Forecast Projection

How it computes

Holt's linear-trend exponential smoothing — ETS(A,A,N) — with model-derived confidence cones.

se = σ·√(1 + Σⱼ₌₁ʰ⁻¹(α + jβ)²);  band half-width = z(confidence) · se

Reference: Hyndman & Athanasopoulos, Forecasting: Principles and Practice

When to reach for it

Use the Revenue Forecast when someone needs a number and its uncertainty — a board target, a hiring plan, a runway check. It draws the Holt-linear projection as a dashed median inside 50% / 80% confidence cones, so the range is visible instead of implied by a single confident-looking line.

How it thinks

From dated history it fits Holt's linear-trend exponential smoothing (holtLinear, level α + trend β) and extends the final level + trend forward. Around that median it draws a cone whose half-width is z(confidence) × se, where:

• se is the ETS(A,A,N) forecast standard error σ·√(1 + Σⱼ₌₁ʰ⁻¹(α+jβ)²) (Hyndman & Athanasopoulos, Forecasting: Principles and Practice). The residual σ comes from the in-sample one-step errors; the 1 + Σ(α+jβ)² term is how a trend model's uncertainty actually compounds — it widens faster than a naive σ·√h, so a 12-month cone is honestly wide rather than falsely tight.
• z(confidence) is the real two-sided normal quantile (zForConfidence, an Acklam probit), so 95% → 1.96. No fudge factors, no stretched bands.

The bands are Gaussian and symmetric by default — fine for most financial series, but a strongly skewed or bounded metric is only being approximated. For those, flip intervalMethod to "conformal": instead of z × σ, the cone is sized from the empirical quantile of the model's own out-of-sample one-step errors — split / inductive conformal prediction (Vovk, Gammerman & Shafer, Algorithmic Learning in a Random World, 2005; Angelopoulos & Bates, A Gentle Introduction to Conformal Prediction, 2021). That drops the normality assumption entirely — the width comes from how wrong the model has actually been on held-out data, not a bell curve — and the "signed" mode yields an asymmetric band that captures skew the Gaussian one can't represent. The horizon-widening still follows the ETS variance law, so conformal changes the cone's level, not its shape; and when there's too little history to certify the level (95% needs ≥19 held-out errors) it falls back to the Gaussian band rather than printing an over-confident one. Its own honest caveat: the conformal guarantee assumes residual exchangeability, which a trending series only approximately meets — the rolling-origin design keeps the scores out-of-sample to soften that. The accessible table mirrors the median and the band edges.

Prior-calibrated — borrow strength from your cohort

When the history is short, the model's own error estimate is noisy — exactly when a confident-looking band is most dangerous. Pass a prior (a SegmentPrior from an installable Benchmark Pack, sextant-insights/benchmark-packs.ts) and the cone is prior-calibrated: the band half-width becomes a precision-weighted blend of the model's own empirical error and the cohort's typical error, q = w·qₚ + (1 − w)·qₑ with w = priorPrecision / (priorPrecision + n). So a brand-new series leans on its segment's prior, and the prior fades like ~1/n as your own history accrues (priorShrinkConformal). It moves only the band — the point forecast stays your own Holt path — and the horizon-widening still follows the ETS law, so the prior resizes the cone's level, never its shape. The honest check is built in: rollingCoveragePrior() reports coverage against the prior band too, flagging cohort divergence when the borrowed prior doesn't actually fit your series. (The built-in packs ship illustrative defaults; install your own measured pack via loadBenchmarkPack.)

How honest is the band?

A band built from in-sample σ can read tighter than reality. rollingCoverage() is the check: a one-step rolling-origin backtest that re-fits at each historical origin and reports how often the next actual landed inside each band. The footnote surfaces it — "80% band — backtested 78% coverage" — so an evaluator can see whether the interval is calibrated instead of taking it on faith. Persistent under-coverage there is the signal to widen σ, revisit the model, or switch to intervalMethod: "conformal" — which sizes the band from those same held-out errors directly, so it can't be over-confident in the way an in-sample σ can.

Good to know

• alpha / beta, the horizon, the confidence level, and the interval method (intervalMethod: "gaussian" | "conformal") are all live knobs; the headless <ChartForecastProjectionPlot> takes the same props and derives the same series via buildForecastSeries, so you can drop the plot into your own layout.
• Below two data points the widget renders an explicit empty state — never a flat-zero forecast or a silently-missing narrative.

Use it with your data

import { ChartForecastProjection } from "@/components/blocks/chart-forecast-projection/chart-forecast-projection";

<ChartForecastProjection history={myData} />

const { data, isLoading, error } = useSWR("/api/metric", fetcher);
// route isLoading/error through <ChartStates>, then:
<ChartForecastProjection history={data} />

const history = await fetch(url, {
  next: { revalidate: 3600 },
}).then((r) => r.json());

<ChartForecastProjection history={history} />

import { mapRows } from "@/lib/sextant/column-map";

const out = mapRows(rows, { /* contractKey: "yourColumn" */ }, dataSchema);
if (out.ok) <ChartForecastProjection history={out.data} />;

Make it yours

Every customization below is an optional prop — omit them all and you get the defaults shown above. Recolor, reformat, restyle, or drop the card chrome without forking the component.

<ChartForecastProjection
  series={{ actual: { color: "var(--chart-3)", label: "…" } }}
/>

const eur = new Intl.NumberFormat("de-DE", {
  style: "currency",
  currency: "EUR",
  notation: "compact",
});

<ChartForecastProjection valueFormatter={(n) => eur.format(n)} />

<ChartForecastProjection narrative="Your own one-liner." />

// …or hide it entirely:
<ChartForecastProjection narrative={false} />

<ChartForecastProjection
  className="max-w-xl"
  classNames={{ body: "bg-muted/20" }}
/>

import { ChartForecastProjectionPlot } from "@/components/blocks/chart-forecast-projection/chart-forecast-projection";

// Just the chart body — bring your own card/layout:
<ChartForecastProjectionPlot history={history} config={config} />