Golf's Data Void: ShotLink, OWGR and the Limits of Analytical Models

At 7:42 on the morning of 13 August, in Nagoya. The familiar dashboard opens, waiting for data to land after an overnight automated run. It carries 288 columns — enough room for driving distance, fairway rate, GIR, scrambling, strokes per round and four branches of Strokes Gained. Row count: zero.

I checked the logs before blaming the interface. The query completed without error. It returned exactly one populated field: the domain label, golf. Article title: blank. Source: blank. The list of information points was empty, not a single item. Entities involved: nothing to list. Time sensitivity: not assessed. Source quality: undetermined.

Seventeen years of watching this industry taught me to endure nights when a model returns poor results. An entirely empty table is a different matter. It does not tell me I was wrong. It tells me I had nothing yet to be wrong about.

Golf analytics runs on one assumption

Every professional golf model stands on a single belief: each shot leaves a trace, and the trace is measurable. ShotLink — the PGA Tour's shot-level data collection system — turns that belief into infrastructure. From it, Strokes Gained splits a round into four buckets: off the tee, approach, around the green and putting.

Among those four buckets, SG: Approach is the metric most strongly correlated with scoring. A good approach shortens the putt, and putt distance decides the probability of holing out. That causal chain is short enough to be hard to dispute. Data Golf, an independent analytics platform, is normally used to cross-check the official numbers — when two sources disagree, the divergence is the part worth reading.

The entire machine depends on one condition: there has to be a shot to measure. No shot, no SG. No SG, no form ranking. No form ranking, and every judgement is just a feeling dressed up in terminology.

A two-tier pipeline and the worth of a single information point

The system I run all year has two tiers. Tier one reads a source and breaks it into information points — small, citable claims, each with entities, timeframes and provenance. Tier two builds eight analytical dimensions: technical and data, player and form, tournament system, industry governance, rules and equipment, risk surface, public narrative, and industry transmission.

The boundary between the tiers is a hard rule: tier two may only reason from what tier one returns. An information point is the atomic unit of evidence. Without it, every tier-two conclusion is fabrication with good formatting.

That morning, tier one returned one label and nothing else. I could have built all eight dimensions immediately — they would have looked highly professional, full of tables and arrows. And they would have been entirely wrong, because no column stood beneath them.

Gaps in a data table can speak too, if we are willing to listen. But this one could only say a single thing: there was nothing here yet to hear.

What a golf data table needs before it can speak

To analyse one player's technique, you need at minimum a player name and a venue name. Only then can you pull SG: Off the Tee, SG: Approach, SG: Around the Green, SG: Putting, average driving distance, fairway rate, GIR and scrambling.

To place that player in a wider picture, you need an OWGR position, the tournament being played, and the corresponding ranking system. The OWGR determines entry into majors and many elite events — skip it and you skip the player's entire decision-making motive.

To read a course, you need the course type: a coastal Links with firm fairways and shifting wind, a heavy-rough penalty course, a distance-rewarding layout, or a bespoke setup such as Augusta. The same set of metrics, placed on two course types, yields two opposite conclusions.

To position an event, you need the cut line. On the PGA Tour, the cut sits around the top 65 and ties after 36 holes. For players living near the Tour Card boundary, a missed cut is not a minor detail — it is the difference between membership and qualifying.

Those four data groups are the minimum. That morning, all four were empty.

Eight dimensions and the quiet death of each

The technical dimension could not identify its subject. There was no way to know whether the question concerned an individual's technique, a single round, a swing overhaul, or the effect of an equipment change. Four technical risk flags — a short-term putting hot streak extrapolated linearly, an unfinished swing-transition period, a technical profile mismatched to the course, strength in one segment masking decline in others — could not be tested. Not because they are rare. Because there was nothing to test.

The player and form dimension vanished entirely. No name, no OWGR position, no FedExCup or Race to Dubai standing, no tour tier. The industry's most important analytical axis — delivery at majors versus performance in regular events — was out of reach.

The tournament-system dimension fared the same. With no event name there is no tier: major, The Players, Signature Event, regular event, team event, or feeder tour. Without a tier, OWGR points, prize money and Tour Card consequences cannot be modelled.

The governance dimension carries the highest diagnostic value in its collapse. It is essentially an entity-relationship exercise: PGA Tour versus LIV Golf, the position of the DP World Tour, regional tours, the role of investment funds, the reactions of sponsors and broadcasters. No entities, no map. The total collapse of this dimension is the strongest single signal that the entity-extraction layer failed outright rather than merely underperforming.

The rules and equipment dimension does not hold either. There is no way to know whether the source concerns an on-course ruling, an Equipment Standards question, a slow-play enforcement case, or eligibility conditions. The flight-distance reform initiated by the USGA and The R&A — commonly called the ball rollback — has an implementation timetable split by level of play, and its effects differ sharply across player groups. But to analyse it, you first have to know whether the source mentions it. The source mentions nothing, because the source never arrived.

The risk dimension left exactly one scoreable entry: system failure. Competitive, psychological, injury, commercial and governance risk all went unrated. A risk matrix with one row is a risk matrix about the process itself, not about golf.

The last two dimensions — public narrative and industry transmission — had nothing to attach to. No narrative to place on the curiosity-acceleration-peak-backlash life cycle. No transmission channel to trace, from course economics and equipment brands to sponsorship, broadcasting, and finally data and information markets.

What can be inferred from an empty result

The domain label golf was the only surviving field. That means the classifier ran and the content fetcher did not. A domain label can usually be inferred from a URL, a page title, or a meta description — things that exist even when the article body fails to load. When the title, author stance, article purpose, entities and time sensitivity all come back blank, the most parsimonious explanation is a single point of failure in the content-retrieval layer, not five independent bugs.

When data hides its face, error becomes the guide. Here the error did not come from the model. It came from transport. And transport error is the most dangerous kind in this trade, because it is silent. A wrong model indicts itself with skewed results. A broken pipeline indicts no one. It simply returns zero and waits to see whether anyone is patient enough not to replace that zero with an invented figure.

One gap, two data cultures

In Vietnam, most amateur and semi-professional golf events have no shot-level data collection at all. No ShotLink, no SG, often not even a standardised scorecard. Analysts there are used to starting from very little: direct observation, handwritten notes, a few rough metrics, and a correctly calibrated dose of humility. What does NOT happen often tells the truth more clearly than what did — for them that is a survival principle, not a slogan.

In Japan, the data layer is dense enough that practitioners can forget what scarcity feels like. The J2 League, domestic tours, GPS systems in training — everything is recorded. Precisely because of that, when a pipeline returns empty, the first reflex is to suspect the interface, then the model, then yourself. Someone working in Vietnam would recognise it at once: with no data, do not analyse. Someone working in Japan takes an extra hour to accept the same thing.

That cultural gap is only worth writing about when the divergence is large enough to mean something. This time it was: one side treats emptiness as the default, the other treats it as an incident.

Three sentences of self-criticism, with corrective data

In 2026 I built a manual xG model from video for Nagoya Grampus and omitted home-advantage weighting, which made my forecasts wrong in 6 of the last 10 rounds. In 2026 I read the PPDA figures from the Japan–Belgium match and ignored the opponent's running distance after the 70th minute, and my pressing model collapsed in extra time. Both times my question was not technically wrong — it was wrong because a variable had never been collected.

Data is never wrong; I simply asked the wrong question. This morning I had not yet earned the right to ask one.

The counter-intuitive angle: the temptation to fill the gap

The first reflex on seeing an empty table is to fill it in. The golf label is still there, and labels suggest things. It suggests a player, an event, a controversy. From one label, three inferential steps are enough to produce an analysis that reads very smoothly — and has no foundation whatsoever.

The troubling part is that this fabricated analysis will be harder to detect than a numerical error. It does not contradict the data, because there is no data. It only contradicts reality, and reality does not automatically appear in a validation table.

This is also where sports metaphor becomes dangerous. I still keep the habit of borrowing football's pressing vocabulary to dissect golf's rhythm of play — but only when the numbers prove the similarity. Gegenpressing does not break the data; it breaks my assumptions. To speak of recovery after a bogey as a counter-press, I need a scoring distribution by hole and a recovery rate after errors. Without those two things, the word gegenpressing is decoration placed next to a blank sheet.

What is scarier than a wrong number

A wrong result can be caught by cross-checking. An empty result is not caught — it is overlooked. So in a multi-tier pipeline, the receiving tier must have the right to refuse. Without a non-empty validation gate, the downstream tier will quietly construct a complete analysis out of nothing, and the final reader will never know.

There is one thing worth questioning myself on: are eight analytical dimensions too many for a single golf article, or just enough? I lean towards painfully just enough. The eight dimensions do not exist so that every one of them has content. They exist to show which one is empty — and to point at precisely that empty one instead of filling it with inference.

The same principle applies to the subjects I follow long term. A closed competitive ecosystem, with no progression places and no open competition, struggles to produce genuine stars — because stars are forged under the pressure of elimination. And at the youth-development level, pushing an unformed body into adult competitive rhythm produces impressive short-term numbers and then charges the bill in injuries. Both claims need data to defend them. Without data, they are merely tidy prejudices.

I do not believe in luck; I believe in probability that has been cultivated. Probability is cultivated with data collected under discipline. This morning, the greenhouse was empty.

The signal for the next cycle

The work to be done is not in the analysis. It is at the gate: a mandatory condition that tier two may only run when tier one returns at least a title, one information point, and a populated entity list. Better to stop early than to publish an analysis that is hollow inside but immaculate in form.

The question I keep for myself: if a pipeline can return zero without anyone noticing, then over the past seventeen years, how many times did I read a table that looked complete, when underneath there was nothing at all?