The rolling forecast that explains itself
A forecast nobody trusts is a spreadsheet with opinions. A rolling forecast that shows its drivers, its accuracy and why it changed gets used. Here is how to build one.
Every finance team forecasts. Few forecasts get used. The reason is rarely accuracy alone; it is that nobody can say why the number is what it is, or why it moved since last month.
A forecast that explains itself answers three questions without a meeting: what drives the number, how good the number has been, and what changed.
Roll, don't restart
An annual budget ages from the day it is approved. A rolling forecast always looks the same distance ahead, usually 12 to 18 months, and is refreshed every month with the latest actuals.
- Less effort: each cycle updates a model instead of rebuilding a plan.
- Better timing: decisions are made on a view that is weeks old, not months.
- Honest targets: the budget stays the commitment; the forecast stays the truth.
Forecast drivers, not lines
Forecasting every account line is slow and noisy. Forecast the few drivers that move the business and let the lines follow.
| Business | Typical drivers |
|---|---|
| Manufacturing | Volumes, price per unit, input costs, capacity |
| Distribution | Orders, mix, rebates, freight per unit |
| Services | Headcount, utilisation, billing rates |
| Subscriptions | New customers, churn, price changes |
Show the accuracy
Publish how far each past forecast was from the actual, by driver and by horizon. People trust a forecast that admits its error more than one that hides it, and the error tells you where to improve the model.
Explain the change
Every month, the first question is why the forecast moved. Break the change into its causes:
- Actuals: the month closed differently than forecast.
- Drivers: volumes, prices or costs were updated.
- Assumptions: someone changed a judgement, and recorded why.
A bridge from last month’s forecast to this month’s, with these three bars, ends most debates before they start.
Where machine learning helps, and where it doesn't
Statistical and machine-learning models are good at patterns in history: seasonality, trends and the effect of known events. They are poor at things that have never happened, such as a new product or a lost contract. The best forecasts combine both: the model proposes, the business adjusts, and every adjustment is visible.
Vittora runs rolling forecasts this way, with each number traceable to its drivers. Our forecasting and FP&A team sets up the drivers with you and runs the first cycles alongside your team.