Why response curves go stale

A response curve is the best planning tool we have. It's also fit entirely to the past — and the past can lie. The honest version of model-driven planning isn't a more confident forecast. It's a faster admission of when the forecast has expired.

Every model is built from yesterday

As Rory Sutherland puts it in Alchemy, all big data comes from the same place: the past. A curve fit to last year's campaigns assumes next year's world rhymes with it. Usually it does — until a competitor, a price move, or a cultural shift rewrites how people respond. Instant coffee sat on a flat, commoditised price curve for decades; then Nespresso re-framed coffee as a premium ritual and the old price-response history was worth nothing. No amount of historical precision would have seen it coming.

All big data comes from the same place: the past. — Rory Sutherland, Alchemy (2019)

A curve is a prior, not a prophecy

This is why Curvio never reports a curve as a single deterministic line. Every response curve carries its uncertainty — directional predictions with confidence intervals, calibrated on your own campaigns, and an explicit line on what the model can't yet see. A curve is a well-evidenced prior you plan against and keep testing, not a prophecy you obey. Sutherland's warning is worth keeping on the wall: a business incapable of action except on the basis of numerical information looks clever, but is at permanent risk of acting dumb.

The answer is cadence, not certainty

You don't beat staleness with a cleverer one-time forecast. You beat it by re-fitting often. Curvio re-calibrates every quarter as new results land, so the curves track the market you're in now — not the one you measured a year ago. When a context shift shows up in the data, the model moves with it; when it hasn't shown up yet, the confidence intervals widen and the plan stays humble. Honesty about time is part of the product, the same way honesty about attribution is.

Why this makes the model worth trusting

It sounds like a caveat. It's actually the point. A planning model you can trust is one that tells you when to stop trusting it — that re-calibrates on a clock, states its uncertainty, and flags the moments a curve might be lying. Precision you can't defend fails silently, after the budget is committed. A calibrated prior that's re-fit every quarter fails loudly, early, and cheaply — which is the only kind of failure worth having.