The generate_features_teamwins function reads the games table populated by fetch_games_teamwins, computes per-team rolling averages over a configurable window, and writes the results to a team_game_stats table.
Function signature
generate_features_teamwins(rolling_window: int = 5)
Parameters
| Parameter | Type | Default | Description |
|---|
rolling_window | int | 5 | Number of most recent games to average over when computing rolling statistics. |
Usage
from Data.generate_features import generate_features_teamwins
# Use the default 5-game rolling window
generate_features_teamwins()
# Use a shorter window for more sensitivity to recent form
generate_features_teamwins(rolling_window=3)
# Use a longer window to smooth out variance
generate_features_teamwins(rolling_window=10)
A larger rolling window smooths out game-to-game variance and captures longer-term team quality. A smaller window makes features more sensitive to recent form. Start with the default of 5 and adjust based on model validation performance.
Rolling features
For each of the following raw stats, a rolling mean is computed per team using the last rolling_window games (with min_periods=1 so early-season rows are still included):
for stat in ['team_points', 'opponent_points', 'team_reb', 'opponent_reb', 'team_ast', 'opponent_ast']:
df[f'{stat}_roll'] = (
df.groupby('team')[stat]
.rolling(rolling_window, min_periods=1)
.mean()
.reset_index(0, drop=True)
)
| Feature column | Description |
|---|
team_points_roll | Rolling average of points scored by the team. |
opponent_points_roll | Rolling average of points scored by the opponent. |
team_reb_roll | Rolling average of team rebounds. |
opponent_reb_roll | Rolling average of opponent rebounds. |
team_ast_roll | Rolling average of team assists. |
opponent_ast_roll | Rolling average of opponent assists. |
Derived feature: points_diff
After computing rolling averages, points_diff is derived as:
df['points_diff'] = df['team_points_roll'] - df['opponent_points_roll']
team_game_stats table schema
The output table is fully replaced on each run (if_exists='replace'):
CREATE TABLE IF NOT EXISTS team_game_stats (
game_id TEXT PRIMARY KEY,
team TEXT,
opponent TEXT,
points_diff REAL,
elo_diff REAL,
home INTEGER,
win INTEGER,
team_points_roll REAL,
opponent_points_roll REAL,
team_reb_roll REAL,
opponent_reb_roll REAL,
team_ast_roll REAL,
opponent_ast_roll REAL
)
| Column | Type | Description |
|---|
game_id | TEXT | NBA game identifier (primary key). |
team | TEXT | Team abbreviation. |
opponent | TEXT | Opponent abbreviation. |
points_diff | REAL | team_points_roll minus opponent_points_roll. |
elo_diff | REAL | Elo rating difference (currently 0, reserved for future use). |
home | INTEGER | 1 if the team played at home, 0 if away. |
win | INTEGER | 1 if the team won, 0 if they lost. |
team_points_roll | REAL | Rolling average points scored by the team. |
opponent_points_roll | REAL | Rolling average points scored by the opponent. |
team_reb_roll | REAL | Rolling average team rebounds. |
opponent_reb_roll | REAL | Rolling average opponent rebounds. |
team_ast_roll | REAL | Rolling average team assists. |
opponent_ast_roll | REAL | Rolling average opponent assists. |
The team_game_stats table is fully replaced every time generate_features_teamwins runs. You do not need to manually clear it before re-running.
Pipeline order
generate_features_teamwins reads directly from the games table. Run fetch_games_teamwins first to ensure the source data is up to date:
from Data.fetch_games import fetch_games_teamwins
from Data.generate_features import generate_features_teamwins
fetch_games_teamwins(season="2025-26")
generate_features_teamwins(rolling_window=5)