**The most surprising finding is that this isn't really a "top-player" tennis dataset, and the serve-stat gaps between surfaces are tangled up with who plays there.** Evidence comes from the per-surface summary of `raw.raw_kaggle` (237,205 player-match rows). 1. **The typical player is far from the elite.** - The median player rank is 237 on hard courts and 288 on clay, and the average rank is about 311 on hard and 361 on clay. - Most of the 117,571 hard-court and 107,650 clay-court rows therefore come from Challenger-level or lower-ranked players, not the ATP headliners most people picture. - Anyone building a model from it would be modelling the lower tiers. 2. **Surface effects on serving are confounded with field strength.** - Grass has the highest-ranked field (average rank about 151, median 109) and averages 7.57 aces per match. Clay averages 3.38, a gap of about 2.2×. - That could be read as grass simply drawing stronger servers. - Carpet undercuts that reading. It has the weakest field (average rank about 517) yet the most aces, at 7.94 per match. - So fast surfaces seem to boost aces regardless of player quality, but the simple averages can't fully separate the two effects. 3. **Double faults barely depend on surface.** They sit between 2.8 and 3.2 per match on every surface, while aces vary more than 2×. Surface changes how often a serve can't be returned, not how often a player misses. 4. **More than a third of the rows have no match statistics.** 86,793 of 237,205 rows (about 37%) have null TP, Aces, DFs, SP, 1SP, 2SP and vA. Any serve-related average, including those above, rests on only the roughly 63% of matches that have stats. I did not test whether the missing statistics are concentrated in lower-tier events. That would be the next thing to check, because it would show whether the data is even more skewed toward elite matches than the averages suggest.