Three Rounds Are Not Enough to Tell a Story: The Small-Sample Problem in Golf Data

Three Rounds Are Not Enough to Tell a Story: The Small-Sample Problem in Golf Data In February, I sat in front of a Strokes Gained table for a young golfer who had posted three consecutive rounds above 2.0 in SG: Putting. Placed side by side, those three numbers formed a clean upward line, the kind of chart any content producer wants at the top of a piece. Last season, I built a small model to forecast early-season putting form based on exactly this kind of three-round data. By June, seven of the ten golfers I had rated highly had returned to around their career averages, and the remaining three had fallen below what I once treated as the floor. The numbers never changed. The reader of the numbers is the one who changed the question. The first question I should have asked was not how well this golfer was putting, but what exactly three rounds were enough to say. When I rewrote that question properly, an entire season of analysis behind it finally began to make sense. The regular season is the phase in which golf data is in its most dangerous state. Courses change constantly, winds shift by the week, turf transitions from Bermuda to Poa within a few flights, and the sample each golfer accumulates usually sits around three to six rounds. That is the zone where human intuition operates most strongly and errs most often. One golfer opens with two rounds under par, and the media instantly calls it form. Another leads the Strokes Gained: Putting board after four rounds, and people instantly write about a putter catching fire. But my job is to stand inside that current and ask: how much of that number is signal, how much is noise. I work as a sports data analyst in Nagoya, covering golf for the Japanese market, after leaving the course as a transitioning athlete. I did not start with an academic grounding in probability. What I had was a decade sitting beside spreadsheets, teaching myself to distrust them. In my early years, I wrote about golf by adding and subtracting basic metrics: fairways hit, greens in regulation, putts per round. I presented those tables as though they were truth. It was only when my model collapsed in a football context that I understood the same table can lead to two opposite conclusions, depending on the question behind it. Gaps in a table speak, if we are willing to listen. The problem is that most readers of sports data are not trained to hear the gaps. They are trained to hear the number. And the number, in raw form, always appears more confident than it deserves to be. Strokes Gained works on a fairly simple principle. For each shot, one compares the number of strokes a golfer actually takes to finish the hole from that position against the average number a tour-level golfer needs from the same position. The differences accumulate into SG. A shot better than average yields positive SG, a worse shot negative SG. The beauty of the system is that it separates skill from raw outcome: a golfer can score poorly and still post positive SG: Approach because he placed the ball correctly and only lost strokes on the greens. The flaw is that it does not protect the user from himself. Over an 18-hole round, a golfer takes roughly 30 putts. Three rounds is about 90 putts. That sounds like a lot, but putting is the highest-variance category in the entire game. For the same golfer, same technique, same turf, the make rate from around four meters can swing widely between two consecutive weeks because of a few millimeters of contact. This is why, in the research of Mark Broadie, the founder of Strokes Gained, the year-to-year correlation of a golfer's putting performance typically lands between 0.2 and 0.4. Compare that with SG: Approach, where the season-over-season correlation usually sits around 0.6 to 0.7, or with driving distance, where the figure jumps to nearly 0.9, and a very different picture emerges. In other words: your putting last season barely predicts your putting this season, while your driving distance is fairly stable, and your ability to hit greens sits between the two. This is the fact I must remember before every piece I write about form. The past season gave me a clear enough example. One golfer opened with a three-round stretch of very high cumulative SG: Putting, high enough that the metric itself became a headline. I tracked that run like a fever, and noted something the coverage skipped: within the same three rounds, his SG: Approach was negative in two of them. He was scoring off the putter, not off the approach. When putting returns to average, and it almost always does, the foundation underneath is exposed. Without a solid approach base, a putting hot streak is merely a loan against the future. I do not believe in luck; I believe in cultivated probability. A long putt dropping is the outcome of a probability distribution, not evidence of a new ability. When a golfer repeats it across many rounds and many seasons, that is when I start betting on it. Here I have to tell an old story, because it is the origin of the method I now use for golf. In 2026, at 24, I began doing data analysis for a Japanese football club playing in a lower division after relegation. I built an expected-goals model from video, and I missed the home-ground coefficient. The consequence: across the final ten rounds, my model was wrong in six. I sat back down with the full footage, checked every phase, and understood that raw data is insufficient; it needs tactical context before it can say anything true. That taught me a principle I carried intact into golf: never give a number without a context condition. Data is never wrong, only the question I set was. The same metric, placed in the wrong context, becomes a perfectly legal lie. In 2026, a second rupture arrived. I was a data contributor for a large football outlet in Nagoya. In a knockout match, I collected a pressure metric to prove one team pressed effectively. The metric looked beautiful. I ignored another variable: the opponent's running distance in the closing period. As a result, the team I analyzed was overturned in extra time because of a vast gap in midfield once fitness declined. I publicly criticized myself on my personal page and admitted the model lacked a real-time fitness variable. Gegenpressing does not break the data, it breaks my assumption. Pressure looks the same on every table, but pressure in the tenth minute and pressure in the seventieth are two entirely different things. Since then, every pressure analysis of mine must include an intensity chart in fifteen-minute intervals. I never conclude anything about pressing without fitness data. This may sound far from golf, but the principle transfers almost one-to-one. In golf, a shot on the third hole and a shot on the seventeenth within the same round do not share the same basis for comparison. A round-long cumulative SG: Putting, if not split by hole, can conceal a truth: the golfer putted poorly in the first half because the body was not warm, and superbly in the second half because it found rhythm. Averaged together, it produces a beautiful number, but that beautiful number mislabels the entire round. The 2026 season brought me to another problem, and it is why I am writing this today. When the pandemic emptied stadiums and broke the schedule, a club went two months without playing. I, then a mid-level staffer, had to rebuild a performance-forecast model with no match data. I proposed using GPS training data from the youth team and precedents from historically interrupted seasons. The coaching staff objected. I persisted in proving it with data from a season that had been interrupted before. The result: the club survived relegation, losing only two of ten restart rounds. The lesson applied to golf becomes this: when match data is empty, training data and historical precedent become the guide. When data hides, error becomes the guide. That is why I always keep a separate note on how each golfer's swing changes during a non-competing period. Here I must publicly admit a limitation. I do not have ShotLink for every event I follow. For events in Japan, shot-by-shot data is not always complete, and I have to reconstruct part of it with my own eyes from video. That means part of my data carries error originating from the human, not only from the sample. I state that error in every table I publish. No stated error, no trusted conclusion. Every number is a confession not yet written into prose. The problem is that the confession usually says more about the writer than about the golfer. Now let us return to the three-round story at the top. I want to reconstruct my chain of reasoning transparently, including the part I got wrong, because that is the only way a reader can trust the part I got right. My initial assumption: a golfer with cumulative SG: Putting above 2.0 across three rounds possesses above-average putting ability. I took that assumption and tried to falsify it. I pulled every golfer in my database, found the three-round stretches with SG: Putting above 2.0, and measured what their SG: Putting was over the next twenty rounds. The average fell below 0.5. That means roughly seventy-five percent of the outperformance in the first three rounds evaporated over the next twenty. This is the point I am often misunderstood on when discussing regression to the mean. Regression does not deny that the golfer putted well. He truly did putt well in those three rounds; the strokes do not lie. Regression only says that a large part of that performance came from factors that will not repeat: a few long putts dropping at the right moment, a few balls catching the edge and falling in instead of spinning out. True skill is drawn out of the wave; the rest is tide going out. So where does a golfer's true skill lie? If I had to rank Strokes Gained categories by predictability, my order is: driving distance, approach skill, around-the-green skill, and finally putting. This is the order I use to decide which part of a table to trust. A golfer with stable multi-season positive SG: Approach is a golfer worth trusting. A golfer with positive SG: Putting across three rounds is a story to tell, not a golfer to bet on. Here I want to talk about a data gap I consider the most important and least discussed: data about what did not happen. What did NOT happen often tells the truth better than what happened. When I watch a round, I do not just count putts that dropped. I count putts not taken because the ball already sat too far from the hole after the approach. I count the times a golfer chose not to attack the pin because of wind, and hit to the middle of the green instead. Those decisions do not appear on the scorecard, but they are exactly what distinguishes a golfer playing the right strategy from one playing beautifully and pointlessly. A concrete example. In a windy round in Japan I watched live, a golfer known for driving distance chose an iron on four par-4s instead of a wood, purely to keep the ball under the wind. He did not make birdie on those four holes. Reading the scorecard, one sees an ordinary round. Reading the decisions, one sees a round managed correctly. His SG does not reflect that, because the system measures the shot, not the reason behind the shot. That is the limit of every current golf model, and the reason I never let a table conclude by itself. I have to sit back, read the context, and ask: did this golfer hit that shot because he could not hit better, or because he chose to hit it that way? There is another lens I have carried through my whole career: the difference between Vietnamese and Japanese training culture, and how it shows up in the data. I was born in Vietnam and work in Japan. In Japan, a golfer practices putting by tracked volume, every session, every distance band, and that number is the yardstick of progress. In Vietnam, where I grew up, golfers practice by feel, by intensity, by the sensation of the club meeting the ball correctly. Both approaches produce great golfers, but they produce two different kinds of data. The consequence is that when I read a table from a golfer trained in Japan, I can trust its stability more, because his training process was already standardized. When I read a table from a golfer trained by feel, I need additional evidence before trusting it. This is not a remark about people, but about the degree of control over the variable. Keep the comparison only when the data gap is large enough to matter; otherwise, I leave it out of the piece. And now the most counterintuitive part, the part I must state plainly even though it runs against how most golf content operates. The entire golf content industry, myself included, operates on the assumption that more data means better conclusions. That assumption is false under small-sample conditions. More data early in the season does not make my conclusion firmer; it only makes my conclusion more confident while remaining just as fragile. This is the biggest blind spot in today's sports analytics: we measure certainty by the number of data rows instead of the quality of the context behind them. A model with a million data rows but no fitness variable for the closing period will still be wrong, and it will be wrong with full confidence. I made exactly that mistake in football, and I see it repeated every week in golf in the form of three-round analyses presented as though they were truths about a whole season. My counterintuitive angle is this: in the golf regular season, the most valuable thing is not a correct conclusion but a correct question. A correct question survives across rounds. A wrong conclusion from three rounds is forgotten in a week and leaves a small stain on the reader's trust. Reader trust erodes bit by bit like that, until they no longer believe any table at all. I am not writing this to say golf data is useless. The opposite. Golf data is better than the data of most other sports. The problem lies with the user. We are holding a high-precision measuring instrument and using it to measure the wrong thing. So what is the signal for the next round that I will track? I will not track who leads the Strokes Gained: Putting board after three rounds. I will track who has positive SG: Approach across three or more consecutive rounds, because that is the predictable category, and I will cross-check it against SG: Putting to see how much of their performance comes from stability and how much from tide. I will track the shots that do not appear on the scorecard: the times a golfer chooses safety over beauty. I will track the gaps in the data, because as I said, they speak too. The regular season is not a sprint, but a race of patience. Whoever reads data with patience will see the story before it becomes a headline. Whoever reads data with eagerness will see the headline before there is a story. I am still learning to be patient.