Forecasting demand before you commit budget

Forecasting has a branding problem. The word makes people picture a model that tells you the future, gets it wrong, and embarrasses everyone in the next board meeting. That's not what we use it for. A demand forecast isn't a prophecy — it's a way to make spend follow evidence instead of opinion, and to know, before the money leaves the account, where it's most likely to come back.

The honest framing is this: we are not trying to be right about next quarter. We are trying to be less wrong than a flat assumption, and to attach a confidence band to every number so the team can size its bets accordingly. A forecast that says "demand for this segment is rising, and here's how sure we are" is worth far more than one that confidently names a single figure it can't defend.

Forecasting is a decision tool, not a prediction stunt

The test we apply to any model is simple: does it change what we'd do? If the forecast arrives, everyone nods, and the plan stays identical, the model was theatre. A useful forecast shifts a real decision — it moves budget from a saturating channel to a rising one, pauses a launch until demand firms up, or greenlights inventory ahead of a seasonal swing. We build models backward from the decision they're meant to inform, never forward from the data we happen to have.

That discipline keeps us from over-engineering. A two-week-ahead channel forecast that's directionally right and updated daily beats a twelve-month projection nobody trusts enough to act on. Resolution should match the decision's horizon, not the model's ambition.

How we score demand and audiences

In practice the work splits into two questions: how much demand is coming, and who is most likely to convert. We answer them with a layered approach rather than one heroic model:

• Baseline demand — seasonality, trend and known events, so we're never surprised by a swing the calendar already predicted.
• Leading signals — search interest, on-site behaviour and pipeline velocity that move before revenue does and give us early warning.
• Audience scoring — a propensity model that ranks segments by likelihood to convert, so spend concentrates where the return is most probable.
• Confidence bands — an explicit range on every estimate, because a forecast without uncertainty is just a guess wearing a suit.

The scores aren't precious. They feed a weekly reallocation: budget tilts toward the segments and channels the model rates highly, we watch what actually converts, and the next week's scores absorb that feedback. The model earns trust by being corrected in public, not by being defended.

Audience scoring is where the leverage usually hides. Two segments can look identical in a spreadsheet — same size, same average order value — while converting at wildly different rates once you account for recency, intent signals and how they first arrived. A propensity model surfaces that gap, and the gap is often large enough that simply spending more on the higher-scored segment outperforms any creative tweak you could make to the campaign itself. We've watched a single reallocation, driven by nothing more than honest scoring, move cost per opportunity by double digits inside a month.

A forecast you can't act on is decoration. We judge every model by one question — would the plan change if the number changed? If not, we don't ship it.

— Joana Reis, Head of Analytics

Where forecasts go wrong

The failures are predictable, which is the good news. Models drift when the world shifts and nobody retrains them; they overfit when someone chases last quarter's noise; they get trusted past their evidence when a confident chart outruns the data behind it. We guard against all three the same way — by holding out recent data to check honesty, by retraining on a schedule rather than on panic, and by reporting error openly so a degrading model gets caught before it misroutes a budget. The point of measuring error isn't self-flagellation; it's knowing exactly how much weight a forecast can bear before you lean on it.

The takeaway

Forecasting demand before you commit budget isn't about predicting the future with precision you don't have. It's about replacing "this feels like the right channel" with "demand is rising here, our confidence is moderate, and we'll know within a week if we're wrong." That's a smaller claim than the word forecast implies — and a far more useful one. Make spend follow evidence, attach honesty to every number, and let the model improve every time reality corrects it. That loop, not any single prediction, is what compounds.