We present an approach to forecast customer orders of ready-to-launch new products that are similar to past products. The approach fits product life cycle (PLC) curves to historical customer order data, clusters the curves of similar products, and uses the representative curve of the new product’s cluster to generate its forecast. We propose three families of curves to fit the PLC: bass diffusion curves, polynomial curves, and simple piecewise-linear curves (triangles and trapezoids). Using a large data set of customer orders for 4,037,826 units of 170 Dell computer products sold over three and a half years, we compare goodness of fit and complexity for these families of curves. Fourth-order polynomial curves provide the best in-sample fit with piecewise-linear curves a close second. Using a trapezoidal fit, we find that the PLCs in our data have very short maturity stages; more than 20% have no maturity stage and are best fit by a triangle. The fitted PLC curves of similar products are clustered either by known product characteristics or by data-driven clustering. Our key empirical finding is that, for our large data set, data-driven clustering of simple triangles and trapezoids, which are simple to estimate and explain, perform best for forecasting. Our conservative out-of-sample forecast evaluation, using data-driven clustering of triangles and trapezoids, results in mean absolute errors approximately 2%–3% below Dell’s forecasts. We also apply our method to a second data set of a smaller company and find consistent results. The online appendix is available at https://doi.org/10.1287/msom.2017.0691 .
We provide and describe a data set of N = 8,935 weekly, normalized customer orders over the entire product life cycle for 170 Dell computer products sold in North America over a three and a half year period, from 2013–2016. Total orders for these products exceeded four million units and well over one billion dollars in revenue. While Dell is historically known for fulfilling customer demand with a build-to-order approach, the products in this data set were designated as build-to-stock products. There are three elements in the data that, depending on the research application, researchers may want to identify or mitigate. First, some products have seemingly anomalous orders representing one-time purchases from large customers. Second, there are negative values for some products representing order cancellations. Third, end-of-life sales may be significantly influenced by management action. We present approaches for cleaning the data to address these issues. The supplemental material is available at https://doi.org/10.1287/msom.2017.0692 .
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.