Annotation

You can build a correct, labelled, unobjectionable chart that says nothing at all — the reader is left to work out for themselves what they were meant to notice. Annotation is where you tell them.

ax.text

ax.text(3, 60, "midweek lull", ha="center", color="#9a9a9a")

Words at a point in data coordinates — so they move when the limits change. ha and va decide which part of the text sits on that point, and getting them wrong is why annotations so often look slightly adrift.

Notes should be quieter than the data: smaller, greyer. They're support, not content.

ax.annotate

This is the one to actually learn:

i = int(np.argmax(delhi))
ax.annotate(f"best day: {delhi[i]}",
            xy=(i, delhi[i]),            # what you're POINTING AT
            xytext=(i - 2, delhi[i] + 8), # where the WORDS go
            arrowprops=dict(arrowstyle="->", color="#9a9a9a"))

Two coordinates: xy is the target, xytext is the label. Everything else about annotate follows once those are straight. arrowprops draws the connector, and connectionstyle="arc3,rad=0.2" curves it around whatever's in the way.

Notice that the interesting point is computed — np.argmax — not typed in. Hardcoding xy=(5, 88) works today and points at empty space the moment the data changes.

Axes coordinates

ax.text(0.03, 0.95, "n = 120",
        transform=ax.transAxes, va="top",
        bbox=dict(facecolor="#161616", edgecolor="none", alpha=0.8))

transform=ax.transAxes switches to fractions of the axes: (0, 0) is bottom-left, (1, 1) is top-right. That's how you pin a note to the corner and have it stay in the corner when the data underneath changes. fig.transFigure does the same for the whole canvas — sources and footnotes.

Label the lines directly

A legend makes the eye travel to a key, decode a colour, and travel back. Writing each series' name at the end of its own line removes that trip:

for name, line in zip(names, ax.lines):
    ax.text(x[-1] + 0.1, line.get_ydata()[-1], name,
            color=line.get_color(), va="center")
ax.margins(x=0.12)      # make room on the right

Taking the colour from line.get_color() is the trick that makes it hold together — the label is the key. This is the change that most improves a multi-series chart, and almost nobody makes it.

Highlighting a period

ax.axvspan("Sat", "Sun", alpha=0.1, zorder=0)
ax.text(5.5, 95, "weekend", ha="center", fontsize=9)

A shaded band for a closure, a campaign, a lockdown. Faint, and zorder=0 to keep it behind the data. It carries context the numbers can't.

The title is the claim

Last and most valuable. "Cups by day" describes the axes, and the axes already describe themselves. Promote the finding instead:

ax.set_title("Weekend trade is 40% above weekdays", loc="left", fontsize=13)

Now the chart is evidence for a stated claim rather than a puzzle. loc="left" reads as a headline, and a smaller grey second line beneath it can carry the detail.

The discipline: write the sentence you want the reader to leave with, then check the chart supports it. If it doesn't, you've learned something more useful than a title.