[This article first appeared in the January 1992 edition of _Industrial_Engineer_.] QUALITY FORECASTING DRIVES QUALITY INVENTORY AT GE by Robert M. Duncan Quality forecasting? Everyone knows that forecasts are always wrong. Variation is inherent in this process. After a few years of running an unsuccessful forecasting system and after two years of a successful one, GE Silicones has learned that forecasting is not a black box with history going in and forecasts coming out. It is a process that requires feedback, judicious use of statistical process control, and carefully applied adjustments. Lessons learned the hard way Demand forecasting for was originally a monthly calculation by a homegrown mainframe black box system. Identifying and reviewing exception items was difficult and seldom done. Furthermore, while the system looked for trends and seasonality, it used the outdated base index method. Also, while the company uses a 5-4-4 calendar (weeks per month per quarter) for reporting purposes, the system performed no calendar normalization on the demand data. Safety stocks were computed from forecast error, but the mean absolute deviation (MAD) was used, rather than the standard deviation. These things added up to inaccurate and insufficiently monitored forecasts. A dramatic example of this is an SKU (with an out-of-control forecast). At the end of 1987, the system decided on a model that included a huge trend, partly as a result of successful promotions that generated record demand for December. Based on the trend, the forecast for January was even higher. Meanwhile actual demand for January was very low (because everyone stocked up at the December price). This increased the forecast error, which in turn caused the system to recommend a higher safety stock. In short, while actual demand was low, the forecast was high and the safety stock was increasing. The resulting large inventory was not just undesirable from a financial point of view, it meant that expensive capacity was tied up producing the wrong product. Furthermore, service for other products - which were needed - suffered. Efforts to improve the situation the classical problems seen in production and inventory management. Marketing product managers, in an effort to insure they would get some of the production for their customers, over-forecasted their own products. The biased forecasts drove heavily frontloaded master schedules. The production schedulers, constrained by capacity from scheduling to cover the forecasts, scheduled what they guessed was actually needed - and they became the real forecasters. And when they guessed wrong the real expediters] The net result: service suffered and inventory climbed. Organizing for success Success was achieved by addressing three areas: organization, software and focus. First, a Materials and Logistics department was created and given the responsibility for customer service, inventory, scheduling, purchasing and forecasting. Second, the Finished Goods Series (FGS) software from E/Step Software Inc. of Tieton, Wa., was installed. FGS is a PC-based demand forecasting and inventory management package that has facilities for identifying and reviewing those exception items that require something other than routine handling. The package includes a simulation module that allows the user to view the results achieved with various strategies for handling exceptions but before the changes are actually made. In this way a number of scenarios can be compared. Third, Pareto analysis was used to identify the few critical items that make up the majority of GE's business. About 20 percent of the products account for 80 percent of the orders. The control limits for these items were set to much tighter tolerances than were used for the rest of the items. This enables a focus of attention on the few items where it is merited and where there are substantial benefits associated with any improvements. initialization and overview About five years of demand history was loaded into FGS. It then determined the optimum model for each SKU. The problem of a 5-4-4 calendar was handled by loading calendar period-ending dates (past and future) and letting it perform the normalization automatically. Any seasonality it detects is true seasonality, and not just caused by the accounting calendar. Each month, the demand for the month just completed was loaded. Next, the demand to check for reasonableness was "filtered." The forecast and actual demands were compared with the marketing intelligence to see whether it helped or hurt. Then FGS used the computed error to revise the forecasts and models using adaptive smoothing. This process also identified exception items which were reviewed and, in some cases, modified. Finally interface files were generated and the new forecasts and safety stocks were uploaded to the mainframe to drive Master Production Scheduling. Why not refit monthly? One frequently asked question is, "Why revise? Why not just fit new models every month?" GE's products as do most companies' items - exhibit a fundamental stability most of the time. If a forecast model truly represents the underlying demand for a product, then that model will be effective for some period of time. It is not that one model will work one month, and a totally different one is required next month. There are most certainly changes, but they are usually changes in degree, not in kind. One would not see a product with a level, trend and quarterly seasonality go to just a level and trend next month and to semiannual seasonality the month after that. The adaptive smoothing and error tracking that occurs in the forecast revision process has two purposes. The first is to make those changes in degree (but not in kind), which keep the model up-to-date with reality. This handles situations such as a trend that is gradually flattening or seasonality that is becoming less pronounced as the company expands into new markets in what used to be the off season. The second purpose of the revision process is to identify those items where the chosen model is suspected of being no longer appropriate (i.e., changes in kind, not just degree). It is vital for the system to identify what used to work for an SKU but no longer does. With that knowledge, one can investigate the cause of the change, be it new competition, product changes or whatever. This is counter to the belief that every model should be tried on every SKU every month. Spending effort on exceptions Each SKU's forecast is an output of a process. Once a month the forecast can be measured against the actual demand. The difference is forecast error and is almost always normally distributed. There is not enough time to review every SKU's error. But using techniques of statistical process control and control charting, it is possible to detect exceptional errors. Then one can research the special causes, correct the forecasts, reduce the projected forecast error and tighten the control limits. Since safety inventory is directly proportional to forecast error then reduced error results in safety stock reduction and reduced inventory. Simulation to handle exceptions A monthly meeting with each inventory planner allows for a review of the exceptions and any questionable forecasts they have identified. The exceptions are reviewed in descending order by safety stock value, which means starting with the products that have the greatest potential inventory savings. The products are reviewed using the FGS Simulation Facility, with the goal of generating forecasts that based on knowledge of the products and markets - look correct on the graph (level, trend, seasonal pattern, etc.). In addition, the safety stock is monitored in an attempt to achieve reductions (via error reductions, not service reductions) wherever possible. This process is continued until either exceptions or time run out. Every item reviewed is not always changed, sometimes waiting until next month to see if the exceptional demand continues or returns to its prior pattern proves correct. When changes are made there is usually a significant reduction in safety stock. Filtering last month's demand When a control chart identifies samples outside the control limits, the data is not random variation. This signals the operator to review the cause and adjust the process. The same can be said of forecast error. The Demand Filter Report is a defense against order entry errors corrupting forecasts and can also help spot trend changes or outliers. Each SKU is given a filter sensitivity in standard deviations: Actual Demand > Forecast +/- 3 Std Deviations Actual demand falling outside the set values (either too high or too low) are not random occurrences and cause the SKU to appear on this report. The SKUs on the report are sorted in order from largest to smallest error in dollars. Trend changes If a trend change of an SKU is used then either service will suffer or inventory will build. The revision process identifies the SKUs for which there is some degree of suspicion that the model in use may no longer be appropriate that is, it looks for bias in the forecasts. The traditional method for identifying trend changes in forecasting is the Tracking Signal: Running Sum Error / Std Deviation - 4 An unbiased forecast has both positive and negative forecast errors. These errors "wash" in the running sum error and the tracking signal is insignificant. But if the demand exceeds (or falls short) of the forecast for a few months, then the remaining sum error accumulates and a tracking signal is hit. A sophisticated tracking signal called the parabolically masked cumulative sum of errors technique Reference 1 is used - quite a mouthful, but it does a good job. In statistical process control, the CUSUM chart is a highly sensitive tool for identifying a process beginning to go out of control. Monthly, the tracking signals are printed sorted by planner, and within that by descending dollar value of the safety stock. This is a favorite report as it is an early warning of changing demand patterns. It is used to focus attention and direct actions. An example of a changed trend is the product called F40 (See Figure IA and IB). Adaptive smoothing reduces forecast nervousness (overreacting to noise), but, as a result, it always lags a change. The tracking mechanism caught the explosive growth in the prior month that adaptive smoothing was too slow to catch. Merely refitting the model sufficed to correct the situation. As a result both the forecast level and trend increased. The lower error in the new model reduced safety stock by 18 percent. Eventually when this product matures, it is likely that a decreasing trend by the tracking signal will serve as a warning. Initially, the tracking signal sensitivity was set for all SKUs to four standard deviations Oust like the text books recommend) and two pages of exceptions were produced. That was all GE had time to deal with. A couple of months later, as users learned to use the tools better and the forecasts improved, the list dropped down to one page; and then to a half page. The sensitivity was heightened to three standard deviations and the list grew to two pages again. Bear in mind that as one tightens the sensitivity, one does a better job of monitoring the forecasts, and the exceptions generated are based on either smaller errors or fewer biased periods or both. Another way of looking at it is that tighter limits mean there is a greater probability that an exception is spurious and requires no action, but it does mean catching changes earlier. After a while the sensitivity was tightened again to 2.5 standard deviations and eventually, with further improvement, to the current 2.2. The greater sensitivity makes it possible to identify trends shortly after they change, which serves to further improve forecasts. This process cannot happen all at once, it has to happen over time. It is the continuing improvement in the ability to forecast which allows a user to tighten the sensitivity. Otherwise, only a longer and longer list of exceptions would be generated. Also, not all products are set to this tight sensitivity - just the ones that merit the close scrutiny. The others are set to wider limits and consequently require less attention. High forecast error Another indicator of forecast that is out of control is to compare the standard deviation of forecast error against its average forecast. Those SKUs that exceed a certain value are opportunities for improvement: Std Deviation/Monthly forecast > .75 Months The product 132 exhibits a case where the standard deviation was .8 months (See Figure 2A and 2B). Demand patterns are most easily identified by viewing a graph of the history. This first example is a classic illustration of the life cycle of a product (grow, mature, decline). The eye can readily pick up this pattern in the graph, and the planner - knowing the story behind it handles this exception by setting a demand history (DH) limit. This limits the history considered when refitting the model. Here, the SKU is treated like a young six month product. The system moved the item to fast smoothing (a higher alpha factor) which means the forecast will automatically drift down more rapidly if the demand continues to fall. Since the variability in the last 6 months was small compared to the prior months, the error reduced and the safety [SysOp note: no graphics were included with the electronic version of this article.]