Inventory Forecasting Methods That Turn Sales into a Buy

Last month's buy for lamp L-14 was 22 a week because that was the average. This week sold 30, a holiday is three weeks out, and the same average is still sitting on the PO. You will be short, then you will overbuy the dip. Inventory forecasting methods are the different ways to turn that history into the next number. They disagree on purpose. Picking one and checking it on a few SKUs beats blending five of them into a chart nobody buys against. What are inventory forecasting methods? They estimate what you will sell next from recent history, a seasonal pattern, or a judgment call on a new SKU. A moving average treats recent periods as equal. Exponential smoothing leans toward the latest week. Check one method against what you actually sold before you buy the whole catalog that way. The forecast is not the purchase order. It becomes the demand input in the reorder point: average demand times lead time, plus safety stock. NC State's reorder point tutorial is about that trigger. A pretty forecast that never reaches the PO is a report. How does a moving average forecast work? Add the last few periods and divide by how many you used. L-14's last four weeks, as an example, were 20, 28, 16, and 24. The sum is 88. Divide by 4 and the forecast is 22 units next week. Each new week, drop the oldest sales figure and add the newest one. The 20 falls off, a new week joins, and the average moves. Four weeks is a choice, not a law. A longer window ignores a real jump for longer, and a two-week window chases one odd order. For a SKU that sells every week at a similar level, four to eight weeks is a sane start. Write the window down so the next buyer does not silently change it. How does exponential smoothing react faster? NIST's handbook defines single exponential smoothing as a weighted blend. The new smoothed value equals alpha times the latest actual, plus one minus alpha times the previous smoothed value. Alpha sits between 0 and 1. The forecasting section writes the same idea as an error correction. The new forecast equals the old forecast plus alpha times the last error, and the error is actual minus forecast. Use the L-14 numbers as an example. The old forecast is 22 and the latest week sold 30, so the error is 8. With an alpha of 0.3, also an example, the adjustment is 0.3 times 8, which is 2.4. The new forecast is 22 plus 2.4, which is 24.4 units. The four-week average still says 22, because 30 has not pushed the older weeks out yet. Smoothing already moved, which is why a higher alpha shows a jump sooner. NIST's advice on the constant is practical. Try values and keep the one with the smaller squared error on your own history. An alpha near 1 chases every spike. An alpha near 0 barely notices the 30. When do you add a seasonal bump? A flat 22 is wrong if this week of the year always runs hot. Build a simple index from history. If this calendar week has averaged 1.4 times a normal week for this SKU, that 1.4 is an example index, not a retail law. Apply it to the baseline. 22 times 1.4 is 30.8. Buy toward 31 for that week, then take the index back off when the week is over. Leaving 1.4 in place for March is how January's spike becomes April's overstock. You need more than one year before an index is anything but a guess. One huge November does not make a seasonal factor. Two or three years that show the same week, after you remove a one-off promotion, can. What do you do with a SKU that rarely sells? A weekly average lies when most weeks are zero. Say a spare part sold in 8 of the last 12 weeks, and the sale sizes were 2, 2, 3, 1, 2, 4, 2, and 2. That is an example. Average sale size is 18 divided by 8, which is 2.25 units. The gap between sales is 12 divided by 8, which is 1.5 weeks. Demand per week is about 2.25 divided by 1.5, which is 1.5 units. A four-week average that happens to land on four quiet weeks forecasts 0, and you will not reorder a part customers still buy. Split size and timing for those SKUs. Forecast how many units a sale usually is, and how often a sale shows up. A plain average is for items that sell in most periods. Stockouts distort the history too. If you were at zero for 2 of 7 days and sold 10 on the other 5, those 10 are not a full week of demand. A rough scale-up is 10 divided by 5, times 7, which is 14. Label it as a sketch. It is better than treating a stockout as a customer who wanted nothing. Which method should you start with? What the SKU does Start here Sells most weeks, no obvious season A 4 to 8 week moving average Just changed level and the average is late Exponential smoothing, alpha chosen by smaller error on past weeks Same weeks run hot every year Baseline forecast times a seasonal index, then remove it New, with no sales history A written judgment for the first buys, replaced once you have about a month of sales Long gaps between orders Sale size and gap, not a weekly average full of zeros I'd rather run one method per SKU and write down the error than average three methods into a number nobody can explain. After a month, compare forecast to actual on the SKUs that matter. If the average error is consistently high or low, the method is biased, not unlucky. Demand planning in OneChannelAdmin is where a forecast can drive the replenishment buy instead of staying in a sheet. The stock that forecast is trying to cover is inventory management. The PO it should create is order fulfillment. How do you test a forecast on one SKU? Pull weekly sales for one SKU that was in stock. Mark weeks you hit zero so those weeks are not treated as true zeros. Compute a 4-week moving average and write the forecast for the next week. Compute an exponential smooth with one alpha, using NIST's form: old forecast plus alpha times the last error. If the week is seasonal, multiply by an index you can point to in prior years. If you cannot point to it, skip the index. Buy or pretend-buy that one number. A week later, subtract forecast from actual. Keep the method with the smaller absolute error over the next four weeks. Change alpha or the window only after that, not mid-week. Questions about inventory forecasting methods What are the main inventory forecasting methods? A moving average, exponential smoothing, a seasonal index on top of a baseline, and a judgment forecast when the item is new. Items with long gaps between sales need sale size and timing split apart, because a weekly average fills those gaps with zeros. How do you calculate a moving average for inventory? Add the last several periods of sales and divide by the number of periods. Four example weeks of 20, 28, 16, and 24 sum to 88, and 88 divided by 4 is a forecast of 22. What is exponential smoothing in plain numbers? The new forecast equals the old forecast plus alpha times the last error. With an old forecast of 22, an actual of 30, and an alpha of 0.3, the error is 8 and the new forecast is 24.4. NIST describes that error adjustment, and a smaller squared error is how you judge alpha. Should you forecast a stockout week as zero demand? No. A zero because you had nothing to sell is not a zero because nobody wanted it. Scale the in-stock days up as a sketch, or exclude those days, before the average learns the wrong lesson. How do you know the forecast is good enough to buy against? Compare it with the next period's actual sales for a month. A method that is always high or always low is biased. Keep the one with the smaller absolute error, then feed that demand number into the reorder point. OneChannelAdmin Team writes about demand forecasts, reorder points, and the purchase orders that follow them.