Demand forecasting: where the accuracy actually comes from

When a distributor asks us to improve its demand forecasts, the conversation often starts with algorithms. Should we use machine learning? Which model is best? These are reasonable questions, but they are rarely where the biggest gains are.

The usual starting point

Most mid-sized distributors forecast in spreadsheets, often with moving averages and a planner's judgement on top. This works surprisingly well for stable products. It breaks down on everything else — and the cost shows up twice: stockouts on fast movers and excess stock on slow ones, often at the same time.

Gain one: put promotions into the forecast

In a recent project for an FMCG distributor managing around 4,500 SKUs, the single biggest improvement came from adding the promotion calendar to the forecast. Swings that buyers had treated as unpredictable were, in large part, explained by promotions — the client's own and the retailers'. Holidays, seasonality and weather added further accuracy, but promotions mattered most.

If your forecast does not know when a product is on promotion, it is not forecasting demand. It is averaging it.

Gain two: stop forecasting every product the same way

Products behave differently. A simple segmentation changes the approach:

• High-volume, stable products can be forecast precisely with relatively simple models.
• High-volume, volatile products need drivers — promotions, seasonality, events.
• Low-volume, irregular products are better managed with safety stock rules than with point forecasts.

We test several forecasting approaches per product group and keep the one that performs best on recent history — not the one that sounds most advanced.

Gain three: measure the right thing

A single accuracy figure across all SKUs hides more than it shows. We track:

• Forecast error (MAPE) by product group, weighted by volume or value.
• Bias — whether the forecast is systematically too high or too low. A small error with a consistent bias still produces stockouts or overstock.
• Business outcomes: stockout rate on top sellers and excess inventory value.

In the distributor's case, forecast error fell from 34% to 21%. More importantly, stockouts on top-selling products fell by 28% and excess inventory by 15%.

Gain four: put the forecast where decisions are made

A forecast that lives in a data scientist's notebook changes nothing. The buyers need it in their daily workflow: recommended order quantities, clear exceptions to review, and the ability to see why the forecast says what it says. Trust is built when a buyer can check a number and understand it.

Common pitfalls

• Training models on history that includes stockouts — the data shows what was sold, not what was demanded.
• Ignoring supplier lead times, so a correct forecast still arrives too late.
• A one-off model with no automated data refresh. Forecasts decay quickly without fresh data.