Golf IQ
Win, top-5, top-10 and make-the-cut probabilities for PGA Tour events, from a simulation of the whole field. An experiment in how far round scores alone can go.
Key Features & Outcomes
- A rating from round scores only, time-decayed and shrunk towards a prior for players with few rounds, then 10,000 Monte Carlo simulations of each event
- Beats the 'everyone is equal' baseline by about 8% on win probability, held out from tuning: 136 events in 2023 to 2025 and 40 in 2026
- Checked for leakage two ways: rewriting the future with noise changed none of the earlier forecasts, and shuffled results scored worse than chance
- Honest about limits: never compared with bookmaker odds (bookmakers are much sharper), favourites look under-rated, and make-the-cut rates run low
The idea
Golf is hard to forecast because the winner is usually someone you would not have picked. The question was how much signal there is in round scores alone, without course, weather or form data, and whether a model can be honest about how little that is.
The build
A Python engine rates every player from field-relative round scores, adjusts for the strength of the events they played, and simulates the tournament 10,000 times to get win, top-5, top-10, top-20 and make-the-cut probabilities. The forecasts are frozen before each event and scored afterwards, so the model cannot be tuned in hindsight. A small Cloudflare Worker and D1 database publish them.
Several tuning experiments, including adding other tours' history, did not help and were dropped rather than kept. The first tournament to be scored with frozen forecasts is the next step. It is a hobby experiment, not a betting tool, and the public page is not promoted yet.