The week the model went 10 for 10
The make-cut accuracy on our golf history board read 72.0%. The honest number — counted only over tournaments where a cut actually happened — is 64.3%.
The gap comes from 38 of the 169 tournaments we graded between 2023 and mid-2026. At all 38, no player missed the cut. Every golfer in the field played four rounds. Our model still ranked the players and predicted who would advance, and it went 100% every time.
That performance and what we published are two different things. Here's a clear example of what actually happened.
After the FedEx St. Jude Championship ended in summer 2026, we published a recap of how the model had performed that week. One line read: "All 10 of the top 10 graded players made the cut."
The FedEx St. Jude Championship has no cut. Every golfer plays all four rounds. We had compared our top 10 ranked players to a field of 125 who all made it through the week. They were necessarily among the 125. A hundred percent accuracy, at an event where nobody could fail.
What a cut is and why it matters to grade
Most PGA Tour events work the same way. Around 150 golfers tee off Thursday morning. After two rounds, the bottom half go home. The remaining 65 or so — the low scorers, plus anyone within 10 shots of the leader — play the weekend. That line is the cut.
It's a natural prediction target. Before the event starts, rank every player in the field. After 36 holes, check which half advanced. Count how many of the model's top-ranked players were actually among the survivors. A binary outcome, a binary grade, checkable in the results.
A meaningful chunk of the tour schedule doesn't work that way. The three FedEx Cup playoff events — the late-season tournaments with the most points at stake — have no cut. Neither do most of the tour's signature events: smaller-field, invitation-only tournaments where the top players in the world show up and everyone stays for Sunday. At those events, there's nothing to predict. The grading system didn't know the difference.
How the signal worked against us
We don't receive a direct flag in the data telling us whether a specific tournament has a cut. Nothing in the golf results data identifies an event as no-cut.
So the cut signal was inferred from outcomes: a player who made the cut played at least three rounds of golf. That's accurate in the forward direction — someone who missed the cut went home before their third round. The trouble is what happens in reverse.
At a no-cut event, every golfer in the field plays four rounds. Under the inferred rule, every single one of them made the cut. When the model ranked 125 players at St. Jude and we went to check accuracy, we compared our highest-ranked 65 to a pool of 125 players who all "made it" — because all 125 did. The top 65 were necessarily among them. A perfect score, by construction.
The Travelers Championship ran without a cut that same year. Same result.
Thirty-eight events. All perfect.
Thirty-eight of the 169 tournaments in the graded sample had no cut we could measure. All three FedEx playoff events each year. Most of the signature events — the ones with elite invitations and no qualifying pressure to send half the field home.
Each one registered a free 100% in the cut-accuracy column. When those were averaged with the 131 events where the model was actually being tested, the blended result landed at 72.0%.
Remove the 38 and the number is 64.3%. Those 8 points weren't the model performing better at prestigious events. They were the model not being graded at all.
We published that number alongside the top-5 and top-10 finish rates, which were calculated correctly against the same events. The cut column wasn't.
Telling the two apart without a label
The fix needed a way to separate no-cut events from real ones without being told directly — which we weren't.
Withdrawals turned out to be the signal. At a real cut, somewhere between 15% and 55% of the starting field doesn't make the weekend. That's the full range across the 131 measurable events in the sample. The tightest real cut — the Arnold Palmer Invitational in 2024 — had 11 of 69 graded players fall short: 15.9%. The widest was an event where just over half the field was sent home.
No-cut events look completely different. The only players who don't complete four rounds are those who withdraw through injury or personal reasons. Across the full no-cut sample, the highest that share ever reached was 7.7% — nine players at a qualifying school event with an unusual structure that caused a run of withdrawals. Most no-cut events land close to zero.
Ten percent sits cleanly in the gap between 7.7% and 15.9%.
A tournament where fewer than one in ten players stops short of a full week gets flagged as unmeasurable: no grade, drops from the averages. St. Jude 2026 had exactly one withdrawal across 125 starters — less than 1% of the field. It now shows a "no cut" note on the history board instead of a 100%. So do the other 37.
The published history reads 64.3%, covering only the tournaments where the model was actually being tested.
What this doesn't say
It doesn't say the model's cut predictions are wrong. 64.3% is the real number, measured across 131 events where a cut happened and we could grade it honestly. What that compares to — how often a random player ranking would put the right half through at baseline — is a separate question and a separate piece.
Two situations still break the signal, and both get silence rather than a published number.
One is a cut that falls after three rounds instead of two. At events like The American Express, the bottom half of the field is trimmed after 54 holes of play. Every player who eventually misses that cut still played three rounds. Under the inferred rule, they read as survivors. It can't be seen through automatically, and known events with this format are handled directly.
The other is a total round count that isn't four. A qualifying school runs six rounds with a cut after four, and the three-round threshold wasn't built for that.
In both cases, the failure mode is silence. A cut we can't verify gets no grade. That's deliberate: the threshold between "no cut" and "real cut" was set at 10% so that the safe direction to fail is toward publishing nothing. A real cut misidentified as unmeasurable costs one data point we decline to claim. The reverse — publishing a number that means nothing — is what ran on the history board for two years before we fixed it.
Sources
Every number here comes from the graded ledger the boards publish. Research from a statistical model, not betting advice; no outcome is guaranteed. Method: how MatchWiz works.