Charts
Statistical charts
Distributions, agreement and uncertainty. Every value here is prepared by the caller: the library draws the ECDF it is given, the quantiles it is given and the interval it is given, because that keeps a number in a chart answerable to exactly one place.
Compare two samples
The empirical CDF and the Q–Q plot answer different questions about the same pair. The CDF shows where one sample sits relative to the other; the Q–Q plot shows where the shape stops agreeing, which is usually the tails.
Empirical CDF
240 synthetic sessions per sample. Drawn as a step — interpolating between order statistics invents mass the sample does not have.
Q–Q plot
Sample quantiles against normal quantiles. Departure from the 45° line is the finding.
Shape, spread and uncertainty
A violin is a density outline the caller computed, an error bar is an interval the caller supplied, and a residual plot is the model's own leftovers. None of them fit anything.
Violin
Three prepared density outlines with a supplied median marker.
Intervals
Point estimates with supplied intervals. The interval drawn is exactly the interval given.
Residuals
Residual against fitted value, with the zero line a well-behaved model sits around.
Classifiers
ROC, calibration and the confusion matrix are the three views of one classifier. The AUC is printed from the caller's number rather than integrated here, and the matrix can normalise per row so a large class cannot dominate a small one.
ROC
Two curves with their supplied AUC. The diagonal is chance, not a score.
Calibration
Observed frequency against predicted probability. Bubble area is the bin count, so a sparse bin cannot masquerade as a strong one.
Confusion matrix
Row-normalised, so a 900-sample class and a 40-sample class are comparable.