Where ShotLink Doesn't Reach: The Data Problem of the Regular Golf Season

At 2:40 in the morning in Nagoya, I replayed round-three footage from a DP World Tour event for the fourth time. The spreadsheet beside me was still empty: not a single row of shot-level data. No ball-landing coordinates, no distance to the pin, no green slope. Only images, and a camera placed slightly off-axis at the 14th hole that made every green-depth measurement meaningless. It took me 7 hours and 36 minutes to label 1,240 shots by hand, an average of 22 seconds per shot, sorting them into eight technical categories. The result: the model's SG: Approach figure deviated by 0.31 strokes per hole from the internal baseline. Three times before, I blamed the cameraman. This time I could not. Golf's data infrastructure is divided into three tiers, and each tier carries a different level of confidence. At the top tier, ShotLink — the shot-tracking system the PGA Tour has operated since 2026 — covers almost every round on that circuit. Each stroke is stored with distance, direction, finishing position and lie. That is the foundation that made Strokes Gained the standard language of golf analysis from around 2026 onward. In the middle tier, the DP World Tour runs its own collection system but does not publish full shot-level data for every event. The Japan Golf Tour, which I track on site, offers basic scoring, fairway, green and putt counts — not enough to build SG: Approach. At the bottom tier, the Korn Ferry Tour and regional tours are little more than leaderboards. LIV Golf creates a gap of a different nature. The tour launched in 2026 and does not publish shot data to ShotLink standards. In October 2026, OWGR rejected LIV's application for ranking points, partly on access and data-quality criteria. For an analyst, the consequence is concrete: dozens of players can complete an entire season without leaving a data sample deep enough for technical evaluation. Then comes the variable that breaks historical comparability. The USGA and the R&A announced their golf-ball distance limitation in December 2026, applying to elite competitions from January 2028 and to recreational play from 2030. That means the driver-distance data series will break at a defined point, and every model built on pre-2028 data will need its baseline recalibrated. Three tiers, three confidence levels, and a regular season with no truly empty week. That is the problem I solve every month. The question I write at the top of every spreadsheet: can a Strokes Gained metric reliable enough for decision-making be built from 2D video on tours without ShotLink? After 14 months, the answer is yes, but only for four of eight technical categories — and with error margins that must be declared openly on every row. The first surviving category is putting inside two metres. It has the lowest variance in the whole model: distance is fixed by marks on the green, slope matters at the scale of a few centimetres, and a success rate over 40 putts is stable enough to compare across rounds. The second is greens in regulation, because a binary outcome does not depend on image quality, only on whether I can see the ball roll to the green's edge. My labelling error here is 1.2 holes per round, measured by cross-checking against the official scorecard. The third is scrambling — getting up and down after missing a green. The fourth is putts per round. SG: Off the Tee, by contrast, is the first category discarded. I measured driver-distance error on 2D video against radar data from three Japan Golf Tour events with on-site equipment: an average error of plus or minus 11 metres. That is enough to flip the ranking of two players separated by 7 metres. So the only driving metric I keep is fairway-hit rate — a binary variable indifferent to absolute values. The principle I drew from this runs against most data practitioners' instincts: the highest-variance metric is also the one most likely to be wrong when the input is thin, so it must be eliminated first, not prioritised for analysis. Since 2026, after I overlooked the fitness variable in Japan's round-of-16 match against Belgium at the World Cup, every model of mine has had to include an intensity chart by time segment. I carried that principle into golf using three six-hole blocks, but kept the cross-disciplinary lens only after testing whether the parallel was real. Across a sample of 62 Japan Golf Tour rounds I labelled, the average score for holes 13–18 was 0.68 strokes higher than for holes 1–6; among players over 40, the gap widened to 1.14 strokes. Had the deviation been a few tenths with no age-based differentiation, I would have dropped the comparison. One type of data exists on no tour at all: referees' rulings. No tour publishes decision logs as structured data, because rulings are tied to player testimony and on-site judgement and resist standardisation into fields. What replaces it is my manual log from broadcast footage plus post-round interviews, and I always assign it a low confidence rating. A six-hole block contains only about 20 shots, far too few to conclude anything. So I never read a single round's score: I read a rolling 10-round window, roughly 180 holes. In that window, the standard deviation of the closing-block scoring average falls from 1.9 strokes to 0.7, enough to separate signal from noise. Course-fit analysis is the category hit hardest when data is thin. Comparing a player to a venue requires fairway width, grass height, green speed and prevailing wind. On tours without ShotLink, I collect those manually from course maps and weather reports. My workaround is to demote course fit from a forecasting metric to an elimination filter: I do not say who will win, I only remove players whose technical profile cannot adapt. Baselines are where many models die. Strokes Gained only means something against a defined baseline, usually the field average for the same round. With manual data, I must choose a narrower baseline: the average of players with complete data in the same round. That choice biases results in favour of the well-measured group, which is why I always print error margins instead of a bare number. One example shows how thick data illuminates what thin data hides: Hideki Matsuyama, the 2026 Masters champion and the first Japanese player to win a major, has a shot-level record spanning many PGA Tour seasons. Among younger players born in the late 1990s, such as Scottie Scheffler, Masters champion in 2026 and 2026, or Ludvig Åberg, who turned professional in 2026, the age curve is on the rising slope. Rory McIlroy has continuous data going back to 2026. How many players on other tours have comparable record length? Very few, and that is an infrastructure problem, not a talent problem. I once assumed the data gap between tours was cultural. After comparing the operating budgets of measurement systems, I found it is an infrastructure story: an event with radar units and calibrated cameras costs many times more than one with an electronic leaderboard. The distance between the Japan Golf Tour and the PGA Tour on this axis lies not in coaching methods but in the ability to pay for equipment. That is why I have largely dropped cultural comparisons from my data writing. I asked the wrong question for five straight months: I went looking for "who is playing best" when the data could only answer "who is putting best". The correcting evidence: when SG: Putting is stripped out of the model, its correlation with scoring drops from 0.61 to 0.34 across the 62-round sample. I then changed the question to "which metrics still hold when part of the data disappears", and every spreadsheet of mine now starts there. Data is never wrong; I simply asked the wrong question. The greatest temptation in this job is turning a correlation into a cause. When I see a player with high SG: Approach and a good score, the reflex is to say approach decides outcomes. But with thin data I am not measuring that player; I am measuring the infrastructure of the tour he plays on. A tour with good data hands me a better model, produces smoother rankings, and creates the impression that its players are more consistent. The second-order effect is the worrying part. Players on data-rich tours get analysed more, quoted more, and attract more sponsorship. Players on data-poor tours are rated lower not because they play worse, but because nobody can measure them. Gaps in a data table speak too, if we bother to listen: they speak about tournament structure, not about a player's technique. What did NOT happen often tells more truth than what did. A player absent from every metric ranking is not necessarily the season's worst; he may simply play where nobody is recording. The signal I am watching next is not a putt or a drive, but whether the DP World Tour expands shot-level publication, and which baseline the 2028 ball rule will force me to recalibrate. When data hides its face, error margins become the guide — and a good guide states clearly where it is leading.