Demand forecasting, without the crystal ball

What a demand forecast actually does

Demand forecasting estimates the quantity of a product or service customers may request over a future period. In logistics, that estimate helps people make decisions about purchasing, replenishment, warehouse space, labour and transport capacity. It is not the same as deciding how much stock to hold. Forecasting estimates likely demand; inventory planning combines that estimate with lead times, supply risk, service goals and cost to choose a stock policy.

A useful forecast starts with historical orders and sales, but history is only one clue. Teams may add seasonality, promotions, product launches, price changes, known lead times and relevant external signals. A retailer might see a familiar increase around a holiday, while a distributor may account for customer ordering cycles. External information can add context, but only if it is relevant and reliable. More columns do not automatically mean more insight.

Data needs a little housekeeping before modelling. Duplicate orders, missing dates, returns, substitutions and one-off bulk purchases can distort patterns. Teams should define what “demand” means: orders placed, units shipped, or customer requests including unfilled orders. Those measures differ, especially during a stockout. If sales history records only what could be supplied, it can understate what customers wanted.

Horizon, granularity and variability

The forecast horizon is how far into the future an estimate looks. A short horizon may support near-term picking or replenishment; a longer one can inform supplier capacity and seasonal planning. Longer horizons usually carry more uncertainty, so forecasts should be interpreted accordingly rather than treated as precise instructions.

Granularity describes the detail of the forecast: by product, location, customer, day or week. Fine detail can help local planning, yet sparse product-location histories may be noisy. Aggregating related items or periods can reveal steadier patterns. The right level depends on the decision being made. A weekly national forecast cannot directly schedule tomorrow’s warehouse shifts, and an hourly estimate for a slow-moving item may be false precision.

Demand variability matters because two items with the same average can behave very differently. One may sell steadily; another may alternate between quiet periods and sharp spikes. Forecast errors—the difference between an estimate and what actually occurred—should be monitored over time and across meaningful groups. A single overall score can conceal recurring misses for a particular location or product family.

From simple methods to machine learning

Simple approaches can be remarkably useful. A moving average smooths recent observations; seasonal methods account for recurring calendar patterns; a planner’s baseline may be adjusted for a known promotion. These methods are easier to explain and can be a sensible starting point when data is limited or demand is stable.

Machine-learning approaches can model relationships among many signals, identify nonlinear patterns or generate forecasts for a larger assortment. They may be useful when the data is sufficiently complete and the planning problem justifies the added complexity. They also require careful validation, monitoring and maintenance. A complex model can learn noise, react poorly when circumstances change, or be difficult to explain to the people expected to act on it. Complexity is not a quality badge.

Compare approaches using historical periods that were not used to fit the model, and look at errors in the context of real decisions. Consider bias as well as average error: a model that consistently overestimates may create excess stock, while consistent underestimation can contribute to shortages. Review exceptional periods separately so unusual events do not silently define normal behaviour.

Putting predictions to work

Forecasts can inform purchase plans, replenishment triggers, warehouse slotting, seasonal labour and transport capacity. They should be presented with a time horizon, data date and uncertainty where possible. A planner needs to know whether a value is an expected quantity, a range, or a manually adjusted estimate.

Model updates should follow changes in the business and the data. New products have little history; promotions can change demand; suppliers may alter lead times. Human review is valuable for known events that the data does not contain, but adjustments should be recorded so teams can later distinguish model output from a manual override. Forecasts are inputs to operational judgement, not instructions to follow blindly.

Inaccurate forecasts can mean extra inventory, shortages, unnecessary transfers between sites and avoidable warehouse pressure. The response is not always “buy a better model.” Sometimes the definition of demand is inconsistent, product records are poor, or replenishment decisions are too slow to use the available signal.

What to know

A forecast is a conditional estimate, not certainty. The best approach is the simplest one that supports the decision reliably, is understandable to its users, and can be maintained with available data and skills. Forecasting and inventory policy work together, but they answer different questions. Keep people in the review loop, especially when unusual conditions or high-impact decisions are involved.