Pandas Timeseries Subplots with Rectangle

Question

I would like to create a plot using a pandas timeseries in one subplot and a rectangle in another subplot.

If I don't include the subplots, I can achieve this pretty easily:

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib.dates as mdates
import matplotlib.patches as mpatches

N = 100
np.random.seed(N)
dates = pd.date_range(start='2018-01-01', periods=N, freq='D')
one_third_delta = (dates[-1] - dates[0])/3
one_third_stamp = dates[0] + one_third_delta
ts = pd.Series(index=dates, data=np.random.randn(N))

def add_rectangle(ax, x, y, width, height, **kwargs):
    ax.add_patch(mpatches.Rectangle(
        (x, y),
        width,
        height,
        **kwargs
    ))

args = [one_third_stamp, -1, one_third_delta, 2]

kwargs = {
    'facecolor': 'orange',
    'edgecolor': 'None',
    'alpha': 0.5,
}

# Plot 1: 1 subplot with ts plotted first (Working)
fig, ax = plt.subplots()
ts.plot(ax=ax)
add_rectangle(ax, *args, **kwargs)
plt.savefig('plot1.png')
plt.close(fig)

Plot 1

However, things already start to get weird when I try adding the rectangle first:

# Plot 2: 1 subplot with ts plotted second (Not Working)
fig, ax = plt.subplots()
add_rectangle(ax, *args, **kwargs)
ts.plot(ax=ax)
plt.savefig('plot2.png')
plt.close(fig)

Plot 2

If I try splitting out the two plots, neither approach works:

# Plot 3: 2 subplots with ts plotted first (Not Working)
fig, axes = plt.subplots(2, sharex=True)
ts.plot(ax=axes[1])
add_rectangle(axes[0], *args, **kwargs)
plt.savefig('plot3.png')
plt.close(fig)

# Plot 4: 2 subplots with ts plotted second (Not Working)
fig, axes = plt.subplots(2, sharex=True)
add_rectangle(axes[0], *args, **kwargs)
ts.plot(ax=axes[1])
plt.savefig('plot4.png')
plt.close(fig)

Plot 3

Plot 4

I've found two work-arounds.

The first involves casting everything to a float with matplotlib.dates.date2num:

# Plot 5: 2 subplots with date2num (Working)
two_thirds_stamp = one_third_stamp + one_third_delta
args_date2num = [
    mdates.date2num(one_third_stamp),
    -1,
    mdates.date2num(two_thirds_stamp) - mdates.date2num(one_third_stamp),
    2,
]
df = ts.to_frame().reset_index()
df.columns = ['date', 'value']
df['num'] = df.date.apply(mdates.date2num)
fig, axes = plt.subplots(2, sharex=True)
add_rectangle(axes[0], *args_date2num, **kwargs)
axes[1].plot_date(df.num, df.value, ls='-', marker=None)
axes[0].set_ylim(axes[1].get_ylim())
plt.savefig('plot5.png')
plt.close(fig)

Plot 5

This isn't great for two reasons:

1. I lose the nice ticklabel formatting that pandas uses.
2. As far as I can tell, date2num is incompatible with how pandas internally represents datetimes as floats. So if I use date2num at all, all other datetimes must be converted too.

The other work around involves a dummy plot:

# Plot 6: 2 subplots with alpha=0 dummy (Working)
fig, axes = plt.subplots(2, sharex=True)
dummy_ts = ts[::(len(ts)-1)] + 10 # make it out of sight
dummy_ts.plot(ax=axes[0], alpha=0) # and invisible for good measure
add_rectangle(axes[0], *args, **kwargs)
ts.plot(ax=axes[1])
axes[0].set_ylim(axes[1].get_ylim())
plt.savefig('plot6.png')
plt.close(fig)

Plot 6

My question (finally) is why is this necessary? What changes between doing this on a single subplot vs. multiple? Is there a better, more canonical way?

---

Python version:

Python 3.6.3 (v3.6.3:2c5fed86e0) 
[GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin

Pip freeze:

cycler==0.10.0
kiwisolver==1.0.1
matplotlib==2.2.0
numpy==1.14.2
pandas==0.22.0
pyparsing==2.2.0
python-dateutil==2.7.0
pytz==2018.3
six==1.11.0

Answer 1

I think you found the reason yourself: Pandas datetime representation for the matplotlib axes (may) be completely different from the matplotlib date units (this is not always the case and depends on the span of the data).

Since I don't know of any way to convert the rectangle's coordinates to the pandas units, the only option is to plot the pandas plot in matplotlib units.

The problem

But let's start at the beginning. Case 1 and 2 work fine for me.

For the third case, the rectangle is added to the other axes, which does have a different scale. This can be seen by printing the transform.

def add_rectangle(ax, x, y, width, height, **kwargs):
    rect = mpatches.Rectangle( (x, y), width, height, **kwargs )
    ax.add_patch(rect)
    return rect

# Case 1 - working
fig, ax = plt.subplots()
ts.plot(ax=ax)
r = add_rectangle(ax, *args, **kwargs)
print r.get_transform()

# This prints
# BboxTransformTo(
#        Bbox(x0=17565.0, y0=-1.0, x1=17598.0, y1=1.0)),

# Case 3 - non-working
fig, axes = plt.subplots(2, sharex=True)
ts.plot(ax=axes[1], x_compat=True)
r = add_rectangle(axes[0], *args, **kwargs)
print r.get_transform()

# BboxTransformTo(
#        Bbox(x0=736728.0, y0=-1.0, x1=736761.0, y1=1.0)),

In the second case, the units are the matplotlib date units, because pandas did not change the transform for the axes in which it did not plot anything.

The solution

The easiest option is probably to tell pandas not to change the scale. This would be done using

x_compat=True

This has essentially the same effect as plotting everything in matplotlib units.

# Plot 3: 2 subplots with ts plotted first
fig, axes = plt.subplots(2, sharex=True)
ts.plot(ax=axes[1], x_compat=True)
r = add_rectangle(axes[0], *args, **kwargs)

# Plot 4: 2 subplots with ts plotted second
fig, axes = plt.subplots(2, sharex=True)
add_rectangle(axes[0], *args, **kwargs)
ts.plot(ax=axes[1], x_compat=True)

So indeed the nice pandas formatting is gone. But you may replicate it with the matplotlib.dates formatters. E.g. in this post. an easy solution to add the days is presented. Here, you would maybe rather use a FuncFormatter as follows:

fig, axes = plt.subplots(2, sharex=True)
ts.plot(ax=axes[1], x_compat=True)
r = add_rectangle(axes[0], *args, **kwargs)

import matplotlib.dates as mdates
import matplotlib.ticker as mticker

def f(val, _):
    d = mdates.num2date(val)
    if d.month == 1:
        return d.strftime("%b\n%Y")
    else:
        return d.strftime("%b")

axes[1].xaxis.set_major_locator(mdates.MonthLocator())
axes[1].xaxis.set_minor_locator(mdates.WeekdayLocator())
axes[1].xaxis.set_major_formatter(mticker.FuncFormatter(f))
fig.autofmt_xdate(rotation=0,ha="center")

producing