Every inventory decision — how much to order, how much buffer to hold, when to reorder — starts from a guess about future demand. Forecasting is the discipline of making that guess systematically instead of by gut feel, and the good news is that you don’t need advanced statistics to get meaningfully better than “whatever we sold last month.” Three methods cover the vast majority of real-world forecasting needs, and picking the right one is mostly about matching the method to how your demand actually behaves.
The Three Methods, in Plain Terms
Moving Average takes the average of your last N periods and uses that as next period’s forecast. If you sold 100, 110, and 105 units in the last three weeks, a 3-period moving average forecasts 105 for next week. It’s simple, transparent, and works well when demand is fairly stable with no strong trend or seasonality. Its weakness is that it treats old and recent data equally — a demand shift three weeks ago pulls the forecast just as hard as a shift last week.
Weighted Moving Average fixes that by letting you assign more weight to recent periods. If this week matters more than three weeks ago, you can weight it 50%, last week 30%, and two weeks ago 20% — the forecast reacts faster to recent changes while still smoothing out single-period noise. The tradeoff is that you have to choose the weights yourself, and a bad set of weights can make the forecast worse than a plain moving average.
Exponential Smoothing takes the “recent matters more” idea further with a single tuning parameter, alpha, that controls how much weight the newest data point gets versus everything that came before. A high alpha (close to 1) makes the forecast react quickly to recent changes — good for volatile, fast-moving demand, but noisy. A low alpha (close to 0) smooths out noise but reacts slowly to real shifts. Most practitioners start around alpha = 0.2-0.3 and adjust based on how the forecast performs.
How to Pick One
The honest answer is: try more than one and measure which performs better on your actual data. But as a starting point:
- Stable, low-noise demand (a mature product with steady sales) → simple Moving Average is often good enough, and it’s the easiest to explain to a stakeholder who wants to understand the number.
- Demand with a mild trend or recent shift (growing or declining sales, a promotion effect fading out) → Weighted Moving Average or Exponential Smoothing, since both let recent data dominate.
- Volatile, fast-changing demand → Exponential Smoothing with a higher alpha, since it adapts fastest — though nothing here handles strong seasonality well on its own; seasonal patterns need a seasonally-adjusted model, which is a step beyond these three basic methods.
Measuring Whether a Forecast Is Any Good
Picking a method isn’t the end of the job — you need to know whether it’s actually working. Two metrics do this: Mean Absolute Deviation (MAD) and Mean Absolute Percentage Error (MAPE). Both measure the same underlying thing — how far off your forecasts have been from actual demand — but express it differently.
MAD gives you the average forecast error in the same units as your demand (e.g., “off by 12 units on average”). It’s intuitive but hard to compare across products with very different volumes — being off by 12 units matters a lot for a product that sells 20 units a week and barely at all for one that sells 2,000.
MAPE expresses the same error as a percentage of actual demand, which makes it comparable across products regardless of scale — “off by 8% on average” means the same thing whether the product sells 20 units or 2,000. Its main limitation is that it behaves badly when actual demand is close to zero (dividing by a tiny number inflates the percentage), so it’s less useful for intermittent or very low-volume items.
A reasonable workflow: forecast with your chosen method, wait for actual results, then run both MAD and MAPE against the forecast. If the error is consistently large or trending worse, either the method is wrong for this product’s demand pattern, or the underlying demand has shifted enough that your parameters (like alpha) need to be re-tuned.
A Worked Comparison
Suppose actual weekly demand for a product over the last 6 weeks was: 100, 95, 120, 110, 130, 125. A simple 3-period moving average forecasting week 7 would average the last three (110, 130, 125) to get 121.7. An exponential smoothing forecast with alpha = 0.3, starting from an initial forecast of 100 and updating each week, would land somewhere in the low-to-mid 110s by week 7 — reacting to the upward trend, but more slowly than the moving average because each new data point only nudges the forecast partway.
Which one is “right” depends on what happens in week 7. If demand keeps climbing, the moving average’s faster reaction to the recent 130-125 pair will likely score better on MAPE. If week 7 demand drops back toward 100, exponential smoothing’s more conservative estimate will look better. This is exactly why you measure — theory tells you which method should work in general, but your own error metrics tell you which one is working for this specific product.
Where This Feeds Into Inventory Decisions
A forecast isn’t the end goal — it’s an input to the inventory calculations covered in What Is Inventory Management?. Your forecasted demand rate feeds directly into EOQ (how much to order) and, combined with a measure of forecast error, into Safety Stock (how much buffer you need against the forecast being wrong). A better forecast doesn’t just look good on a report — it directly reduces how much safety stock you need to carry for the same service level, which is real money saved.