What we did
- Collected historical PGA tournament data, player stats, and event-level shot records.
- Generated forecasting questions anchored at pre-tournament and round-by-round timestamps.
- Labeled with the actual tournament outcome — cut made, winner, hole-in-one, head-to-head.
- Fine-tuned with RL and evaluated on 855 held-out golf questions.
Example datapoint
A sample training example — question, source, and outcome-derived label.
DATASET Tournament outcomes
RL
Question
Will world No. 1 Scottie Scheffler win the 2025 Players Championship outright given the Mar 6 draw, TPC Sawgrass firm-and-fast setup, and current Strokes Gained: Tee-to-Green rankings?
Question source
PGA Tour Mar 6, 2025
Weather window firms; firm conditions favor longer hitters off the tee
Label
No.
Type
binary
Confidence
Label source
Leaderboard Mar 9, 2025
Playoff decided on second extra hole — top seed eliminated in R3
Results
Benchmark comparisons against frontier models.
ProphetArena Sports Leaderboard
Foresight V1 32B ranks #1 on ProphetArena Sports ahead of Grok 4, Gemini 2.5 Pro, and multiple GPT-5 variants. On 855 held-out golf questions, it achieved Brier Skill Score +17.0% vs. +12.8% for GPT-5, with 41% lower calibration error (ECE 0.062 vs. 0.106).
Read more
Papers, models, datasets, notebooks, and write-ups for this case study.
Blog post reddit.com
Golf forecaster beats GPT-5
Reddit thread with community discussion and full methodology breakdown
Blog post news.ycombinator.com
Hacker News discussion
HN discussion on the golf forecasting approach and results
Notebook github.com
Golf forecasting fine-tune
Step-by-step notebook for training a golf outcome predictor
Model huggingface.co
Golf-Forecaster
Fine-tuned model for golf tournament outcome prediction
Dataset huggingface.co
Golf forecasting dataset
Tournament questions paired with resolved finishing positions