Demand forecasting is not about guessing next month's revenue. It is about how much stock to order, when to launch a promotion and at what point to start clearing inventory. An error in one direction means stockouts at peak season; an error in the other means cash frozen in a warehouse. That is why predictive analytics pays back faster in e-commerce than most other AI projects: the result is measured in working capital, not in clicks.
What is actually being forecast
"Demand forecasting" is an umbrella covering at least four distinct tasks with different difficulty and different value.
| Task | Horizon | Decision it supports |
|---|---|---|
| Sales per SKU | 2–12 weeks | Purchasing, replenishment |
| Sales per category | 1–6 months | Budgeting, assortment planning |
| Seasonal peak | a season | When to start the advertising campaign |
| Overstock risk | 4–8 weeks | When to discount and by how much |
Start not with the most accurate model but with the task where an error costs most. In most online stores that is purchasing for the top 20 SKUs that generate the bulk of revenue.
What data you need
A demand model is mostly a data problem, not an algorithm problem. The minimum set looks like this:
- Daily sales history covering at least two years. One year does not let the model see seasonality.
- Stock availability history. Without it the model reads out-of-stock days as falling demand and underestimates next season.
- Price and discount history. Otherwise a promotional spike gets interpreted as organic growth.
- Daily advertising spend. Lets you separate bought demand from organic demand.
- An events calendar: holidays, weekends, new product launch dates, shipping changes.
The first two are critical. Stores that do not record stock history usually spend the first month of the project reconstructing it from indirect data, and forecast accuracy suffers for it.
Why a simple model often beats a complex one
The pattern is consistent: for products with steady demand, classical statistical methods produce results close to complex models and are an order of magnitude easier to maintain. Complex models win where many external factors act on demand at once — seasonality plus promotions plus competitor pricing plus weather.
The sensible sequence: build a baseline forecast with a simple method, measure its error, and only then decide whether a heavier model is justified. Without a baseline you cannot prove the complex model improved anything at all.
New products with no sales history are a separate category. Their forecasts are built from comparable items in the catalogue, and accuracy is always lower. Assigning a new SKU the same confidence as a product with two years of history is a standard planning mistake.
How to measure forecast quality
The real question is not "how accurate is the model" but "how much better is it than how we planned before". Compare against a naive forecast — for example, "we will sell what we sold last month". If the model does not beat that consistently, the implementation is not worth keeping.
Second: errors in opposite directions cost different amounts. Underestimating demand on a fast mover means lost sales and lost search position. Overestimating means frozen cash, but the stock is still there. The model should be tuned to that asymmetry rather than simply minimising average error.
Three traps that stop forecasts from working
The forecast exists but nobody uses it. The most common failure mode: the model is built, the dashboard is live, and the buyer keeps ordering by instinct. A forecast has to be embedded in the decision — in the order form, in the weekly meeting, in the replenishment rule — or it stays a report.
The model was trained on an anomalous period. If the history contains months with a sharp spike or collapse caused by one-off circumstances, the model will treat them as a pattern. Those periods need to be flagged explicitly, not quietly averaged in.
The forecast is never revisited. Demand shifts, competitors run promotions, the assortment changes. A model that is not retrained drifts away from reality — and this is usually noticed after the warehouse is already full of the wrong product. Reviewing accuracy quarterly is the minimum discipline.
Where to start
The realistic first step is not a platform but a pilot on one category with a long sales history and a high cost of error. Three or four months of that pilot produces enough data to judge whether the effect is real and what maintaining it costs.
A hard prerequisite is correctly configured analytics. If sales and traffic data do not reconcile, there is nothing to forecast from. The basics are covered in Google Analytics 4 setup and web analytics, while consolidating data across sources is end-to-end analytics.
A broader view of which decisions can be automated from data is in AI analytics: data-driven decisions and in our AI analytics service description. To discuss a pilot, reach us through contact Top Team.
