Q4 Possession Dataset for Elite Women's Basketball: 2025 Continental Championships and 2026 World Cup Qualifiers

```This dataset contains 6,512 verified fourth-quarter (Q4) possession records from elite women's basketball tournaments, extracted using our verified automatic possession segmentation pipeline achieving ≥95% accuracy. COVERAGE:- EuroBasket 2025: 36 games, 1,301 possessions- AmeriCup 2025: 32 games, 1,211 possessions - AfroBasket 2025: 28 games, 1,049 possessions- Asia Cup 2025: 20 games, 741 possessions- World Cup Qualifier 2026: 15 games, 2,210 possessionsTotal: 131 games, 6,512 Q4 possessions DATA STRUCTURE (13 fields): Temporal & Contextual (5 fields):- start_time_sec: Possession start time (seconds remaining in Q4)- start_score_diff: Score differential (offense score - defense score)- defensive_foul_count: Defensive team fouls at possession start- offensive_foul_count: Offensive team fouls at possession start- offense_from_backcourt: Binary (1=started from backcourt, 0=frontcourt) Strategic Features (5 fields):- first_shot_points: Point value of first shot attempt (0/1/2/3)- timeout_count: Timeouts called during possession- offensive_foul_drawn_count: Defensive fouls drawn by offense- offensive_rebound_count: Offensive rebounds during possession- points_scored: Possession outcome (0-5 points) Identification (3 fields):- game_name: Tournament and competition name- game_no: Unique game identifier (1-131)- possession_id: Possession sequence number within game QUALITY ASSURANCE:- Extraction accuracy: ≥95% validated against manual expert annotation- Self-verification: 4 independent constraints enforced (score consistency, quarter-score consistency, free-throw consistency, possession-count symmetry)- All constraint violations flagged and re-processed through LLM arbitration CLUTCH-TIME SUBSET:563 possessions (8.6%) meet clutch criteria (time ≤ 300s, |score_diff| ≤ 5), spanning 34 clutch-game situations. USE CASES:- Validating possession segmentation methods- Clutch-time performance analysis across confederations- Machine learning model training for possession outcome prediction- Gender-comparative basketball research- Benchmarking automated extraction pipelines DATA FORMAT:- File: Excel (.xlsx) with 13 columns- Encoding: UTF-8- Missing values: 0 for numerical fields, empty for missing identifiers EXTRACTION METHOD:7-stage pipeline combining:1. Dual-path event extraction (template parser + LLM schema extraction)2. Event cleaning and sorting3. Candidate boundary identification (5 types)4. Rule-based arbitration (3 checks)5. LLM arbitration for ambiguous cases (Claude-3.5-Sonnet)6. Possession annotation (offense/defense roles, termination type)7. Multi-constraint verification with rollback mechanism CITATION:Mulin Yang, Dandan Cui, Yuqiao Qian, and Wenchao Yang. 2027. From Public Play-by-Play to a Deployed Clutch-Possession Retrieval System: Verified Possession Extraction and an Open Dataset for Women's Basketball. In Proceedings of the 33rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '27). ACM, New York, NY, USA. CONTACT:[e-mail]