Curve Smoothing
The lines between your data points are a design choice.
A line chart is mostly not data. Fifteen measurements produce fourteen connecting segments, and every pixel of those segments is invented - the only honest parts are the points. Since you're inventing them anyway, the question is what the invention should communicate.
Same fifteen points. The curve still passes through every one of them - smoothing only replaces the straight segments between measurements, which were never data to begin with.
Choosing a curve
- Straight segments say "discrete samples": precise, jagged, a little technical. Right for sparse data where each reading matters (deploys per day, invoice totals).
- Catmull-Rom / cardinal curves pass through every point with smooth tangents - the default "designed" look, and what the demo above draws. The risk: between points the curve can overshoot above or below any actual measurement.
- Monotone cubic (d3's
curveMonotoneX) smooths without overshoot - it never invents a peak that isn't in the data. The safe choice for dashboards where someone might read values off the line. - Shu Ding's adaptive fitting goes further for dense, noisy series: fit low-degree polynomials over sliding windows and blend them, tuning smoothness by error tolerance rather than by eye.
The rule
Smoothing is presentation, so it obeys presentation's rule: it may not lie. Show the true points when they're sparse, keep tooltips reading raw values, and if a smoothed line would cross a threshold the data never crossed, use monotone. The curve should feel calm - a chart used for decisions shouldn't be interesting between measurements.