Biltmore Championship: Davis Thompson and the Empty-Data Problem at The Cliffs at Walnut Cove

On October 6, as I downloaded the PGA Tour's Strokes Gained table to prepare for the FedExCup Fall opener, one number made me reopen the file three times. Davis Thompson ranked No. 1 in the field in SG: Approach over the last 36 rounds. Right behind him sat Ben Kohles at No. 2. Two names sitting side by side in exactly one metric, at exactly one golf course for which no analyst in the field has modern ShotLink data. The Cliffs at Walnut Cove in Asheville, North Carolina, designed by Jack Nicklaus, measures 7,249 yards at par-71. The last time it hosted a PGA Tour-level event was nearly two decades ago. That span is longer than the professional career of most golfers in this week's field. I stared at those two data rows and asked my usual question: am I reading a model, or am I reading a coincidence dressed up as one? On weeks like this, I always pull out my old notebook and write down the original question before writing anything else, so I do not change the question midway when the data pushes back. My original question today: does the approach metric actually predict results on a course for which I have never had a modern data sample? To understand why this is a hard week for anyone working in golf data, you have to start with the structure of the FedExCup Fall. It is the post-playoff stretch where golfers hovering near the Tour Card cut line fight to preserve their status. For a player like Kohles, a top-40 target is not a modest goal. It is an equation about next season's card, about priority access to Signature Events, about standing within the system. Put another way, the FedExCup Fall is the least glamorous but systemically heaviest stage of the entire season. No majors, no large ranking points, very little television light. In exchange, this is where the careers of mid-tier players get decided. The Biltmore Championship is an entirely new event. By itself it carries both variance and opportunity. Commercially, the PGA Tour is bringing a new event to a new market to sustain attention after the playoffs end. In data terms, this is a blind spot. I encountered this situation once, in 2026, when the pandemic emptied stadiums and my club Nagoya Grampus went two months without a match. I had to rebuild the performance prediction model under conditions with no match data. I chose to use GPS training data from the youth team and precedents from historically disrupted seasons, specifically J.League 2026 after the earthquake disaster. When match data disappears, you are forced to find substitute data at a different layer. That is exactly what is happening at Walnut Cove this week. The course has no modern history. The field has no recent-season data. My job is to shift from the question "who fits this course" to "which metric remains trustworthy when the course context is unverified." This is where the original PGA Tour article takes a different approach from mine. It calls its candidate group "horses for courses." On a course with no history, that concept only works if you use a golfer's technical profile to reason about the course, not use course history to reason about the golfer. That is an important inversion, and it is where I want to linger. The course profile, per the organizers' description, has three distinctive traits. Extremely wide fairways. Elevation above sea level that makes the ball travel farther than normal. Rough about 5 inches deep, a number in the severe category for modern golf. And a par-71 layout, meaning three par-5s and five par-3s, generating more scoring chances than a standard par-72. These three traits do not point in the same direction. Wide fairways let players swing freely off the tee, reducing the penalty for errant drives. But 5-inch rough punishes precisely those errant drives. Elevation pushes the ball farther, turning par-5s into genuine birdie chances. And par-71 creates a scoring structure where birdies are almost mandatory to contend. I imagine the precedent. When a parkland course with wide fairways and elevation appears on the PGA Tour schedule, metrics tend to amplify in two directions: driving distance gains value, and iron quality determines scoring quality. That is why I want to examine the ordering of SG metrics in this week's field. Davis Thompson ranks No. 1 in SG: Approach over 36 rounds. He also ranks 6th in birdies-or-better gained over 24 rounds, 2nd in scoring opportunities, and 2nd in birdie-or-better percentage from rough. This is a rare data set. It does not merely show he hits well from the fairway; it shows he converts chances once the ball has left the short grass. Ben Kohles ranks No. 2 in SG: Approach over the same 36-round window. He has had a summer with several contention weeks, including a T3 at the John Deere Classic and another top-10 at the 3M Open. At the U.S. Open, where field quality is far higher, he gained nearly two strokes on approach. That is a critical data point, because it shows his ball-striking held up against stronger opposition. My question now is: is leading the approach metric enough to predict a good week on this course? This is where I want to dig a little deeper, because it touches methodology, not just outcome. In post-break weeks, data is always thinner than normal. The PGA Tour offers three different sample windows: 24 rounds, 36 rounds, and 50 rounds. None is an absolute standard. 24 rounds captures recent form but is noisy. 50 rounds is more stable but blends in older form. 36 rounds is the midpoint, which is why the original preview chose it. But the mere existence of three windows side by side is itself a signal: the writer is managing a narrow data window. Gaps in a table can speak too, if we are willing to listen. The fact that neither Thompson nor Kohles was given an SG: Putting or SG: Around the Green figure in the original preview tells me two things. First, their profiles are incomplete in the writer's hands. Second, it is possible the writer selected metrics that flatter the thesis rather than presenting a full picture. I am not accusing anyone here. I am talking about genre. The original preview belongs to the betting commentary genre, and in that genre, metric selection is normal. But the reader needs to know what they are reading. Thompson has a weakness the author himself names: his putting is questioned. The author writes that he is not sure Thompson's putter can cooperate for four rounds. That is an important sentence. It explains why the analytical frame tilts toward the FRL market, the First-Round Leader, rather than the tournament winner market. Logically, this is consistent. If a player has elite scoring upside but unproven four-round reliability, placing him in a one-round frame makes sense. But there is an underlying layer to read: Thompson's putting is a gap not yet filled by data. For Kohles, the story lies in the distance between contending and winning. He has had several weeks near the lead but has not yet sealed a first PGA Tour win. In behavioral data, this is a type of signal about conversion ability. It does not say he lacks the skill. It says that when psychological pressure compresses to the final moment, he has no successful sample to draw on. That is why a floor-type market like top-40 fits Kohles's data better than a win market. Top-40 does not require him to pass the biggest psychological test. It only requires him to maintain a stable ball-striking level across four rounds. And that is what his data supports. I have a concrete lesson about this kind of gap, from J.League 2026. At 24, I began doing data analysis for Nagoya Grampus, then in J.League 2 after relegation. I built a manual xG model from video. But I missed a four-match losing streak because I failed to account properly for the home-venue factor. The result: I got 6 of the final 10 rounds wrong. When I sat down, reviewed all the footage, and cross-checked every sequence, I realized raw data was not enough. It lacked tactical context and lacked the psychological variable of competition. Data is never wrong; I just asked the wrong question. In 2026, I asked "which team is stronger." The right question should have been "which team is stronger under this match's specific conditions." The difference between the two questions is the context layer. At Walnut Cove this week, a similar question appears. I could ask "who hits approach best." The answer is Thompson, then Kohles. But the better question is "how much value does the approach metric retain when the course context is unverified." To answer that better question, I have to walk through each variable. Variable one: wide fairways. This reduces the value of SG: Off the Tee in the sense that the penalty for errant drives falls. But 5-inch rough reverses part of that effect. So the metric that matters is not raw driving distance but the ability to convert from rough. Thompson ranks 2nd in birdie-or-better percentage from rough. That is a data point that maps directly onto the course's character. Variable two: elevation. The ball travels farther, meaning par-5s become more attackable. Thompson ranks 2nd in scoring opportunities, meaning he consistently creates birdie chances. Kohles, with a solid approach foundation, has the same structural edge. Variable three: par-71. This structure generates more chances than a par-72. But it also means the winning score will be lower, and every mistake costs more. On a course demanding continuous birdies, whoever cannot score will fall back fast. These three variables converge on the same technical conclusion: the course profile favors players with high iron quality and the ability to score from multiple positions. That is exactly the profile of both Thompson and Kohles. But this is also where I must stop and check myself. There is a trap in the entire argument above, and I want to name it before someone else does. The trap is this: I am building a model about a course for which I have never seen data. Everything I did above is reasoning from a course description, not from course data. Wide fairways, elevation, 5-inch rough, par-71 — those are design parameters, not operating parameters. A golf course in competition can play very differently from its blueprint, depending on pin placements, rough cutting, and green speed. With Walnut Cove, none of us will know how the course plays until the first round ends. This is a genuine data void, and it cannot be filled by inference. When data hides its face, error becomes the guide. What I can do is not pretend I know how the course will play. What I can do is determine which variable is most likely to carry error, and which variable retains value even if the model fails. The approach metric is the variable most likely to carry error. Why? Because iron quality depends on many contextual factors: elevation, wind, grass thickness, pin placement. On an unverified course, a player ranked No. 1 in approach across other courses may not hold that position here. Conversely, the birdie profile — the ability to create and convert chances — is a more stable variable, because it reflects behavioral patterns rather than shot conditions. Thompson ranks 6th in birdies gained over 24 rounds. That is a pattern signal, not a moment signal. In other words, if I had to pick one variable to trust this week, I would choose Thompson's birdie profile over his No. 1 approach ranking. Not because approach matters less. Because approach is more sensitive to course context, and course context is the unknown. This is a small shift from the conventional reading. Conventionally, people put Thompson on top because he ranks No. 1 in approach. I want to separate the two. He leads a metric sensitive to context. What is more trustworthy is his scoring behavioral pattern, sitting near the top in birdies. With Kohles, the trap runs in reverse. He ranks No. 2 in approach, but his real story is not approach. It is the gap between contending and closing. That is a behavioral variable, not a technical one. And behavioral variables are also sensitive to competitive context. On a new course, with a new field, Kohles's closing pressure could rise or fall depending on how the week unfolds. If he leads entering the final round, the old question returns. If he sits in a safe group, his floor is very solid. That is why I find the top-40 frame more reasonable than the win frame for Kohles, but I also do not treat it as a certain conclusion. It is a conditional conclusion, tied to Kohles not being pushed into a direct contention position over the final two rounds. There is one more layer I want to make clear. The original preview presents a ball-striker group including Thompson, Kohles, Blades Brown, William Mouw, Tom Hoge, and Chris Kirk. Choosing a group rather than one individual is methodologically correct, because it acknowledges that the data is not enough to settle on a single name. But it also means the "horses for courses" argument is doing more rhetorical work than analytical work. A course with no modern history cannot be used to prove the idea of "horses for courses" in the traditional sense of the phrase. Traditionally, you compare a golfer's profile against course history. Here you compare a golfer's profile against a course description. The distance between those two acts is the distance between evidence and inference. Every number is a confession not yet written into prose. The "No. 1 approach" line confesses that the player hit irons well within a specific sample window. It does not confess that the player will hit irons well on a course that is elevated, wide, and thick with rough. I must distinguish the two. This is where I return to the original question I wrote in my notebook before beginning this piece. Original question: does the approach metric actually predict results on a course for which I have never had a modern data sample? My most honest answer after walking the entire chain of reasoning is: the approach metric predicts player quality, but it may not predict results on this course. It is a necessary condition, not a sufficient one. The sufficient condition will appear when the first and second rounds end, when we know how the course actually plays. I will track four signals this week to test my model. Signal one is the first-round scoring level. If the whole field scores well below par, the birdie-fest thesis is confirmed. If the course plays severe, the entire analytical frame must adjust. Signal two is Thompson's SG: Putting by round. This is the gap the preview author himself named. If his putting runs hot, his FRL frame has a basis. If it runs cold, the rest of his profile cannot compensate. Signal three is Kohles's closing ability in the final round. This is the behavioral variable I want to see with real data, not just historical data. Signal four is the actual punitive effect of the rough. If 5-inch rough truly decides scoring, Thompson's "birdie from rough" profile gains even more value. I do not believe in luck; I believe in cultivated probability. What I am doing this week is not predicting who wins. It is defining in advance which variables will confirm or break my model, so that when the first round closes, I know where I was right and where I was wrong. And if the model is wrong, I will say so publicly. Not because I want to appear humble, but because that is the only way I ask a better question next time.