The Empty ShotLink Cell and the Honesty Limit of Golf Analytics

5:48 a.m., third-floor data room of a training centre in Nagoya. On the second monitor, the ShotLink table I had left open overnight held 1,412 rows, and all 1,412 cells in the strokes gained column carried the same character: N/A. Nothing was broken. The vertical tracking camera at the 7th hole had been misaligned for three rounds, the ball-landing sensors had recorded nothing, and the software, exactly as designed, refused to interpolate what was missing. It left the column empty. I sat looking at a table that looked very much like a real table, and wondered how many analyses in the world have been written from tables like this one. At 6:10 the head coach called. He asked one question: “What did you conclude?” I said: “Nothing.” The line went quiet for four seconds. Those four seconds lasted longer than any analysis meeting I have ever sat through. Golf analytics today produces more data than every previous era combined. ShotLink covers nearly the full PGA Tour schedule; the DP World Tour runs its own radar system; regional Asian tours follow a few seasons behind. Every shot at the top level generates dozens of measurements: club speed, launch angle, spin rate, apex height, roll-out distance, time to rest on the green. The betting and sports-data layer has separated from broadcast rights into a revenue stream of its own, with its own contracts and its own vendors. More data does not mean enough data. Three kinds of shortfall rarely announce themselves. Missing baseline data: an event never wired, or wired but not operated correctly. Missing context data: numbers exist, but wind, green firmness and pin positions do not. Missing temporal data: the numbers exist, but the sample is too small to say anything. The difficulty is that none of these three produces a system error. They produce an empty cell. And an empty cell, inside a content industry that runs on publishing schedules, is treated as a professional failure rather than a technical one. The original question that morning was narrow: is this player hitting approach shots better than three months ago? I wrote it on paper before opening the data, because I had learned something expensive in 2026. That year, building an xG model for a second-division Japanese club, I constructed it from raw video and forgot the home-advantage coefficient. Over the final ten rounds I got six matches wrong. The data was not wrong; the data was right. I had asked the wrong question and then blamed the data. Data is never wrong — I am simply asking the wrong question. Since then, my analytical question is written down before I touch a table, and it stays unchanged until a conclusion exists. Three years later, in 2026, I did data work for a football site in Nagoya. For Japan against Belgium in the World Cup round of 16, the PPDA figures showed Japan pressing effectively, and I wrote exactly that in a bulletin sent at midnight. I ignored Belgium's midfield running after the 70th minute. Belgium won 3-2. I criticised myself publicly in three sentences: I lacked a real-time fatigue variable; I had concluded beyond the data; I would not repeat it. No further apology was needed — only a running-intensity chart in fifteen-minute blocks in every report afterwards. When I moved into golf full time, I carried both scars. But golf makes the problem harder. A football match is one independent sample with ninety dense minutes. A golf round is eighteen discrete samples ten minutes apart, subject to rotating wind, greens that firm up by the hour, and pin positions reset daily. Same shot, same yardage, same club, two entirely different outcomes for reasons that live outside the table. Back to that Nagoya morning. I split empty cells into three groups. Group one is genuinely empty: the system could not measure it and nothing can recover it. Group two is recoverable: data absent from the main table but present elsewhere — video, paper scorecards, refereeing records. Group three is structural: data that was never designed to be collected, such as workload for fifteen-year-olds, or actual putting stats for regional women's events. Collapsing all three into a single N/A is the most common error I see in reports sent to coaching staff. Group one I discard. Entirely. No interpolation, no field average, no borrowing from the previous round. Group two I rebuild by hand — the dullest and most valuable work in the job. What did not happen often tells the truth more plainly than what did. Cross-checking the 7th-hole video against the organiser's paper scorecards, I recovered forty-three shots the system had missed. Forty-three out of 1,412 rows. That is not a rounding error. That is a chapter torn out of the book. Group three is the worrying one, because it never reveals itself. Nobody raises an alarm about data that was never designed to be collected. Gaps in a table can speak, if we are willing to listen — but only when we go looking for them. I now spend twenty minutes a week simply listing the metrics that should exist and do not. That list is longer than the list of metrics I actually use. One example I still use in internal training. When Hideki Matsuyama won the 2026 Masters, searches for his strokes gained putting data spiked across the system. The same week, a regional women's event in Asia finished with almost nothing published beyond a final leaderboard. Names like Hideki Matsuyama, Scottie Scheffler and Rory McIlroy are attached to thousands of rows per season, while most of the sport cannot produce a single hole-by-hole distance chart. One sport, one week, two levels of data presence separated by orders of magnitude. Fans see the lit part and assume the rest does not exist. That asymmetry breeds a closed ecosystem wherever data is thin. I once tracked a system that published only its own internal metrics with no third-party verification. After three seasons its average figures sat above every peer system, while the players who left it underperformed on the international stage relative to expectations built from those very figures. Nobody cheated. There was simply no control. A system that only compares itself to itself has one direction for its numbers to move. Between Vietnam, where I was born, and Japan, where I work, the golf data gap is so wide that direct comparison is meaningless. A national amateur event in Vietnam may offer a leaderboard and photographs; a prefectural amateur event in Japan already has hole-by-hole distance data. I keep the comparison only where the gap is large enough to carry tactical meaning — completion time for the final group over eighteen holes in heat, for instance — never at the advanced-metric level that does not exist on one side. Cultural comparison earns its place only when both sides can measure it. The 2026 season taught the same lesson at scale. With stadiums empty and schedules broken, my form-prediction model lost its inputs. I proposed using youth-team GPS training data and the precedent of historically interrupted seasons. The coaching staff pushed back. I argued it through with data from the season disrupted by the 2026 earthquake, when teams returned at a visibly different rhythm. The team lost only two of ten restart matches. The lesson was not the result but the process: when the primary source dries up, publish the substitute source and publish its limits alongside it. Moving from football to golf, I carried one translation across, used only when the numbers permit it. In football, gegenpressing is the art of recovering the ball immediately after losing it. In golf its nearest equivalent is bogey recovery speed: after losing a stroke on one hole, how many holes does a player need to regain rhythm? Gegenpressing does not break the data; it breaks my assumptions. I assumed fast recovery was a psychological trait. The data showed it tracks closely with tee order and approach quality on the next hole — technique, not mentality. Once again my assumption collapsed and the numbers held. There is one further layer I consider the most serious gap, and it concerns junior players. Workload, tournament rounds and movement intensity for players aged fourteen to seventeen go largely unrecorded across most systems, precisely during the years when bodies are unfinished but already pushed into adult competitive rhythm. When a sixteen-year-old plays three rounds in four days, nobody holds data on what that does to their spine ten years later. The silence of the table here is not neutral. It is a debt not yet entered in the ledger. At market level, a week without operational ShotLink is not academic. Data vendors sell weekly packages; bookmakers price with models running on those packages; and when the underlying data is thin, their margins widen rather than close. The cost of the empty cell does not fall on the vendor. It falls on the buyer, and ultimately on the viewer, who receives only a smoothed number. The split between tour systems in recent years adds another layer. When a group of players competes outside world-ranking recognition for a period, any direct comparison with the rest becomes indirect, mediated by conversion models. Every conversion model carries assumptions, and assumptions are rarely published alongside the number. Every number is a confession not yet written down. I do not believe in luck; I believe in cultivated probability. And probability is only cultivated with verified data. Stopping there, however, would mean deceiving myself once more. The greatest temptation for anyone working with data is not to invent numbers. It is to turn the gap into a doctrine. After years of saying “not enough data”, it is easy to slide into the opposite state: treating every conclusion as suspect, every model as a con, and caution as intellectual laziness. I made that mistake twice in consecutive seasons, and noticed late both times. A gap earns its value only when it answers two questions: why the data is missing, and which direction a conclusion would move if the data existed. Without those two answers, a gap is just an excuse dressed in terminology. Correlation and causation live squarely in that zone. In golf it is easy to find a beautiful correlation — high strokes gained approach wins more — and then treat it as cause. But the causal order can invert: frequent winners are placed in the final groups, play in more stable green conditions, and therefore post prettier numbers. Elimination is the real key. When I removed tee order from the model, the gap between the leaders and the middle of the field narrowed sharply on several metrics. What remains is the part worth discussing, and it is much smaller than the metric leaderboards still present. Three weeks after that Nagoya morning, I sent the coaching staff a nine-page report. The first four pages existed only to say that the forty-three shots at the 7th hole supported no conclusion, and to explain why. The next four presented the substitute data I had rebuilt by hand. The final page posed three new questions I would not answer alone. When the data hides its face, the error becomes the guide. One question stays open: if an honest empty table is worth more than a full and wrong one, why does sports analytics still pay the person who fills empty cells fastest, rather than the person who knows when to leave them empty?