# Which columns are most related? Partial answer **I never computed a correlation, so I can't rank column pairs by strength.** All three tool calls returned the same column listing. The notes below come from the schema and the null counts, not from any correlation statistic. ## What the evidence shows **1. Six columns share an identical null count, so they are missing together.** `TP`, `Aces`, `DFs`, `SP`, `1SP`, `2SP` and `vA` each have exactly **86,793 nulls**. That is seven columns, not six. The match suggests they come from the same source: match-level serve statistics that were either recorded for a row or not recorded at all. This is a relationship in *missingness*, and it is the only one the data directly confirms. **2. The rankings columns are probably related, but I didn't test it.** `Rk` (the player's rank, 1–2225, 4,388 nulls) and `vRk` (the opponent's rank, 1–2252, 10,390 nulls) sit on the same scale. Their null counts differ, so they are not missing together. **3. Some serve-statistic columns are likely related by definition.** These are plausible from the column names and ranges only. I did not check them against the data. - `SP` (service points, max 253) should be roughly `1SP` + `2SP` (max 165 and 112). - `TP` (total points, max 499) should be at least as large as `SP`. - `Aces` and `DFs` are probably tied to the number of service points. - `vA` (range 0–61) is probably the opponent's aces, which would relate it to `Aces` and to match length. **4. Some columns are redundant by construction.** - `_ingestion_timestamp`, `_batch_id`, `_source_file` and `_source_system` each have 1 distinct value. They carry no information and have no meaningful correlation with anything. - `Name` and `against` both identify players, and `against` is a messy free-text field (~252,129 distinct values). They may overlap in content, but I didn't test that. ## What remains uncertain - The actual Pearson or Spearman correlations between the numeric columns (`Rk`, `vRk`, `TP`, `Aces`, `DFs`, `SP`, `1SP`, `2SP`, `vA`) were not computed. - I did not test whether rank differences relate to serve stats, or whether `Surface` or `Rd` relate to ace counts. - I don't know whether the 86,793 null rows are the same rows across all seven columns. The matching counts make that likely but don't prove it. ## Suggested next step Run a correlation matrix over the nine numeric columns, restricted to rows where `TP` is not null. I'd expect the strongest pairs to be `SP`–`1SP`, `SP`–`TP`, and `Aces`–`SP`. Treat that as a hypothesis until it's computed.