jackbenn / graphing_tools Goto Github PK

The CDF would be more useful with quantile lines (horizontal to the left of the curve, vertical under the curve)

Also it should match multihist better in its options.

Create a plot that shows the KDE together with some confidence interval for the density at each point, so there's a minimum and maximum curve as well as the base curve. Presumably it would have a greater range in low-density areas, though also the overall distance between the upper and lower bounds would be a good measure of the uncertainty.

The QQ and PP plots handle distributions, but not the CDF plots or the QP matrix.

As part of that, the various plots should change scatter->plot to allow a line between points (appropriate for dist vs dist plots).

The probability plots aren't quite aligned right. In particular, the spacing on matching distributions should probably be at the midpoints of the bins, and the probability plot shouldn't use a strict inequality.

Add options for best-fit line. Consider all the statsmodels options (https://www.statsmodels.org/stable/generated/statsmodels.graphics.gofplots.qqplot.html) in particular, the 45-degree line and regression line, and also a PCA line (since linear-regression treats the horizontal and vertical axes differently).

The histogram should be able to do bounded kdes, either on one side or both. It might have options between reflection and transformation, though there may be a better option. Possible parameters:

kde_bounds: tuple of lower and upper bounds, with None meaning that side is unbonded.
kde_bounding: 'reflect', 'transform'

Add an option to encourage the bins to line up on integers, powers of tens, or (less frequently) small powers or two/five times those. There's a tradeoff between getting the right number of bins and minimize the power of two/five.

Possible multiples of powers of 10:
0: 1
1: 2, 5
2: 2.5, 4
3: 1.25, 8
4: 1.6, 6.25
5: 3.125, 3.2

We then choose the step size that minimizes the sum of
alpha times the number of steps, plus
the square of the log of the ratio between the requested number of bins and the number of bins using that ratio

Not sure of the proper value of alpha yet, that will take experimentation, and maybe it will be adjustable. Likely the default will prevent anything past 1 or 2, and if adjustable it will probably be capped at 5.

Numpy has some algorithms for choosing bins. Matplotlib uses them, but multihist doesn't recognize them.

https://numpy.org/devdocs/reference/generated/numpy.histogram_bin_edges.html

Also: it have an option to detect discrete values and just plot those as a bar chart.