Short and hard-to-predict product life cycles make it important that firms master New Product Introduction (NPI) demand planning. A good New Product Introduction plan should account for:
- Sales volume over the long term at normal operating capacity, otherwise known as the baseline demand
- The demand profile during the initial launch period.
Let’s start with the baseline demand and then we’ll look at the launch profile.
Forecasting baseline demand for a new product is difficult due to the absence of historical sales information. So one common approach is to forecast it by projecting from past histories of similar products. If the new product is replacing another product, we can use the history of the previous item. Essentially it’s a 1 to 1 substitution process. There are various names for this technique, but we call it “super sessions”.
If there is no direct substitute, then another technique is to identify attributes of the product that are similar to existing products, so that you can group demand characteristics from those items to forecast baseline demand. You use the aggregated past history to forecast a new product under what is sometimes called a “dummy item”.
In order to do determine the most important characteristics and the weighting of those characteristics, we have to define the “split coefficients”. In the past, this has been done manually, but new techniques are being developed that can perform this process automatically. For instance, with some of our fashion industry customers we have been doing “advanced attribute clustering” using a mix of attributes to define product groupings. SKUs with the same sales profile are clustered into groups, and then machine learning software selects the relevant attributes and rules and assigns the new item to previously defined clusters.
The second challenge is defining the demand profile during the initial launch period. Because the launch period can be a period of very disruptive demand, being able to forecast the demand peaks is crucial for some businesses such as consumer technology or fashion retail. Using web analytics and product attributes, a rules-based machine learning model can predict the potential performance of a product launch. Early indicators such as web page visits, social media, control groups and data sources can suggest the contours of a new product launch early on. These are hard to account for with traditional forecasting techniques, but can be automatically identified by machine learning embedded within the demand planning software.
For example, a global leader eyewear company adds some 2000 new styles to its collection annually and defines their launch window as the first 90 days of sales. Previously, forecasts tended to over-estimate sales during the launch. Now demand planning software automatically clusters behaviors from past launches and selects the most probable behavior for the new product in the first launch period. These detailed, daily demand profiles helped reduce the global WMAPE (Weighted Mean Absolute Percent Error) by 10% and improve the forecast baseline on new launches by about 30%.
In another example, an electronics wholesaler introduces more than 50,000 new items each year. It uses web demand analytics and product attributes to build a rules-based machine learning model that accurately predicts potential product performance more than 85% of the time. In the process, the company has developed a clearer view of rules that define “dead dogs” and “rising stars,” and improved forecast accuracy for quantities over five months by 15%.
Forecasting demand peaks during the launch period is essential to businesses that realize the bulk of their sales revenue during the first months of the launch. That makes advanced demand planning software not only a better forecasting tool, but a revenue and growth generator.