Choosing the right chart — and the right color

Most dashboards don't fail because the numbers are wrong. They fail because everything is a bar chart, the colors fight each other, and by the fifth panel your eye has stopped reading. A great dashboard is a designed object: the chart type answers a specific question, the color carries meaning, and the mix of forms keeps the page legible. Here's the field guide we follow when building the PrimaForja demo dashboards on the distributor dataset.

1. Match the chart to the question, not to habit

Every panel answers a question, and the question dictates the form. A quick mapping, with the example from our own data:

• Comparison ("who is biggest?") → a sorted bar. Revenue by customer segment, top-10 customers, cost per order by DC.
• Trend over time ("which way is it going?") → a line. Monthly revenue and gross margin, fill-rate trend, DSO over the quarter.
• Part-to-whole that changes → a stacked area or stacked bar, not a series of pie charts. Revenue mix by segment across months.
• A single part-to-whole snapshot → one donut, sparingly, with ≤5 slices. More than that and a sorted bar always reads better.
• Ranked contribution ("what drives the 80%?") → a Pareto. ABC inventory classification, concentration risk.
• Step-by-step build-up → a waterfall. The pocket-margin bridge from list price → invoice → pocket → contribution → EBITDA.
• Distribution ("how spread out?") → a histogram or box plot. Order sizes, lead-time variability, rep quota attainment.
• Correlation ("do these move together?") → a scatter / bubble. Vendor defect rate vs on-time delivery, sized by spend.
• Geography → a choropleth. Revenue by US state.
• Flow between stages → a Sankey. The order pipeline from ordered → picked → shipped → delivered.

If you can't say what question a panel answers in one sentence, it doesn't belong on the dashboard — or it's the wrong chart for the question it's trying to answer.

2. Vary the form on purpose

A page of ten bar charts is exhausting even when every bar is correct, because the reader has to re-learn what they're looking at each time the axes look identical. Deliberately alternate forms: pair a trend line with a ranked bar, a KPI tile with a sparkline, a map with a scatter. The variety isn't decoration — it signals "this is a different kind of question" before the viewer has read a single label. On our executive tab, no two adjacent panels share a chart type for exactly this reason.

The counter-rule matters too: when two panels answer the same kind of question, keep them identical so they're instantly comparable. Consistency where the questions match, variety where they differ.

3. Make color mean something

Color is the most abused channel in analytics. Three palettes, three jobs:

• Categorical — for unordered groups (segments, regions, carriers). Pick a set of distinct hues, cap it at ~6–8, and reuse the same hue for the same category everywhere. If Enterprise is teal on one panel, it's teal on all of them.
• Sequential — for an ordered quantity (revenue on a map, utilization in a heatmap). A single hue ramped light→dark. Never a rainbow: rainbow ramps invent boundaries that aren't in the data.
• Diverging — for a value with a meaningful midpoint (budget variance, forecast bias, margin above/below target). Two hues meeting at a neutral center.

Beyond palette choice, three rules save most dashboards:

• Encode, don't decorate. If two bars are the same category, they should be the same color. Rainbow bars where color carries no information just add noise.
• Reserve a semantic pair. Keep one color for "good/revenue/on-target" and one for "bad/cost/risk," and never use those two for anything else. Red should mean attention, not "the third category."
• Design for color-blindness and grayscale. ~8% of men can't separate red from green. Use a color-blind-safe palette, and lean on position, ordering, and labels so the chart still reads if the color is stripped out entirely.

4. Let the data breathe

Remove chartjunk: heavy gridlines, 3-D effects, redundant legends, and axis labels that repeat the title. Direct-label the important series instead of forcing a legend lookup. Give the ink to the data, and keep a light, consistent theme — including a proper dark mode, since a lot of analysts live in one.

The mix, in one sentence

A dashboard that works alternates chart types so each question announces itself, uses a small, consistent, meaningful color system, and spends its ink on data rather than decoration. Everything on our demo dashboards is built to that standard — and every metric behind them is defined in the methodology reference so the picture and the number always agree.