The situation
An e-commerce business has to decide ahead of time how much stock to hold, how many people to roster and how much to spend on marketing. All of those decisions depend on a guess about how many orders are coming.
The problem
The guess was usually last year's number plus a growth figure that felt about right. That works until a promotion lands badly or a busy season arrives early. Nobody could say how good the guess was, so nobody could say whether a better method was worth the effort.
What we did
We started with the simplest possible forecasts, such as "next week will look like this week" and "next week will look like the same week last year". These are the yardstick. A cleverer method only counts if it beats them.
Then we trained forecasting models on past order history, picking up the weekly rhythm, the seasons and the effect of holidays and promotions. To test them fairly, the most recent stretch of history was hidden, the models predicted it, and we compared the predictions with what really happened.
Finally, the forecast was shown as a range, not a single number. "Around this many, and anywhere in this band is normal" is much more useful to someone ordering stock than a figure that pretends to be exact.
The outcome
Planning conversations began from a forecast the team could check, with a record of how far off it had been in the past. When the forecast was wrong, there was something to learn from instead of an argument about whose guess was better.
What it meant for the business
The forecast did not replace people's judgement about things a model can't see, like a competitor's price cut. It gave that judgement a sturdier starting point and made it easy to see when a plan was relying on something unusual.
Could this work for you?
If your plans depend on a number that lives in one person's spreadsheet, a first forecast is a good small project. We begin with the baseline, measure honestly and add complexity only when it earns its place.
Related services: Predictive modelling and forecasting, Data analysis and insights
This project was led by our founder, Achyutanand Jha, in an earlier role. The client is not named, and figures are those reported at the time.