CS271 ML project to predict outcome of nba game
to generate data use run raw_script.py, edit season and output_file in main function

Game-Level Feature Engineering Overview

The goal of this pipeline is to transform raw per-team per-game NBA data into a fully leak-free feature set that supports:

• Game-level win prediction models (XGBoost, Logistic Regression)
• Opponent-aware modeling (pre-game matchup strength)
• Sequence models (RNNs / LSTMs reading team trajectories game-by-game)

This new version is fully redesigned for per-game prediction, ensuring:

• No label leakage (features use only games before the current game)
• Time-aware features (chronological sorting, game index)
• Opponent pre-game strength (their own season-to-date form)
• Short-term form and volatility (rolling windows)
• Rest/fatigue modeling (days since last game)

---

1. Input CSV Structure (Raw Stats)

Each seasonal input file (e.g., nba_team_stats_2020_21.csv) contains one row per team per game with:

• Identifiers: season, game_id, game_date, team_id, team_name, opponent_team_id, home, win
• Team advanced stats: offensive/defensive/net rating, eFG%, TS%, TOV%, ORB%, FTr, pace
• Opponent advanced stats: same fields prefixed with opp_

These are current-game stats collected after the game, so they cannot be used directly as features.

---

2. Chronological Ordering & Game Index

All games are sorted by: season → team_id → game_date

A per-team game_index is assigned:

• 1 = first game of season
• Increases sequentially for each team

This gives models awareness of season progression, letting them learn differences between early-season and late-season performance.

---

3. Convert Raw Stats to Game-Level Differentials

The script computes internal team minus opponent metrics:

• eFG_diff, TS_diff, TOV_diff, ORB_diff, FTr_diff
• net_rating_diff

These reflect performance gaps but are used only inside rolling/expanding windows.
They are never fed directly into the model (to avoid leakage).

---

4. Season-to-Date Features (Shifted = Leak-Free)

For each team and season, the script computes expanding means of:

• team_offensiveRating, team_defensiveRating, team_netRating
• eFG_diff, TOV_diff, ORB_diff, FTr_diff
• team_pace
• win → becomes win_pct_todate

All are shifted by 1 game:

For game t, features summarize games 1..(t–1) only.

This produces long-term form features that are 100% pre-game.

---

5. Rolling Windows (Short-Term Form & Volatility)

Using 5-game and 10-game windows, the script computes leak-free rolling:

• Means: team_netRating_last5, win_last10, eFG_diff_last10
• Standard deviations: performance volatility, e.g.
  team_netRating_std_last10, eFG_diff_std_last5

All rolling calculations are shifted so they contain only prior games.

These features capture:

• Momentum
• Slumps
• Consistency vs volatility

---

6. Rest / Fatigue Metrics

For each team:

days_since_last_game = difference in days from previous game

This encodes:

• Back-to-backs
• Long rest periods
• Fatigue effects

---

7. Opponent Pre-Game Strength

For every game, the opponent’s own season-to-date features are merged in:

• opp_netRating_avg_todate
• opp_win_pct_todate

Then matchup feature gaps are computed:

• matchup_net_rating_gap
• matchup_win_pct_gap

This lets the model compare teams based on pre-game strength, not final-season stats.

---

8. Output Datasets

After running the script:

• game_features_raw.csv — full leak-free dataset
• X_game.npy — standardized numerical features
• y_game.npy — per-team game outcome (0/1)
• game_feature_columns.txt — names of input features for modeling

These outputs are directly compatible with:

• Logistic Regression
• XGBoost / Random Forest
• Neural networks
• RNN/LSTM team-sequence modeling

---

Summary

This pipeline converts raw NBA per-game stats into a robust, leak-free dataset suitable for game-level prediction.
It includes temporal structure, opponent context, short-term form, long-term form, rest effects, and matchup strength — all computed using only information available before each game.