A Golf Leaderboard With Every Column Filled and Nothing Inside

At 3 a.m., the export file from the scoring system of an amateur tournament held in Nha Trang landed on my machine. 156 rows, 14 columns: par, gross score, fairways hit, greens in regulation, putts, driving distance, round completion time, four strokes gained categories. Not a single empty cell. My automated validator ran the whole file and reported no structural error, no type error.

I opened the strokes gained column. All 156 players held a value of 0.000 across all four categories: off the tee, approach, around the green, putting.

The leaderboard was valid. The format was correct. And it carried not one unit of information about any player's shots. The cause was simple: the tournament had no shot-level tracking system, so strokes gained could not be computed. But the export template still generated those four columns, filled them with default zeros, and included no field stating that the column was empty because no source existed.

Whoever receives that file — a coach, a sponsor, a trader — sees a complete table. They do not see the gap. In my trade, this is silent failure: the system does not crash, only the conclusions go wrong.

Golf has the thinnest data sample of any mainstream sport. One 18-hole round is roughly 70 to 80 shots; a four-round tournament yields about 290. Set beside football, where a single match produces thousands of ball touches, golf forces the analyst to work with individual sample units. There is no room for junk data, and no room for fake data either.

The strokes gained framework, published by Columbia Business School lecturer Mark Broadie and adopted by the PGA Tour from the 2026 season, was built to solve exactly that. It does not count; it prices. Each shot is converted into expected strokes gained or lost against the field average from the same distance and the same lie. To compute it, the system must know the start point and end point of every shot, in metres, along with grass type and lie type. On the PGA Tour, ShotLink does this on every hole, every round, every competition day.

Outside the PGA Tour, coverage collapses fast. The LPGA has narrower coverage. Asian tours, amateur tours, and essentially every domestic tournament system have none. But scoring software still exports every column, because the template is shared across all events.

Three years of tracking domestic rounds taught me something no university teaches: a record can be perfectly structured and completely empty, and the software has no mechanism to report that. I keep one rule before every analysis: check coverage first, read numbers second.

Coverage is shots recorded divided by shots actually played. Across the 412 rounds I logged by hand and cross-checked against video, average coverage was 62%. More than a third of all shots vanish. And they vanish systematically: balls into trees, into water, into sand, balls replayed from a penalty spot — the worst shots are the hardest to record. Golf data does not go missing evenly. It goes missing in the direction that flatters the picture.

| Displayed column | Value in file | Actual information | Coverage | |---|---|---|---| | Strokes gained, 4 categories | 0.000 | cannot be computed | 0% | | Fairways hit | 9/14 | denominator distorted | 11/14 | | Putts per round | 31 | count complete, distance distribution missing | 100% count, 0% context | | Driving distance | 246 m | only 8 holes had measurement | 57% |

The four rows describe four different failure types, and only the first is an obvious one.

Fairways hit is a denominator failure. When a ball goes into water or out of bounds, that shot is not counted in the denominator of “fairways possible”. A player who hits 6 of 8 chances looks worse than one who hits 8 of 14, even though the latter lost two balls. A column that appears neutral rewards caution and punishes aggression, the opposite of the meaning readers assign to it. In my own data, the group averaging four strokes better in handicap posted a lower fairway rate, purely because they hit driver on narrow holes.

Putts per round is a context failure, and the single most misread column in golf. Putting count depends almost entirely on where the ball lands on the green. A player with a sharp approach leaves it near the pin and takes few putts; a player with a poor approach putts repeatedly from eight metres. Counting putts without a distance distribution measures the output of a different skill. Broadie splits scoring differences among elite players into four buckets: approach about 40%, off the tee about 28%, around the green about 17%, putting about 15%. Domestic amateur leaderboards allocate three or four columns to putting and one to greens in regulation. That ratio is inverted against the real contribution ratio.

Driving distance is a sampling failure. Only holes with measuring equipment produce a number, usually straight, wide, easy holes. The figure 246 m describes the eight best drives a person hit, not the person.

Then comes strokes gained at 0.000. This is the most dangerous failure because it resembles a neutral value. A quick reader concludes “this player lost nothing on the greens”. The truth is nobody measured anything.

In the 2026 season, Scottie Scheffler led the PGA Tour in total strokes gained by a wide margin over the field, while his putting sat near tour average. Read the putting column alone and you see a mid-tier player. Read strokes gained split by category and you see nearly the entire gap coming from approach and off the tee. One player, two opposite conclusions, and the difference is whether the data was split by dimension. Ludvig Åberg is the reverse case: he rose with top-tier ball speed and very high strokes gained off the tee, but his recorded round count is still thin, so any comparison with a player holding 200 rounds must carry a confidence coefficient. Rory McIlroy has held a stable strokes gained off the tee across multiple seasons, and that is usable data precisely because the sample is thick enough.

Three consequences of silent failure, in the order I encounter them. First, a strokes gained ranking for a tournament without tracking will align almost perfectly with the gross leaderboard, since both sort by stroke count, yet readers believe they are viewing an independent metric. Second, forecasting models ingest the 0.000 column and learn the wrong lesson: they conclude strokes gained has no predictive value, when the problem sits in the source. Third, the report sent to the sponsor has clean charts, a colour palette, and was built out of nothing.

How I handle it in reports: every metric carries a coverage flag, and any metric below 70% coverage is dropped from the model rather than interpolated. Interpolation in golf is more dangerous than a blank, because it turns a shot nobody saw into a data point that looks measured.

Data is never in a hurry; it simply waits for someone who can read it. An empty column waits for no one. It just sits there, correctly formatted, ready to be misread.

The counterintuitive angle: more columns make analysis worse. I ran a small test with two groups receiving the same 40-round dataset. Group A received all 14 columns. Group B received 6 columns plus one line noting coverage per column. Group B caught the fairway denominator error markedly faster; Group A reached conclusions with more confidence and got two of four questions wrong. Schema completeness manufactures trust, and that trust is not underwritten by the data.

Correlation does not rescue it either. High strokes gained correlates with winning, but it does not cause winning; and the correlation collapses once coverage falls below 70%. In events where my coverage data is thin, strokes gained rankings after reweighting shuffle up to four of the top ten. Half of a leaderboard is a product of method, not of shot-making.

To those who say “I have 20 years in this game, putts per round is enough”, I answer with figures: putting accounts for roughly 15% of scoring difference, and raw putt count is contaminated by approach quality. I do not argue with tone. I put the comparison table on the desk, point out the sample limits, and let the other person decide. An empty stadium does not lack noise; it lacks one dimension of data — and a tournament without tracking is the same: it does not lack shots, it lacks the dimension to read them.

A bounded prediction: within the next 12 months, at least one tour-level golf data provider will add a coverage field to its default leaderboard, because report buyers have started asking. Until then, every report I send opens with a single line: what percentage of shots in this dataset were recorded. I write the report, close the file, and the market reopens on its own.

And if the leaderboard in front of you has a strokes gained column, do you know who measured it, with what equipment, and across how many holes?