Does either team get a home-court advantage?
    w_ad, l_ad, = 0., 0.
    if wloc == "H":
        w_ad += HOME_ADVANTAGE
    elif wloc == "A":
        l_ad += HOME_ADVANTAGE
    
    # Get elo updates as a result of the game
    pred, update = elo_update(elo_dict[w] + w_ad,
                              elo_dict[l] + l_ad, 
                              margin)
    elo_dict[w] += update
    elo_dict[l] -= update
    preds.append(pred)

    # Stores new elos in the games dataframe
    compact.loc[i, 'w_elo'] = elo_dict[w]
    compact.loc[i, 'l_elo'] = elo_dict[l]

Gathering only the final elo for each team in each season¶

Since the primary purpose of elo rating is to determine the march madness tournament results before the tournament begins.

In [8]:

def final_elo_per_season(df, team_id):
    d = df.copy()
    d = d.loc[(d.WTeamID == team_id) | (d.LTeamID == team_id), :]
    d.sort_values(['Season', 'DayNum'], inplace=True)
    d.drop_duplicates(['Season'], keep='last', inplace=True)
    w_mask = d.WTeamID == team_id
    l_mask = d.LTeamID == team_id
    d['season_elo'] = None
    d.loc[w_mask, 'season_elo'] = d.loc[w_mask, 'w_elo']
    d.loc[l_mask, 'season_elo'] = d.loc[l_mask, 'l_elo']
    out = pd.DataFrame({
        'team_id': team_id,
        'season': d.Season,
        'season_elo': d.season_elo
    })
    return(out)

Creating a csv of the Season, TeamID, and Elo Rating to be used in prediction modeling¶

In [9]:

df_list = [final_elo_per_season(compact, i) for i in team_ids]
season_elos = pd.concat(df_list)

season_elos.to_csv("elos_season.csv", index=None)

Elo Rating for NCAA 2018 March Madness Predictions¶

Importing necessary libraries¶

Setting K value for elo formula and setting the home advantage increase on elo score¶

Loading game results and filtering data to only grab games after 2001¶

Creating dictionary for elo ratings per team¶

Calculating point margin per game¶

Defining elo formulas¶

Running elo functions for all regular season games and updating elo¶

Gathering only the final elo for each team in each season¶

Creating a csv of the Season, TeamID, and Elo Rating to be used in prediction modeling¶

Click here to see the March Madness prediction model.