Fan Chart

Trends Forecast

"What will revenue be next quarter?" The honest answer is never a single number - it is a range, and the range is the whole point. Hand an exec one number and you are guessing; hand them this fan and you are telling the truth about the risk. The solid line is what already happened, the dashed line is the single most-likely path, and the shaded bands show how sure the forecast is - widening as they reach into the future, because the further ahead you look, the less anyone can know.

Example

canvas · responsive

How to read it

• Solid line - history. The actuals you already have.
• Dashed line - the most-likely path (the forecast median): one best guess, never the whole story.
• Nested bands - confidence. Inner to outer = 50% / 80% / 95%. The real value should land inside the 95% band about 19 times out of 20. You plan against the band, not the line.
• Why it fans out. Next month is fairly knowable; a year out is not. Uncertainty compounds with distance, so the bands widen.

Read your worst case off the bottom of the outer band and your best case off the top. The fan is your base / upside / downside in a single picture - no separate scenario tab needed.

The maths, in plain terms

You do not need the equations to use the chart, but here is what is under the hood - and why you can trust it:

• The median comes from Holt-Winters exponential smoothing. It tracks two moving quantities, the current level and the trend (slope), and rolls them forward; if the series has a repeating season, it tracks that too. (Prefer a straight line? method: "linear" fits an ordinary least-squares regression instead.)

  ℓₜ = α·yₜ + (1−α)(ℓₜ₋₁ + bₜ₋₁) · bₜ = β(ℓₜ − ℓₜ₋₁) + (1−β)bₜ₋₁ · ŷₜ₊ₕ = ℓₜ + h·bₜ

• The bands come from the model's own past errors. It measures how far its fitted values missed (the residual spread σ) and widens the interval as ŷ ± z·σ·√h - z = 1.96 for 95%, and the √h is exactly why the fan opens with the horizon h.
• Should you trust it? A backtest hides the last few real points, re-forecasts them, and reports the error (MAPE, RMSE). You get an honesty score before you bet on the number, not after.

All of it runs in the browser - no data-science backend, no server round-trip. (Power BI, by contrast, only forecasts on a line chart and stops where real modelling begins.)

Build the data in one call with forecastFan() from @michi-vz/insights/forecast, or hand it series (history + certainty:false median) and nested bands.

Usage

Insights (one call)

import { mountFanChart } from "@michi-vz/core";
import { forecastFan } from "@michi-vz/insights/forecast";

// history = DataPoint[] of actuals; build the fan (median + 50/80/95% bands)
const item = forecastFan(history, { method: "holt-winters", horizon: 4, levels: [0.5, 0.8], level: 0.95 }, "Revenue");
const chart = mountFanChart(el, { dataSet: [item], xAxisDataType: "date_annual" });

Vanilla JS

import { mountFanChart } from "@michi-vz/core";

const chart = mountFanChart(el, props); // props.dataSet = FanDataItem[]
chart.update(next);
chart.getContext(); // renderer-agnostic, LLM-ready
chart.destroy();

Web component

<script type="module" src="https://cdn.jsdelivr.net/npm/@michi-vz/wc"></script>

<michi-vz-fan-chart id="c"></michi-vz-fan-chart>
<script>
  Object.assign(document.getElementById("c"), props); // dataSet (series + bands), title, …
</script>

Data shape

A FanDataItem is a familiar line series plus nested bands:

interface FanDataItem {
  label: string;
  color?: string;
  series: DataPoint[];   // history (certainty:true) then forecast median (certainty:false → dashed)
  bands: { level: number; series: RangeDataPoint[] }[]; // drawn widest-first, graduated opacity
}

API

Props are typed as FanChartProps in @michi-vz/core and mirror LineChartProps (width, height, margin, colors / colorsMapping, renderer, highlightItems, disabledItems, fillOpacity, and the on* callbacks). onChartDataProcessed / getContext() return the renderer-agnostic ChartContext.