Dispersion.Golf
Live — Pre

How the numbers are made — and gated.

The quant desk for golf prediction markets. Transparent, validated, contract-native. Here is the whole pipeline, including its limitations.

01
Data
Public per-round scoring data across four tours — PGA Tour, DP World Tour, Korn Ferry Tour, and LIV (2022–present, 210,000+ player-rounds) — and live contract prices from Kalshi's public market-data API, including order-book depth. Cross-tour rounds connect the difficulty surface, so players arriving from other tours carry real skill estimates instead of a generic prior. No paid data feeds.
02
Player skill model
Raw scores are adjusted for course and round difficulty with a two-way decomposition, then converted to per-player skill estimates using recency weighting (half-life ≈ 29 rounds) with empirical-Bayes shrinkage calibrated to the four-tour population's skill spread. Where public data is insufficient, we say so on the number rather than silently degrading: thin-history players are flagged and excluded from analysis.
03
Monte Carlo field simulation
Every tournament is simulated at least 10,000 times with correlated scoring: shared daily condition shocks, Thursday/Friday wave shocks, integer-stroke rounds (tie clusters at the cut line are real and priced), event-specific cut rules, and a final-round pressure adjustment near the lead. Live mode re-simulates the remainder of the tournament from the current leaderboard. Win, top-10, make-cut and matchup probabilities are read directly off the simulated paths.
04
Contract pricing
Edges are computed against executable prices — the ask when evaluating YES, the bid when evaluating NO — never the midpoint or last trade, and expected value is net of exchange fees. Apparent edges on sub-10¢ longshots are held to a stricter threshold: the favorite-longshot bias in prediction markets is well documented, and cheap contracts carry the worst fee-to-price ratios.
05
Statistical validation gates
No model version publishes until it passes the López de Prado harness on historical data: Deflated Sharpe Ratio above 95%, probability of backtest overfitting below 0.5, and out-of-sample performance thresholds — plus calibration (Spiegelhalter Z within ±2, published reliability curves) and Brier skill versus baseline with tournament-clustered significance. Golf offers ~45 tournaments a year — small-sample discipline is the whole game.
06
Immutable publication
Every probability shown on this site is simultaneously written to an append-only, hash-chained ledger with its timestamp, the live contract price, and the exact model version (MLflow reference) that produced it. The track record you see is the record — there is no other copy.
07
Honest limitations
  • Public data has per-round totals only — no shot-level strokes-gained decomposition, so ball-striking versus putting form cannot be separated.
  • Players without recent tour history (many major-championship qualifiers) get prior-dominated estimates and are excluded from analysis.
  • The market's closing price embeds injury and withdrawal news the model cannot see.
  • Team events, Stableford scoring, and 54-hole-cut formats are not modeled; the engine declines to produce numbers for them.

What we publish is model probability versus market price — impersonal analytics. We do not make picks, and nothing here is advice to buy or sell any contract.