Demand Forecasting: A Practical How-To Guide
Every operational decision downstream of the customer depends on a forecast: how much stock to hold, how much capacity to book, when to issue purchase orders, and how much buffer to build in. Bad forecasting is expensive in both directions — too high and you carry surplus inventory and tie up cash; too low and you miss sales, disappoint customers, and pay premium prices for emergency supply.
This guide covers the main forecasting approaches, how to measure whether your forecast is working, and the avoidable mistakes that undermine even sophisticated models.
Qualitative vs. Quantitative Methods
Quantitative methods use historical data to project future demand. They work well when the pattern is stable and history is a reliable guide. Qualitative methods bring in human judgement — useful when you are launching a new product, entering a new market, or when external factors (legislation, competitor moves, a new retail listing) mean the past is a poor predictor of the future. In practice, good forecasting combines both.
The three most common quantitative techniques are:
Moving average. Average the last N periods of demand. Simple and transparent, but it weights old data equally with recent data and lags behind trend shifts. Best for stable, low-variation demand.
Exponential smoothing. A weighted average that gives more weight to recent periods via a smoothing factor (α). The higher α is, the more responsive the forecast — but also the more it chases noise. Variants like Holt-Winters also capture trend and seasonality.
Causal models. Regression-based approaches that tie demand to external drivers — economic indicators, weather, promotional spend, channel activity. More powerful when the driver relationship is stable, but they require more data and more maintenance.
Measuring Forecast Accuracy
Two metrics matter most:
MAPE (Mean Absolute Percentage Error). The average of the absolute percentage errors across a set of periods. Easy to explain to management, but it breaks down when actual demand is near zero and can penalise over-forecasting asymmetrically.
Bias. Whether the forecast consistently over- or under-shoots actual demand. A forecast with low MAPE but persistent positive bias (always too high) will systematically inflate inventory. Track bias separately — it reveals systematic error that MAPE can hide.
Review forecast accuracy at the right level of aggregation. A forecast that is accurate at the product-family level can mask large errors at the SKU-location level, which is where fulfilment decisions actually happen.
Common Mistakes and How to Avoid Them
Over-relying on history: Historical patterns reflect past conditions. A promotional event, a supply disruption, or a channel shift can make last year's data actively misleading. Always sense-check statistical outputs against current market intelligence.
Ignoring market intelligence: Sales teams and account managers carry information that the data does not — upcoming promotions, customer expansion plans, competitor stock-outs. Build a structured process to capture and weight this intelligence.
Forecasting at the wrong level: Aggregate forecasts are more accurate but less useful. Disaggregate to the level at which you make decisions — by SKU, by location, by channel.
Treating the forecast as a commitment: The forecast is a planning input, not a contract. Build safety stock and lead-time buffers to absorb forecast error rather than expecting the number to be exactly right.
No feedback loop: If no one reviews forecast accuracy regularly, the model drifts and errors compound. Set a monthly cadence to review MAPE and bias by product category and to adjust parameters.
Building a reliable forecasting process is not primarily a technology problem. It is a discipline problem — consistent data hygiene, regular review, and a clear decision about who owns the number.
XNM Consulting helps organisations design and implement supply chain processes that are resilient and responsive. Learn more about our procurement and supply chain services.