Tick formats¶

The locator chooses a tick's value and matplotlib writes it; its two rules are Talbot's legibility criterion at work, and the unit belongs in the axis label, once.

A tick at five million can read 5000000, 5 with 1e6 in a corner, or 5 under an axis labelled "revenue (millions)". The range frame decides where the tick sits and leaves the writing to matplotlib, which follows two rules that stay invisible until a figure comes out wrong.

Matplotlib's two rules¶

The first is a window. From a hundred-thousandth up to a million, every tick is written out in full. Outside it, matplotlib divides all the labels by a common power of ten and parks the factor in the corner of the axes, so revenue from 1 to 9.4 million ticks as 2, 4, 6, 8 with 1e6 sitting in the corner. axes.formatter.limits holds the window's edges.

The second fires inside the window, when the data cover a narrow band far from zero. Matplotlib then subtracts a common value as well, so an index hovering around 100 ticks like this:

99.900 .. 100.075   →   −0.100  −0.075  −0.050  −0.025  0.000  0.025  ...   +1e2

Both rules exist to make the labels shorter. That is Talbot, Lin and Hanrahan's fourth criterion, legibility, which their tick search scores and mizani's leaves out; matplotlib's formatter applies it after the search instead, and whose decision is which says why the library leaves it there.

Say it in the label, once¶

The offset labels are not wrong. They are a claim: that the deviation is what matters and the level is context. For an index that is often exactly right, and the offset makes a comparison legible that 99.900, 99.925, 99.950 would bury. What the formatter cannot know is when the level is the story instead, and it makes its claim silently, with the evidence in a corner the reader may never look at.

So make the claim yourself, and make it where the reader arrives: at the end of the axis, in the label, which is where Tufte and Doumont put a unit rather than on every tick. When the deviation is the story, keep the offset and name it:

ax.plot(year, index)
ax.vzs.apply()
ax.vzs.set_ylabel("index (deviation from 100)")

When the level is the story, divide the data and name the unit:

ax.plot(year, revenue / 1e6)
ax.vzs.apply()
ax.vzs.set_ylabel("revenue (millions)")

The ticks read 2 through 8 either way. The difference is that "millions" now sits at the end of the axis in your own words, where xlabel and ylabel put it, instead of 1e6 in a corner that the range frame does not place: the corner text is matplotlib's, so on a numeric axis it can land far from a spine that stops at the data.

Either rule turns off on its own, after apply, the common factor with style="plain" and the common value with useOffset=False:

ax.ticklabel_format(axis="y", style="plain")
ax.ticklabel_format(axis="y", useOffset=False)

For a whole figure or session, axes.formatter.useoffset and axes.formatter.limits do the same as rc parameters.

Dates¶

A date axis gets the same treatment from a different formatter. apply installs ConciseDateFormatter, so the labels shorten to what changes between ticks, a run of months does not repeat the year, and the shared year is the offset rule in date form: it moves to the right end of the bottom spine, where xlabel goes, and stacks above an xlabel if you set one. The ticks themselves sit on calendar starts, a year, a month, a day, chosen by the locator.

Where every tick should carry the same stamp, set a formatter after apply:

import matplotlib.dates as mdates

ax.xaxis.set_major_formatter(mdates.DateFormatter("%b %y"))  # Jan 26, Apr 26