Where this comes from
In a demand review, forecast error turns into an argument fast. Planning points at the sales override. Sales points at the market. The month closes with no decision, and the next month the same conversation comes back.
What I saw work was showing up with the error already broken down in numbers: where the forecast misses, in which direction and whether the manual override helped. With that on screen, nobody has to defend an opinion. It takes three cuts, always in the same order.
1. Error by product family
This is the first cut I make. A 40% total error does not tell you where to act. Split by family, a Pareto almost always shows up: a few families explain most of the error.
Each family's weight in the error is its absolute error divided by the total absolute error:
Weight in error = sum of |actual - forecast| for the family ÷ total sum of |actual - forecast|
| Family (fictional) | Weight in error | Cumulative |
|---|---|---|
| A | 30% | 30% |
| B | 24% | 54% |
| C | 16% | 70% |
| D to H | 30% | 100% |
In the example, 3 of 8 families explain 70% of the error. The meeting starts with them.
For the total error, I prefer the weighted error (WMAPE: sum of absolute errors divided by the sum of actuals) over a simple average of percentages. In a simple average, an item that sells 2 units and misses by 1 weighs as much as the best seller.
2. Did the manual override help? (FVA)
Every demand process I have worked with had two versions of the number: the statistical forecast from the model, and the adjusted forecast after sales or planning overrides. FVA (Forecast Value Add) shows whether that override made the result better or worse:
FVA = statistical forecast error - adjusted forecast error
| Version (fictional) | Error (WMAPE) |
|---|---|
| Statistical forecast | 25% |
| Forecast adjusted by sales | 32% |
| FVA of the override | -7 points |
A negative FVA means the override made the forecast worse. Then the conversation changes: the model was better, and what needs review is the override process.
I learned to take three precautions so the number is fair to whoever makes the override:
- Calculate it by family and by who made the override. An override can help one family and hurt another.
- Look at least at three months in a row. One month on its own can be an exception.
- Always compare on the same horizon. I tracked Lag 1 and Lag 3: a forecast made one month ahead and one made three months ahead have different errors by nature.
3. Which way the error goes (bias)
Absolute error shows how big the error is. Bias shows its direction. When the error always goes the same way, it is no longer noise.
Bias = sum of (forecast - actual) ÷ sum of actual
With positive bias, the forecast sits above actuals: stock piles up, coverage rises and so does the risk of expiry, which for pharmaceutical products turns into provisions and destruction. With negative bias, the forecast sits below actuals and the result is stockouts.
In the video example, the adjusted forecast was above actuals in 12 out of 12 months, with a bias of +15%. Twelve months in a row is not bad luck. It is an optimistic override repeated every month, and it shows up straight in inventory.
4. From analysis to action
With the three cuts on the table, the meeting agenda changes:
- Which families explain most of the error?
- Where is the manual override making the number worse?
- Does the error always go the same way?
- Who does what before the next meeting?
Each critical family leaves with an action, an owner and a deadline. Overrides with negative FVA are reviewed or stopped. And the same numbers come back at the next meeting to show whether things improved.
What I take from it
Measure before arguing, and always measure the same way: same metric, same horizon, same cut. That idea is also behind the forecast qualification model I built in Power BI and DAX. It classifies each item's demand pattern (Syntetos-Boylan) and recommends, using 8 criteria, when the statistical forecast can go ahead without manual review. That way the team spends its time where the risk is higher.
Month-end checklist
I put the four steps, the formulas and the meeting questions on a single A4 page (in Portuguese).
Download the checklist (PDF, 1 page)
All numbers in this article and in the video are fictional and only serve as an example. No company data was used.