Simplify Supply Chain Forecasting
Is 100% forecast accuracy attainable? Should it be? Wouldn’t that be called an order?
Anyone that has ever had to forecast demand for products or services knows that obtaining a consistently high forecast accuracy is part science and part magic. Minimizing forecast error requires a background in data analysis and statistics while also having a solid understanding of the products, markets and customers you are trying to predict demand for. Even with all of the knowledge we have collected and the vast resources of information at our fingertips, good forecasters realize that producing a forecast that approaches 100% accuracy approaches impossible (and, if it does happen, is purely by chance). This makes the demand planner’s role challenging and rewarding at the same time.
An improvement in forecast accuracy can have a significant impact on the bottom line by reducing inventory buffers, obsolete products, expedited shipments, DC space required and other non-value added work. At the same time we see higher customer fill rates, improved customer satisfaction, and increased revenue and margins. Clearly, forecast accuracy is very important.
With all of these benefits, some have proposed the advent of Lean & JIT (just-in-time) have rendered forecasting obsolete. Even if, and that is a big if, your manufacturing and distribution cycle time is less than or equal to your customers expected order fulfillment time you still have to plan for and acquire raw materials and components. These acquisitions must be planned. At best, Lean and JIT push the push/pull boundary back up the supply chain a bit; they dont eliminate the need to create a demand forecast.
Learn from your Peers
Now that we’ve established a demand forecast is needed and critical to your company’s success, how can a demand planner improve the accuracy of the forecasts they produce? Learn from others’ experiences. In this post, we will look at the forecast methods that have proven time and again to produce superior results.
Many items or groups of items that you forecast will have some type of trend (either up or down) but may not display a seasonal pattern. When you have at least a few months of history the Modified Holt, or double exponential smoothing technique, can be utilized for forecasting trended demand. When demand also has a seasonal component, the Holt-Winters triple exponential smoothing technique can provide very good results if you have at least a year of historical demand data.
On the flip side, your portfolio may have many items with no discernable demand patterns. In this case a simple moving average technique can provide the most accurate forecasts. Recent research even shows that for many items with random demand a Naïve forecast might be the best. A Naïve forecast is created by taking the last period’s demand as this period’s forecast. Remember though, this is not a blanket rule as some items may be better served through a simple KanBan system where product is replenished as needed to keep a certain amount on hand. Inhibiting the forecast and using alternative inventory replenishment means KanBan may be more appropriate for items that are low value and volume, and are easily obtained.
One of the hardest class of items to forecast are those that have intermittent demand. Often demand patterns for these items look digital with random spikes and zero demand periods. A Modified Croston Method handles low and lumpy demand that exhibits either a patterned variation or no pattern by looking at available history to classify each demand element relative to those around it. It classifies the periods into peaks, valleys, plains, plateaus, up-slopes and down-slopes, and then measures the duration of plateaus and plains, as well as the severity of peaks and valleys. Through a pattern-fitting analysis, this method creates regularity over time, attempting to fit the pattern to the history and averaging for low and high points.
The speed and volume of new product introductions continues to escalate creating potential headaches for supply chain teams. Forecasting demand for new products can be a time consuming effort. The silver lining here is that new products are rarely completely new. The product may already exist in a different market, it might be very similar to an existing product or it may replace an existing product or combination of products. An effective method of generating new product forecasts is to use demand variations or extensions from existing products, families or brands. This method, called Derived Modeling, draws on the historical data of existing products and applies all or a percent of that demand to create the new forecast. Since the existing historical data already contains base, trend, and if appropriate seasonality information, the new product forecast will also reflect this information. Derived Modeling combined with causal effects and management overrides can provide a realistic and dynamic forecast for new products.
Attribute-Based Modeling is another effective method to forecast new product demand as well as demand for short-life cycle products or products with limited history. Most products can be associated with another or grouped through common attributes, i.e., functionality, features, size, color, style, performance, material type, etc. A demand profile can be determined for a group of products with a certain attribute. A set of profiles can be maintained and when the demand for new products or short-life cycle products is required the product can be associated with the appropriate demand profile. Advanced demand planning systems will automatically create demand profiles and suggest the appropriate profile. These systems will also automatically monitor forecast accuracy and suggest new profiles as actual demand data is acquired.
A specific type of Attribute-Based Modeling is called Proportional Profile Planning (PPP) which is used to disaggregate higher-level forecasts for groups of items into lower-level forecasts for individual items using user-defined attributes. PPP is often used in the apparel and footwear industries where the number of forecasts can reach into the millions due to size, color, style, location combinations, and more. Often the demand patterns for the lower level items are very random and difficult to forecast; however, as these items are placed into product or location groupings the demand pattern becomes much more forecastable. PPP allows the planner to manage less forecasts while producing more accurate forecasts at some user defined grouping, and then take advantage of this higher accuracy by systematically far-sharing the grouped forecast down to the individual items. PPP enables less work and higher forecast accuracy. Thats a Win-Win.
Finally, for those promoted products, Causal Modeling can help determine the effect or “lift” a certain promotion or set of promotions will have on the demand for a product. A very powerful type of causal modeling is called Neural Network Modeling which uses “backcasting” to pre-train the system to calculate new promotion plans. Backcasting is a technique that uses actual “lift” past promotional events to learn the relationship of future promotional programs to expected “lift”. Backcasting dramatically shortens the amount of time needed to obtain actionable information from a causal modeling system.
There is no doubt that in today’s ever increasing complex supply chain, using the appropriate forecast method for every stage in a product’s lifecycle is more important than ever. Competition is too fierce and customer expectations too great to settle for low forecast accuracy. Demand planners need to be armed with the latest enabling technology and be able to use that technology to continually improve forecast accuracy to drive business benefits.
What forecasting methods does your company use? Are they driving continual improvement in forecast accuracy?