This project is born out of my weathered relationship with the Chicago Bulls. They've had the reputation of only "showing up" for big games for a few seasons now and after some motivation I decided to try my hand at putting data around the question of how teams perform, based on their opponent's record.
So far it's just a modest data transformation effort, taking the game level data from basketball reference and storing each game with a cumulative tally of wins, losses, and win percentage for each team. Naturally, it took longer than I anticipated.
Basketball Reference is a fantastic source of hoops info. I couldn't find what I needed in the format I wanted but I found all the pieces needed to make it happen. The first step was pulling down all the data for the 2016-17 NBA season in a CSV.
- Adding GameID column, column renaming, dropping unnecessary columns
- Adding columns for calculated data
- Winner
Winner - Point Differential
PT_Diff
- Winner
- Create restructured temporary data frame for home and away teams to combine into one master list
- This included game level data (from source data) for each team
- Adding columns to temporary data frame for columns calculated by game by team
- Whether it was a win or a loss
W/L(Win = 1, Loss = 0)
- Whether it was a win or a loss
- Combining home and away data frames into a single data frame containing two records (home and away team) for each game with some additional calculated fields.
- Cumulative sum of wins
Wins - Cumulative sum of games played
GP OpponentandOpponent_PCT
- Cumulative sum of wins
The resultant data frame contains data for all 1230 games played in the 2016-17 season. For each game, it retains the initial information (date, time, home and away team, whether it went into overtime). It also includes the new data of the point differential, and the critical points for this analysis:
- Record of both teams at that point in the season
PCT - Outcome of that game for both teams, whether it was a win or loss
W/L
Developer Note: The record PCT at each game was changed to exclude the outcome of that game in the calculation at v 1.0.1.
Didn't see an immediate need for record at each game including the outcome of that game but this can be obtained from the helpers.season_totals
| Basketball Reference Top 10 | Resultant Top 10 |
|---|---|
 |
 |
Note: This script runs in about 15 seconds, compiling a season's worth of games.
After implementing the performance function, I was able to calculate the record for teams against opponents who were above
and below 0.500. This led to the discovery that the Chicago Bulls actually perform slightly better than 0.500 against below 0.500
teams, challenging my initial assumption. This will be something I will explore more thoroughly with a few seasons worth of data as
the initial data set was restricted to the 2016-17 season.
The issue with the PCT being calculated after the game was played cropped up earlier than I would have thought in that I was not able to reconcile wins and loss counts vs above and below 0.500 teams with the data on ESPN.com. This was related to calculation errors resolved in v 1.0.1 but also ESPN calculates record vs above/below 500 teams based on the end of season record and not record going into the game. This is an important distinction that this data adds to analysis as we couldn't make gametime predictions based on a team's end of season record.
- Clean up file structure
- Unit tests
- Coding style tests
- Look for relationships between team performance and opponent performance
- Consider storing season ids, date/time in other data frames.
- Attempt to fit (hidden) markov model to performance data for a team with the observation being W/L and the hidden state being opponent win pct.
Nothing special needed for installation.
python 2.7
pandas
- channiemills - Initial work - channiemills
- People that believe I can do anything