I received a call from one of my customers this afternoon. “Jon, something is wrong with my demand forecast. We usually have a sharp drop in usage in January because several customers have shutdowns. But the forecast is recommending we stock almost twice as much as we did last year. We need a better forecasting model.”
We had worked extensively with this customer to develop a comprehensive set of forecasting models that I was pretty sure would provide reasonable estimates of future usage. I looked at the customer’s data to try to determine why the system was producing an inaccurate forecast. I paid particular attention to the usage recorded in the previous November through March (the months surrounding January):
“Mike,” I said, “I’m looking at last year’s usage history and I don’t see a decrease in usage in January. In fact, January’s usage of 1,390 pieces is close to 15% higher than December’s usage of 1,210 pieces.”
Mike quickly responded, “That’s impossible. Let me look at the data.”
The phone was silent for a few minutes before Mike sheepishly came back on the line. “Now I remember what happened. Right after our year-end physical inventory we liquidated 790 pieces of excess stock with a broker. We never adjusted our usage history to take out that unusual sale.”
Today’s competitive market requires accurate forecasts for the future demand of each product in inventory. We are constantly looking for ways to reduce the forecasting error (i.e., the difference between a forecast and the resulting usage). But these efforts are often frustrated by unusual activity imbedded in historic usage data.
Most computer systems realize the importance of usage history that is not exaggerated by unusual activity. They try to identify, and possibly correct, possible unusual usage. A common method used by many computer systems compares the usage recorded in the inventory period just completed to the total usage in the several previous inventory periods. One of the proponents of this method is recently retired inventory guru Gordon Graham. He long advocated that a company might have experienced unusual usage in the month just completed if the usage in that month was greater than the total usage in the five previous months*. For example:
The usage of 2,500 pieces in June is greater than the total usage of 1,825 (410+290+375+450+300) pieces recorded in the five previous months. But this method would not have identified my customer’s January usage as unusual:
Note that January’s usage of 1,390 pieces is not greater than the total usage of the five previous months. In fact, unless you knew about the anticipated drop in usage due to customer shutdowns, it would be hard to detect any unusual activity. For this reason, unusual usage should not be identified by comparing usage in the month just completed to the usage recorded in previous months. It should be identified by comparing usage in the month just completed to the forecast of usage in that month.
There are software packages that compare the forecast of demand to actual usage and automatically adjust usage for any suspected unusual activity. One system presumes that any usage is greater than x%, or less than y%, of the current forecast represents unusual activity that probably will not recur. This system will correct the usage for the inventory period to equal the forecast plus x% or the forecast minus y%:
In the example above, both the high and low limits for unusual usage are set to 75%. The high limit for usage is the forecast of 164 pieces plus 75%, or 287 pieces. If the actual usage for the inventory period was more than 287 pieces, it would be corrected with a reduction to equal 287 pieces. The low limit is 164 pieces less 75%, or 41 pieces. If the actual usage for the inventory period was less than 41 pieces, it would be increased to equal 41 pieces. Other systems correct usage in a similar way using statistically calculated standard deviations.
Unfortunately this method does not always accurately correct unusual usage activity. To understand why, we must look at the three reasons why actual usage might deviate significantly from the forecast:
- Unusual usage activity that will probably not occur again in the future.
- The start of a significant new trend in the usage of the product.
- The wrong forecast formula or method being applied to the item.
The process of automatically adjusting history for unusual usage will address most unusual activity that will probably not recur in the future. But if a product suddenly gains or loses popularity, automatic usage adjustment might actually remove significant usage history. For example, if a product suddenly experiences a 150% increase in usage in one month and the product is destined to remain popular, automatic adjustment will take sales that will probably occur again out of usage. The result will be future forecasts that will not accurately reflect your customers’ needs. And, if a system continually corrects usage to be no more than a certain percentage of the forecast, how will you ever know if the usage activity is actually normal, but you are using the wrong forecasting method for the item?
We have found a better method is for a computer system to identify, but not correct, items that might have experienced unusual usage activity. Buyers, planners, and/or salespeople review the list of items and determine whether unusual activity actually occurred, a significant new trend has begun, or the wrong forecasting method has been applied to the item. They can then manually adjust actual usage to reflect what usage would have been had no unusual activity occurred. For example, a system might list items whose usage was more than 300% (i.e. three times) of the forecast or less than 20% of the forecast. The buyer or other appropriate employee would then review the list of products and make usage adjustments as necessary.
Advanced systems allow the user to vary the definition of unusual activity for different types of products. For fast moving “A-ranked” products it might be important to see any situation where usage is more than twice the forecast (for example, if the forecast for an item was 10,000 pieces and actual usage exceeded 20,000 pieces). At the same time, unusual activity for a slow-moving product might be anything more than 400% of the forecast. The forecast for an item might be one piece, and unusual activity might be usage of more than four pieces in the month.
Regardless of how your system identifies unusual usage, in order to accurately forecast future demand of products it is imperative that historical usage be corrected for any unusual activity.
*Graham, Gordon, Distribution Inventory Management for the 1990s, Inventory Management Press 1987, page 77.