Demand forecasting isn't magic or one clever algorithm. It's a small toolbox of statistical methods, each good at a different kind of demand. The real skill is matching the method to the product — here's how.
Before any formula, look at a product's sales history and ask: is it steady, trending up or down, seasonal, or intermittent (lots of zero weeks)? The right method depends entirely on which of those you're looking at.
Moving average — the average of the last few periods. Simple and stable; great for flat, steady demand, but it lags when demand changes and ignores seasonality.
Weighted / exponential smoothing — like a moving average, but recent periods count for more. It reacts faster to change while still smoothing out noise. A workhorse for most steady-to-slowly-changing products.
Linear trend — fits a straight line through the history to project growth or decline. Useful for products on a clear trajectory, risky if you extrapolate too far.
Seasonal naive & Holt-Winters — these explicitly model repeating seasonal shapes (the Q4 spike, the summer dip). Holt-Winters combines level, trend, and seasonality, which is why it shines for products with a strong annual rhythm — provided you have two full years of history.
Croston's method — built for intermittent demand, where a product sells in occasional bursts with many zero periods in between. Ordinary averages badly mis-handle these; Croston forecasts the demand rate instead.
You backtest. Hide the last few months of real history, ask each method to "predict" them, and measure how far off it was (using an error metric like RMSE). The method with the lowest error on the held-out data is the one most likely to be right going forward — for that specific product. A method that wins for a steady staple will lose for a seasonal gift item.
Fopsis runs all of these methods (12 in total) against each product's history, backtests them, and automatically picks the lowest-error method for that product — then applies its seasonal shape on top. You see which method won and why, so the forecast is never a black box.
Every method assumes your sales history reflects true demand. It usually doesn't — stockouts hide demand you couldn't fulfil, and returns inflate it. Correcting for those before forecasting matters more than which method you pick. We cover that in why stockouts wreck your forecast and netting returns out of demand.
Connect your store and get an auto-selected, explainable forecast in minutes.
Start Free Trial