Arviz: Plot visualizing quantiles in time series

[ariddell]

It would be great to have a nice, generic version of this:

See https://github.com/mwaskom/seaborn/issues/450 for discussion.

The suggested matplotlib recipes are wanting by comparison, I think (http://matplotlib.org/users/recipes.html).

[twiecki]

+1

[ariddell]

I'd settle for something simpler too, maybe something with a point at the center of a line, like this:

p + geom_pointrange(aes(ymin = lower, ymax = upper))

from

http://ggplot2.tidyverse.org/reference/geom_linerange.html

Looks like plotnine has this.

[ahartikainen]

We could do this with np.percentile + for loop with fill_between.

Do we want make this smooth? Then kde / spline method could be used, but not sure if it's appropriate to interpolate with smooth function.

[ariddell]

That's a good idea.

(I'm not sure this particular plot is high-priority, it's just something I find myself making often with Stan output.)

[ahartikainen]

Here is a sketch for tsplot. I don't know how much options we should add.

import numpy as np
import matplotlib.pyplot as plt
plt.style.use('ggplot') # this was just used for the examples

# data
t = np.linspace(0,100,100)
y = 5 * np.sin(t/10) + 4*np.random.randn(100*150).reshape(150, 100)
y_ = 5 * np.sin(t/10) + 4*np.random.randn(100*4000).reshape(4000, 100)

t__ = np.linspace(0,100,6) 
y__ = 5 * np.sin(t__/10) + 4*np.random.randn(6*4000).reshape(4000, 6)

def tsplot(x, y, n=20, percentile_min=1, percentile_max=99, color='r', plot_mean=True, plot_median=False, line_color='k', **kwargs):
    # calculate the lower and upper percentile groups, skipping 50 percentile
    perc1 = np.percentile(y, np.linspace(percentile_min, 50, num=n, endpoint=False), axis=0)
    perc2 = np.percentile(y, np.linspace(50, percentile_max, num=n+1)[1:], axis=0)

    if 'alpha' in kwargs:
        alpha = kwargs.pop('alpha')
    else:
        alpha = 1/n
    # fill lower and upper percentile groups
    for p1, p2 in zip(perc1, perc2):
        plt.fill_between(x, p1, p2, alpha=alpha, color=color, edgecolor=None)


    if plot_mean:
        plt.plot(x, np.mean(y, axis=0), color=line_color)


    if plot_median:
        plt.plot(x, np.median(y, axis=0), color=line_color)

    return plt.gca()

Then can be called either in one go:

tsplot(t, y, n=100, percentile_min=2.5, percentile_max=97.5, plot_median=True, plot_mean=False, color='g', line_color='navy')

tsplot(t, y, n=20, percentile_min=2.5, percentile_max=97.5, plot_median=True, plot_mean=False, color='g', line_color='navy')

tsplot(t, y, n=5, percentile_min=2.5, percentile_max=97.5, plot_median=True, plot_mean=False, color='g', line_color='navy')

tsplot(t, y, n=2, percentile_min=2.5, percentile_max=97.5, plot_median=True, plot_mean=False, color='g', line_color='navy')

tsplot(t, y_, n=100, percentile_min=2.5, percentile_max=97.5, plot_median=True, plot_mean=False, color='g', line_color='navy')

or multiple times

# IQR
tsplot(t, y_, n=1, percentile_min=25, percentile_max=75, plot_median=False, plot_mean=False, color='g', line_color='navy', alpha=0.3)
#  90% interval
tsplot(t, y_, n=1, percentile_min=5, percentile_max=95, plot_median=True, plot_mean=False, color='g', line_color='navy', alpha=0.3)

tsplot(t__, y__, n=1, percentile_min=25, percentile_max=75, plot_median=False, plot_mean=False, color='g', line_color='navy', alpha=0.3)
tsplot(t__, y__, n=1, percentile_min=5, percentile_max=95, plot_median=True, plot_mean=False, color='g', line_color='navy', alpha=0.3)

plt.text(0, -5, "n=2", fontsize=14)

[ariddell]

These look great!

I'm inclined to do something very simple like have the posterior mean and single 90% band. This seems to be what is used in Bayesian Data Analysis (e.g., chp 21).

[ariddell]

FYI: https://github.com/jasonhilton/ggfan

Not sure I like this. The simpler envelopes are clearer.

[ChampionZP]

helpful !

[ColCarroll]

I'm leaving this closed as out of date, but feel free to reopen if there is interest in taking it up.

Also of interest, it closed because of

:joy: