Choosing a Color Palette for Data Visualization

The right color palette is determined by your data type, not your taste. Ordered numeric data wants a sequential palette, data with a meaningful midpoint wants diverging, and unordered groups want categorical. Pick the wrong family and the chart will mislead even if the colors are pretty.

The Three Palette Families

• Sequential — light to dark in a single hue. For ordered values (counts, percentages, time).
• Diverging — two hues meeting at a neutral midpoint. For values around a meaningful zero (profit/loss, above/below average).
• Categorical — distinct hues, similar lightness. For unordered groups (countries, product types).

When to Use a Sequential Palette

Use sequential when the values have a natural order and no special midpoint:

• Population density on a choropleth
• Count of orders per region
• Heatmap of correlation magnitudes (0 to 1)

Battle-tested palettes:

• Viridis, Plasma, Inferno, Magma (matplotlib, perceptually uniform, colorblind-safe)
• Blues, Greens, YlOrRd (ColorBrewer)
• Tableau's Blue or Orange sequential ramps

When to Use a Diverging Palette

Use diverging when zero (or a target, or an average) is meaningful:

• Profit vs loss
• Sentiment scores (-1 to +1)
• Year-over-year change
• Correlation matrices that include negatives (-1 to +1)

Battle-tested palettes: RdBu, BrBG, PiYG, RdYlGn (avoid for colorblind audiences — see below).

Always center the color scale on the meaningful midpoint. A diverging palette with a midpoint at the data mean (instead of zero) hides the sign of the values.

When to Use a Categorical Palette

Use categorical for nominal data — groups with no order:

• Product categories
• Countries on a multi-line chart
• Cluster labels from k-means

Battle-tested palettes: Tableau 10, Set2, Dark2, Okabe-Ito (8-color colorblind-safe).

Cap your categorical palette at 7 colors. Beyond that, the eye can't distinguish them and the legend becomes a lookup chore — group small categories into "Other."

Decision Rule in One Sentence

Is your data ordered?
├─ No  → Categorical palette (Tableau 10, Set2)
└─ Yes → Is there a meaningful midpoint?
         ├─ No  → Sequential palette (Viridis, Blues)
         └─ Yes → Diverging palette (RdBu, BrBG)

Accessibility: Skip Red-Green Diverging

Roughly 8% of men and 0.5% of women have red-green color deficiency. RdYlGn and other red-green diverging palettes look identical to those viewers. Safer alternatives:

• Diverging: RdBu (red-blue), BrBG (brown-teal), PuOr (purple-orange)
• Sequential: Viridis family — designed to be both perceptually uniform and colorblind-safe
• Categorical: Okabe-Ito — 8 colors verified for protanopia, deuteranopia, and tritanopia

Test your chart with a simulator like Coblis before publishing.

Code: Picking a Palette in Python

import seaborn as sns

# Sequential
sns.color_palette("viridis", as_cmap=True)
sns.color_palette("Blues", n_colors=5)

# Diverging
sns.color_palette("RdBu_r", as_cmap=True)  # _r reverses

# Categorical
sns.color_palette("tab10")        # Tableau 10
sns.color_palette("colorblind")   # Okabe-Ito-style

Common Pitfalls

• Rainbow palettes (jet): not perceptually uniform — they create false boundaries in the data. Use Viridis instead.
• Categorical for ordered data: using Tableau 10 for "small / medium / large" throws away the order information.
• Diverging on data without a midpoint: if your range is 0–100 with no special middle, the palette implies a meaning that isn't there.
• Too many categorical colors: more than 7 distinct hues is a legend, not a chart.