Forecasting is arguably one of the essential important functions in ensuring that customer demand can be met when requested. The forecast should reflect the best possible projection of future demand. Past sales history, although not always ideal, is one of the best sources of information on which to base forecasts. The advent of powerful computer hardware has made the process of testing many different methods to determine a “best-fit” for each individual item very fast, and has taken much of the guesswork out of the forecasting process.
But a computer calculated forecast is not the end of the forecasting process, it is only the beginning! A calculated forecast based solely on history will only be reasonable if the future turns out to be the same as the past. That is a very shaky premise on which to build a company’s supply chain operations, particularly when there area normal past events which gave rise to actual history data that does not fit the normal demand pattern for a product. The real task of a forecaster is first to evaluate the legitimacy of the calculated forecast, and then to apply their own experience along with information from other departments in the company and from customers to refine the calculated value to more closely reflect future expectations.
This may seem like a daunting task, especially if there area large number of products and/or a significant number of locations to manage. Nowhere is the adage more appropriate that“working smart” is much better than “working hard”. Consider Pareto’s rule of thumb that 20% of a company’s product line accounts for 80% of its business. Since it takes about the same amount of time to review any single item’s forecast, concentrating on those 20% should yield much greater impact than a comparable amount of time spent on the same number of items from among the other 80%. The careful use of ABC analysis (see CP Section 11) can be extremely useful in determining where precious time resources should be spent.
The Avercast forecast calculation process is built on a“tournament” philosophy. There are a total of 204 different methods available which are grouped into“personalities”, each of which is intended to provide optimal results for different situations.
Each item in the database is subjected to an individual “tournament”in which various methods are tested over a period of time using the actual demand history. Based on parameters set inForecast Policies (see Section 4.3) a certain number of most recent periods of history are isolated, and the remainder of history is applied against each method to calculate forecasts for the isolated months. For example, if the number of historical periods to consider is set to 24 and the target percentage (i.e., the percentage of the periods to be isolated) is 15%, then the most recent 4 (15%of 24 = 3.6) periods of history would be isolated and the remaining 20 periods would be used by each method to forecast the most recent 4 periods.
For each method an error factor is calculated called the Mean Square Error (MSE). The lower its MSE, the more closely a method matches the pattern of the historical data. Out of all the methods tested, the one with the lowest MSE is selected for that item and is used to calculate the future forecasts based on the most recent history. In this way, each item in the database is forecast using the best fit method for the item.
The methods which are used in any tournament are determined according to forecast “personalities”. There are 6 defined personalities available, whose profiles are as follows:
· 0 – provides the best pattern matching, but will guess wrong more often than Personality 1
· 1 – gives best compromise between pattern matching and accuracy in short term (1-2 months)
· 2 – (default) best for longer term forecasting,especially when annual seasonality is present
· 3 – includes all methods
· 4 – “Grand Rapids” good for forecasting slow moving products
· 5 – best for weekly forecasting based on 52-week seasonality
When generating monthly forecasts based on monthly history,updates to the historical time series data normally occur once per month after the corporate systems have completed their month-end processing. The process of updating this information is usually done monthly, although occasionally more frequent updates are requiredif adjustments to history occur at other times of the month. When the latest history has been loaded into the database, new forecasts are calculated. For complete details refer to Section 5.
Because the forecasts play an integral role in the replenishment planning process, updating of the forecasts should occur as soon as possible after the updated history becomes available. But as mentioned in the previous section,that is only the start. To the extent that time is available for the purpose, review of the forecasts and making adjustments based on intelligence from across the company is critically important to a successful forecasting process. This should be done whenever relevant information becomes available.