Choosing the right chart

Two chapters of matplotlib and seaborn gave you the power to make any chart. This one — the start of the storytelling chapter — is about making the right one. Because every chart type exists to encode a specific relationship between variables in a visual form that the human eye can decode quickly and accurately. The mistake most people make is starting with a chart they like and fitting data into it. The better move: start with the question, then pick the encoding that answers it most directly.

Start with the question, not the chart

There are five core analytical questions. Each maps cleanly to a family of chart types.

1. Comparison — How does one category’s value compare to another’s?
2. Trend — How does a value change over time?
3. Distribution — How are values spread across a range?
4. Relationship — Do two (or more) variables move together?
5. Composition — What share of a whole does each part contribute?

The diagram below is your decision guide. Read the left column as the question in your head; follow the arrow to the chart family on the right.

The five chart families in detail

Comparison — bar chart

A bar chart encodes magnitude as length. Length is one of the most accurately perceived visual properties humans have — far more accurate than angle or area. When you want a reader to answer “which is bigger?” or “by how much?”, length wins.

• Sort bars by value (not alphabetically) unless there is a meaningful ordinal sequence.
• Use horizontal bars when category labels are long so they stay readable.
• Keep the y-axis starting at zero. Cutting the axis makes the differences look proportionally larger than they are — a genuine misrepresentation.

Trend — line chart

A line chart encodes change over time using slope and position. The connecting line is a promise: “there is continuity between these points.” That promise is appropriate when the x-axis is temporal (a date or time) or a genuinely continuous variable. If you connect bars representing discrete categories with a line, you imply an ordering and continuity that does not exist.

Distribution — histogram, box plot, violin plot

A histogram bins a continuous variable and shows how many values fall in each bin. It answers “where are most values clustered?” and “are there outliers or multiple peaks?”

A box plot summarises a distribution with five numbers: minimum, first quartile, median, third quartile, maximum. It is compact and great for comparing distributions across several categories side by side.

A violin plot is a box plot with a kernel density estimate overlaid — it shows the full shape of the distribution, not just five summary statistics.

Use a histogram when you have one continuous variable and want to show shape. Use box or violin when comparing distributions across groups.

Relationship — scatter plot

A scatter plot places one variable on each axis and draws a dot for each observation. The resulting cloud shows correlation (do the dots slope upward?), clustering (do points group together?), and outliers (which points are far from the mass?).

Add a regression line if you want to make a trend claim explicit — but remember: the scatter cloud carries the honest story, the line is a summary.

Composition — stacked bar (and why pie charts are usually worse)

A stacked bar chart encodes part-to-whole relationships as segments of a bar. The baseline for each segment is its left (or bottom) edge, and length differences between segments are readable because they share a common baseline.

A pie chart encodes composition as angle and arc. The human visual system judges angles and areas much less accurately than lengths with a shared baseline. Experiments consistently show that viewers struggle to tell whether a 34% slice is larger than a 38% slice when they are not adjacent. With more than two or three slices the errors compound further.

When a pie is acceptable: exactly two or three slices where one is a clear majority (for example, 75% vs 25%). For everything else, use a stacked bar — or better yet, a simple bar chart that lets each value stand on its own baseline.

The data-type lens

Before you commit to a chart, classify your variables:

• Categorical — named groups with no natural order (product names, countries, colours). These map to the x-axis of a bar chart or the grouping dimension of a scatter.
• Continuous — measured numbers that can take any value in a range (revenue, temperature, height). These drive histograms, scatter axes, and the y-axis of line charts.
• Temporal — dates and times. A special ordered type that implies continuity and should use a continuous time axis in a line chart — not discrete bars at monthly intervals unless you are making a deliberate aggregation point.

Mismatching axis type to variable type is a subtle but serious error. Plotting monthly revenue as a bar chart is fine; treating discrete survey responses as a continuous line is not.

Bad chart vs right chart — runnable demo

The same five data points, two encodings: a pie chart that forces you to compare angles, and a bar chart that uses length. See which one makes the ranking obvious.

import matplotlib.pyplot as plt
import numpy as np

labels = ["North", "South", "East", "West", "Central"]
values = [31, 28, 17, 14, 10]
colors = ["#6366f1", "#8b5cf6", "#a78bfa", "#c4b5fd", "#ddd6fe"]

fig, axes = plt.subplots(1, 2, figsize=(10, 4))
fig.suptitle("Same data — two encodings", fontsize=13, fontweight="bold", y=1.02)

axes[0].pie(values, labels=labels, colors=colors, autopct="%1.0f%%",
            startangle=90, wedgeprops=dict(linewidth=0.5, edgecolor="white"))
axes[0].set_title("Pie chart\n(compare the angles...)", fontsize=11)

sorted_idx = np.argsort(values)[::-1]
sorted_labels = [labels[i] for i in sorted_idx]
sorted_values = [values[i] for i in sorted_idx]
bars = axes[1].barh(sorted_labels, sorted_values,
                    color=[colors[i] for i in sorted_idx], height=0.6)
axes[1].set_xlim(0, 40)
axes[1].set_xlabel("Share (%)")
axes[1].set_title("Bar chart\n(rank is immediate)", fontsize=11)
for bar, val in zip(bars, sorted_values):
    axes[1].text(val + 0.5, bar.get_y() + bar.get_height() / 2,
                 f"{val}%", va="center", fontsize=10)
axes[1].spines[["top", "right"]].set_visible(False)

plt.tight_layout()
plt.show()

Compact decision guide

When no single type fits perfectly, ask which relationship matters most to your audience and encode that relationship first. A secondary relationship can live in colour, size, or a small multiples layout.

In one breath

Start from the question, not the chart gallery. Five analytical questions each map to a family: comparison → bar (length is the most accurately perceived encoding — sort it, start at zero); trend → line (the connecting line promises continuity, so use it only on a temporal/continuous axis); distribution → histogram for one variable, box/violin to compare across groups; relationship → scatter (the cloud carries the honest story); composition → stacked bar, not pie (angles and areas are judged far worse than lengths on a shared baseline). Then apply the data-type lens — categorical, continuous, temporal — and match the axis to the variable. The two classic lies to avoid: a truncated y-axis that inflates small differences, and a pie with more than two or three slices.

Practice

A question to carry forward

You can now pick the chart that answers your question. But picking the right chart isn’t the same as making it land. The pie-vs-bar demo hinted at it: even the correct bar chart, dropped raw onto a slide, just sits there as five numbers. A reader still has to work out what it means and what they should do.

So the question to carry forward is: once the chart is the right one, how do you shape it into a story — an opening that frames the stakes, a single clear takeaway, the eye guided to the one bar that matters? The next lesson, the narrative arc of a chart, turns a correct plot into a persuasive one: setup, tension, resolution, applied to a single figure so your audience leaves knowing not just what the data says, but what to do about it.