HSDSC NBA Framing
2024-10-07
HS DSC Basketball play-by-play
Task: Based on a set of play-by-play data from a full regular season… predict the wining percentage of a set of playoff match ups.
Step 1: set of ~30 game winning %. Same as 2023,
games are Home-Away and neutral site.
Step 2: presentation deck with methods, etc
(same as HSDSC 2023)
File size:
- teams = 28.
- games / seasons = 68.
- possessions / game ~200.
- events / possession ~2.
number of rows (events): 7.616^{5}
Simplified play-by-play data
type of events from 130 => ~20
Example data: minimum number of fields
| id | game_id | event_seq | type_text | score_value | scoring_play | shooting_play | period_number | clock | time_s | coordinate_y | coordinate_x |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 4014425354 | 401442535 | 1 | Jumpball | 0 | FALSE | FALSE | 1 | 12:00 | 720 | 214748365 | 214748406.75 |
| 4014425357 | 401442535 | 2 | Shooting Foul | 0 | FALSE | FALSE | 1 | 11:46 | 706 | 1 | -36.75 |
| 4014425359 | 401442535 | 3 | Free Throw - 1 of 2 | 1 | TRUE | TRUE | 1 | 11:46 | 706 | 0 | 28.00 |
| 40144253510 | 401442535 | 4 | Free Throw - 2 of 2 | 1 | TRUE | TRUE | 1 | 11:46 | 706 | 0 | 28.00 |
| 40144253511 | 401442535 | 5 | Floating Jump Shot | 2 | TRUE | TRUE | 1 | 11:33 | 693 | 9 | -36.75 |
| 40144253513 | 401442535 | 6 | Turnaround Fade Away Jump Shot | 0 | FALSE | TRUE | 1 | 11:11 | 671 | 6 | 27.75 |
| 40144253514 | 401442535 | 7 | Defensive Rebound | 0 | FALSE | FALSE | 1 | 11:09 | 669 | -6 | -27.75 |
| 40144253515 | 401442535 | 8 | Jump Shot | 0 | FALSE | TRUE | 1 | 11:00 | 660 | 14 | -19.75 |
| 40144253516 | 401442535 | 9 | Defensive Rebound | 0 | FALSE | FALSE | 1 | 10:57 | 657 | -14 | 19.75 |
| 40144253517 | 401442535 | 10 | Pullup Jump Shot | 2 | TRUE | TRUE | 1 | 10:52 | 652 | 11 | 26.75 |
play-by-play model:
event level is sub-possession level
Inputs: Team parameters + Schedule + play-by-play model + ______ + _______
Input 1: Team parameters (season averages?)
Use observed data as starting point, expand observed distributions (maybe)
- FGA rates (FGA_2; FGA_3).
- Blocks & Turn overs.
- Shooting % (FGA_2; FGA_3; FT).
- Rebounding (OREF; DREB).
(last 3 seasson of nba teams)
Input 2: Team schedules (similar to 2023 NSL)
- HomeTeam; AwayTeam; Location
- no temporal information
- no dates, or game sequence
- no changed in team quality over the course of the season
- Unbalanced schedule:
- two conferences with limited cross-conference play
- uneven team quality between conferences
Input 3: delta time = f(event type)
Problems:
My “simple” play-by-play model doesn’t capture needed features. It will regenerate season averages (if I get the timing resolved). Points for-against is all that is needed for predictions.
- no HFA.
- end of game chaos.
- game state changes
- substitutions (less subs in the playoffs).
Input 3: Changes in play
from season average rates per team to team rates that change based on:
- Home - Away.
- Game state: {nuetral; big lead; garbage time}
- Previous possession (in bounded ball; everything else)
Too easy version of the solution: points for - points against is pretty good
There is so much information in points
HSDSC 2023 WH got pretty close on the HFA distance effect.
