Mastering demand forecasting of short lifecycle products

May 22, 2019 5 min

The lifespan of the average magazine ranges from just a couple of days to a few weeks. Books generally stay in demand for a few months before losing relevance. Home electronics such as phones and flat screens sell for a couple of years at the most.

The common tie here is a short lifecycle. Interest in these products typically peaks upon launch or soon thereafter, at which point demand drops, sometimes radically, and dies out because the product loses its relevance, or a newer item becomes available. These rapidly shifting sales patterns make demand forecasting of short lifecycle products enormously difficult.

In this white paper, we’ll first explain why it’s just not feasible to forecast demand for short lifecycle products using their own sales data alone. More importantly, we’ll then show how accurately forecasting for such products is possible through intelligent statistical modeling that uses data from previous product introductions. 

Short Lifecycles Are Difficult to Forecast

Imagine that a new Harry Potter book has been announced, and you’ve been tasked with forecasting its sales. In the first week, your store sells 1000 copies. Using only this information, a logical forecast for the next week’s demand would be 1000 copies again, right? In practice, this forecast would be far too high because so many book-lovers bought the book during the first week. There would be fewer buyers left to account for the second week of sales.

Say, though, that sales during the second week amount to 500 copies. Based on two weeks of sales now, the logical forecast for the third week and the fourth and so on should simply follow the sequence: 1000, 500, … – 0 copies! This approach, of course, leads to far too low a forecast. (See Figure 1 for an illustration of these two forecasting approaches compared with the actual sales pattern.)

Figure 1: Lifecycle profiles are usually non-linear (green curve). Consequently, linear approaches to demand forecasting, such as setting the forecast equal to the previous week’s sales (red curve) or making a linear extrapolation based on the two previous weeks (yellow curve), result in poor forecasts.

Clearly, forecasting demand for short lifecycle products based solely on the products’ own sales data is infeasible.

Much better forecasts can be obtained by estimating the sales profile — i.e. the distribution of sales between the weeks following the product launch — from the historical sales patterns of similar products. Applied to the example above, the drop in sales from Week 1 to Week 2, or even sales over the book’s entire lifecycle, can be forecasted based on sales patterns from previous Harry Potter books.

Optimal Demand Forecasting of Short Lifecycle Products

There are many ways to go about estimating a lifecycle profile from historical sales data. Again, we’ll start with the most obvious solution: to first calculate the average lifecycle profile from past sales of similar products, then apply that average when forecasting demand of new products. Unfortunately, this approach still fails when applied to real-life retail data because it produces random variation in the sales profile that negatively impacts forecast accuracy (see Figure 2). Furthermore, planners may find it difficult to trust forecasts that seem to jump randomly up and down.

Figure 2: Calculating the lifecycle profile directly from past sales data, without any intelligent modeling, leads to noisy forecasts. The statistical noise is emphasized when sales volumes are low.

Optimal results can be attained only through intelligent modeling that makes good use of our a priori knowledge of a sales profile’s shape.

Our research shows that we can capture the shape of a lifecycle profile by using differential equations. We begin with the premise that that short lifecycle products have a fixed market potential. It follows that the number of potential buyers, and consequently the rate of sales, will decrease towards the end of the lifecycle as that potential is consumed. This thought, when expressed mathematically, results in a differential equation, the solution of which is a decreasing exponential function. This simple curve actually captures the general shape of many lifecycle profiles pretty well.

The model gains flexibility and becomes a universal fit for real-life lifecycle shapes by further assuming that those customers who have already bought the product may recommend it to their friends or followers on social media — the “imitators.” The resulting sales curve (see Figure 3) corresponds to the Bass diffusion model.

Figure 3: The Bass diffusion model is very flexible and applicable to most lifecycle patterns.

Proven in Practice

The lifecycle profile forecast model implemented by RELEX estimates the shape of a new product’s lifecycle based on completed lifecycles from similar products. This profile is then used when forecasting demand for the new product.

This kind of lifecycle profile has been validated on many product categories, ranging from fashion to books, with excellent results. Figure 4 shows the forecast model in action. In this case, a full lifecycle forecast was calculated for a product prior to launch.

Figure 4: An example of applying RELEX’s lifecycle forecasting model to forecast demand over a product’s full lifecycle. In this example, the new product is introduced in the beginning of April. The sales for similar products have historically peaked about 6 weeks after the launch, and therefore the highest sales are forecasted for mid-May.

Because of intelligent modeling, this approach can produce reliable forecasts even based on relatively low volumes of historical sales data. When powered by RELEX’s efficient in-memory database, both model fitting and forecasting become very fast operations.

List of resources:

Towards a weather-proactive supply chain

Improve demand forecasts by considering cannibalization and halo effects

Decision Science and Pragmatic AI in Retail

Create meaningful forecasts for new products

Better promotion forecasting in 4 steps

Written by

Tuomas Viitanen

Tuomas Viitanen

Senior Data Scientist