Auto Analyst: Instant Statistical Report

Notes

• Every pair of columns gets the test that fits their types: Spearman’s rank correlation for two numbers, Mann–Whitney or Kruskal–Wallis for a number across groups, and a chi-square test (Fisher’s exact test on small 2×2 tables) for two categories. The p-values are then adjusted together with the Benjamini–Hochberg procedure, so that on average no more than one in twenty reported relationships is a false alarm. The statistics agree with SciPy to about thirteen significant digits.
• Strength uses one scale for every test: ρ, the rank-biserial r, √η² or Cramér’s V below 0.1 is negligible, below 0.3 weak, below 0.5 moderate, and 0.5 or more strong. Findings are ranked by strength, not by p-value, because with enough rows even trivial effects become “significant”.
• The paradox scan compares each association across all rows with the same association inside every grouping column of two to eight groups: a pooled within-group correlation, a Mantel–Haenszel odds ratio and risk difference, or a weighted within-group difference of means. It reports patterns that reverse and patterns that vanish.
• Key drivers come from a random forest grown in the browser. Its score and the importance of each column are measured on out-of-bag rows, the ones each tree never saw, and the curves are partial dependence: the average prediction as one column changes. Drivers are predictive, not causal.
• Data checks: values missing more for some groups than others, columns missing together, exact duplicate rows and IDs, placeholder codes such as 999 or -1 (treated as missing), spelling variants of one label (merged), values beyond three interquartile ranges, ambiguous dates such as 03/04, and columns that repeat each other.
• Large files: the pairwise tests run on a random sample of 20,000 rows and the forest on 6,000. At that size a correlation is pinned down to about ±0.01, so the sample costs almost nothing in precision.
• Sample data: Palmer penguins from the palmerpenguins package by Horst, Hill and Gorman (CC0; collected by Gorman, Williams and Fraser at Palmer Station, Antarctica); Berkeley 1973 graduate admissions from Bickel, Hammel and O’Connell, Science 1975; the customer churn table is invented.
• Runs 100% in your browser — your files and data never leave your device.