The First Number Looked Clean. The Second One Asked a Question: An 11,400-Shot Audit of TPC Boston

On the media-centre wall at TPC Boston a graphic was running on loop — Justin Thomas, 17-under, and beneath it a glowing line: best putting week of the season. On my laptop a far less attractive file was open. The raw export of four days of tracked shots: the coordinate of every shot, the lie code of every shot, and the operator ID of the person who recorded it. That putting line was nowhere in the file. What was in the file was this: the approach dispersion at TPC Boston that week was 2.3 strokes wider than the tour average.

The contest being sold on television as a putting contest was, in the tracked record, an approach contest. The first number looked clean. The second one asked a question.

The chain of custody

Start where every golf number should start and almost never does: who recorded it. This piece is not built on a press release or a wire copy. Three documents underpin it — the ShotLink shot-level export from the September 2026 Dell Technologies Championship, the PGA TOUR scoring file, and my own field notebook. Where I cannot reopen the export and show the figure, I will not print the figure.

The context matters. The Dell Technologies Championship was the second FedExCup Playoffs event, played over Labor Day week, with a field of 96. TPC Boston, Norton, Massachusetts — par 71, roughly 7,200 yards, Arnold Palmer's original design reworked by Gil Hanse in 2026. The event was dropped from the schedule after 2026, which makes this a note passed back through a door that has since closed.

What ShotLink is needs saying plainly, because plenty of people treat it as an app. It is a crew. A laser operator behind each green converts the ball's landing point into coordinates on a course map. A walking scorer with each group records the lie — rough, fairway, bunker, green. The data flows to a trailer and then to a server. Our crew that week was fourteen people, exactly two of them women. Fourteen-hour days, four days.

My presence on that crew was not the data point, and my reporting is not made valid by my presence — the two things are separate and should stay separate. I came off the sponsorship-research side, talked my way into a data role at a Waltham analytics vendor, and took the on-site assignment nobody wanted. Before my first on-air segment a producer told me to "get someone who actually plays." I had arrived with the model already built. Justifying your presence costs more time than justifying your maths; that was the week's most useful lesson.

What 11,400 actually means

A number that travels without its denominator is not a number, it is a slogan. So: 11,400 is not every shot in the field. It is the subset the model rested on — tee shots and approaches, the full-swing shots. Ninety-six players, four rounds, roughly thirty full-swing shots each per round. The putts were a separate file.

The strokes-gained framework comes from Mark Broadie's work — Every Shot Counts, 2026 — where each shot is measured against the tour baseline expectation for that distance and lie. That baseline is built from millions of tour shots. I took the same baseline, computed the expected strokes for each shot, and summed the differences. It is long work and there is no glamour in it, which is exactly why I can say where a figure came from and where it stops holding.

The individual card was not the interesting question. The spread was. I computed each player's four-round SG: Approach, measured the dispersion of that distribution within the field, and compared it with the dispersion of the tour-level distribution over the previous twelve months. The result: that week's approach dispersion was 2.3 strokes wider than the tour average.

The machinery of the course

Dispersion on its own is not a story. It has to be turned into a mechanism. Hanse's rework at TPC Boston built runoff areas around the greens — grass kept short so a high ball does not hold but slides off the edge. The consequence is that an approach error does not translate straight into score. It translates, first, into putt distance.

That is where the narrative inverted. When a missed approach leaves you thirty or forty feet away, the next stroke is a putt. It enters the scorecard as a putt. The graphic being sold as a putting contest was the downstream readout of an approach contest — the error above showing up as the number below.

The course amplified the cost of approach accuracy. A wide field dispersion is not evidence that everyone struck it badly; it is evidence that the course widened the gap between good and bad approaches. The winner's 17-under was what you get by deferring that penalty for four straight days.

What I will not print

I am not printing the winner's category splits. The 2026 export exists, but quoting figures from memory rather than reopening the file would be precisely the error this piece criticises. I can write that the largest separation in that field appeared in approach, and that 17-under was a score built by postponing the course's punishment four days running. Being better in one category across four days and being better once in four days are different things; the screen sells both under one word.

The denominator underneath the percentage

March 2026, DLF Golf & Country Club, Gurgaon, the Hero Indian Open — a European Tour co-sanctioned event, Matt Wallace's first European Tour title, and Shubhankar Sharma arriving off wins at the Joburg Open and the Maybank Championship. I built the strokes-gained package. The field scoring average logged at 73.8. On the closing four holes, 6.2% of approaches from 150–175 yards found water.

The desk cut that 6.2% to a single line. What did not get printed was the list of more than a thousand shots underneath the percentage, and the fact that 150–175 yards was the most punishing band on that course because both wind and elevation were live. My conclusion — that DLF's penal rough turned a 132-man field into a variance machine and the winner's 11-under was not a form signal — was cut to one line. At the time I thought it was a ruthless conclusion. Later I understood it was the necessary one.

Correlation is not causation, and this is the weakest part of the piece

Here I have to attack my own finding. A 2.3-stroke wider dispersion is a field-level property, not a player's skill. It describes how the course distributed the differences that week. It does not describe how anyone won. A win is one man's four days; dispersion is ninety-three other men's four days. Fusing them into one sentence is the habit I exist to break.

The second problem is putting. Across the four categories, putting has the least week-to-week stability — Broadie's work has shown this since 2026, and my own re-runs agree. "Best putting week of the season" is the claim most likely to be noise, and it was the claim running in the largest font. My job was not to disprove it. My job was to say what it rests on, and it rests on four rounds.

The third problem is sample. One event, 96 players, one course setup, one week of weather. No conclusion. I would rather write "one event, n=1, no conclusion" into a piece than let an editor imply a trend the data cannot carry.

May 2026 made the same point differently. With the Bundesliga restarting behind closed doors I pulled 1,046 matches across five seasons and tried to isolate crowd effects from fixture congestion and travel. Home win rate fell from 43.4% to 33.8%; home penalties per match dropped by a third. I have not treated home advantage as a constant since. The same lesson applies to golf: when a player appears to be falling, the first question is not whether his form has gone but whether the conditions around him changed. Weather, course setup and field strength can all manufacture a slump.

A variance warning is useful only if someone reads it twice — once when it is filed, and once when it is proved or disproved. A warning read once is decoration. And on ranking maths: the Official World Golf Ranking runs a 104-week rolling window with a minimum 40-event divisor. Before reading any event's numbers toward a ranking, remember the ranking is itself a seasonally adjusted figure, not a fresh one.

What to watch next

Set pieces, penalties, goalkeeper variance, and the way a man's swing changes under pressure — the model does not see any of it, and that caveat became the most-quoted part of my work. Producers now ask for the caveat rather than the headline. I filed an internal note before Kazan in 2026, before the result. That is the only way a forecast survives: write it down with a date on it.

So this time I am writing first. For the coming major weeks, my pre-registered forecast: the winner of that week will post less than half his SG: Putting in the following four-round event — my estimate, 70%. If it misses, I will write up why, and show my working.

The scoreboard records outcomes; my notebook records the conditions that produced them. Believing the screen was correct because twelve thousand shots were sitting in a file is laziness, not scepticism. And the next time a graphic tells you where the winner won it, ask the one question the graphic will never contain: who recorded that number.