Forecasting New Product Life Cycle Curves: Practical Approach and Empirical Analysis

Before a firm starts manufacturing a new product, it wants to have an accurate forecast of weekly demand over the product’s life cycle. Such a forecast is vital for production and logistics planning. Accurate forecasts are even more critical for products with short life cycles (6 to 12 months) and long lead times (6 to 12 weeks)—a common combination for electronics manufactured overseas—because the lead time constitutes a significant portion of the lifespan and many of the logistics and production decisions need to be made far in advance. However, generating such a forecast is difficult. After all, the product is new. Traditional time-series forecasting methods based on historical data do not apply here; there is no historical data for this product yet! However, all is not lost. Oftentimes, new products are not earth-shatteringly new, and not unlike anything customers have ever seen (the first iPhone being a notable exception). Instead, many new products are similar to some subset of the firm’s existing products. This level of similarity could range from low (a significant change is being made to a product line intended for a new market) to high (a product for which a firm has significant historical data is receiving a very minor upgrade). Yet despite having data on related products, many firms do not leverage this data—even for products that are very similar to existing ones—instead relying on market research, executive opinion, and sales force input.

In our paper published in Manufacturing and Service Operations Management, we propose an approach to forecast the weekly product life cycle demand of (not earth shatteringly) new products before launch. The approach leverages managerial input (such as product life cycle length and sum of total volume) as well as data from other products in an attempt to minimize forecast error. We test our method on data from Dell—whose products tend to have short life cycles—and find that this approach improves forecast error over Dell’s previous forecasts by 2–3%, equating to about $2 savings per unit over millions of units. The contribution of this forecasting approach is validated on three more electronic products of a second firm which produces computer accessories. These improvements come from applying the algorithms directly without any expert post-processing. In reality, the demand planners at Dell may be able to adjust our automated forecasts to include promotions or other known seasonality patterns in order to decrease forecast error by even more.

There are three key steps to generate product life cycle curve forecasts, which we summarize in Figure 1:

1)      Fitting. For each old product in a firm’s catalog, the historical demand must be cleaned and normalized to a time length of one and to a sum of total demand equal to one. In this way, product life cycle curves can be compared to each other even if the products have vastly different lengths or volumes. Then, for each product’s normalized historical demand, a “curve” is fit through the data. We test three versions of curves: Bass curves (one of the most popular “new product” curve shapes), polynomial curves (a versatile smooth curve), and piecewise linear curves (namely, trapezoids and triangles).

2)      Clustering. Once each old product has a fitted product life cycle curve, these fitted curves are clustered together into distinct groups. This clustering can be expert-driven (managers may cluster together products that seem similar) or data-driven (an off-the-shelf time-series clustering algorithm can be applied). The result is a set of clusters, each of which contains a subset of the products whose product life cycle shapes are similar to each other.

3)      Forecasting. To forecast the weekly product life cycle demand of a new product, a manager must first choose the cluster which she thinks is the best match for the product. The individual product life cycle curves within this cluster are aggregated (by averaging or fitting) into a single normalized product life cycle curve. Then, this curve is scaled to be the length of the new product’s life cycle (this length is often known at Dell because timing is tied to the release of new chips or consistent release cycles). Finally, the manager scales the curve to the forecasted volume of the new product (at Dell, a product’s lifetime volume is often easier to forecast than the breakdown of when this demand will occur at the week level).

4 steps are illustrated using graphs

 

We apply this approach to 170 products at Dell, sold over a three-and-a-half year period, covering more than four million units sold. We find two major insights:

1)      Simple piecewise linear curves perform the best in terms of forecast error. They improve mean absolute scaled error by 2.8% over Dell’s previous forecasts, as compared to a 1.6% reduction for smooth polynomial curves and a 0.3% increase for the popular Bass curves. Simple piecewise linear curves are easier to interpret. While the parameters of a fourth order polynomial curve have little meaning in and of themselves, piecewise curves such as triangles and trapezoids map to common stages of product life cycles which are already used at Dell: launch is the first rising portion, sustain is the flat middle portion, and end-of-life is the last decreasing portion.

2)      Data-driven clustering performs the best. In addition to the data-driven time series clustering approach, we also tested clustering by category (desktop, laptop, etc.) and product features (similar chip, hard drive, memory, etc.). The data-driven approach resulted in the lowest forecast error.

We also test our approach on a second smaller dataset with similar results and insights. Additionally, we make the Dell product life cycle data set available to allow other researchers to test other methods of product life cycle forecasting (Acimovic et al. 2018).

More than a quarter of an average firm’s revenue comes from new products but profits lag this proportion of revenue. Because new products are expensive to support, even small cost savings can have a large impact on a firm’s bottom line. Being able to accurately forecast new product weekly demand before launch is one way to cut costs; more accurate forecasts result in less unused capacity or overtime costs at the factory, fewer expedited shipments, more favorable contract terms with the manufacturer and shippers, and so on. We show here one forecasting method which reduces forecast error at two firms, resulting in potential savings of millions of dollars at Dell alone.

 

References

Acimovic J, Erize F, Hu K, Thomas DJ, Van Mieghem JA (2018) Product life cycle data-set: Raw and cleaned data of weekly orders for personal computers. Manufacturing Service Oper. Management, ePub ahead of print May 8, https://doi.org/10.1287/msom.2017.0692.   

Hu K, Acimovic J, Erize F, Thomas DJ, Van Mieghem JA (2018) Finalist—2017 M&SOM practice-based research competition—Forecasting new product life cycle curves: Practical approach and empirical analysis. Manufacturing Service Oper. Management, ePub ahead of print May 7, https://doi.org/10.1287/msom.2017.0691.

Comments