Forecasting SLCP Demand: a Fuzzy Clustering Approach

This research investigates forecasting Short Life Cycle Product demand with a fuzzy clustering approach. The forecasting methods from Basallo-Triana, Rodríguez-Sarasty, and BenitezRestrepo (2017) are verified in this research. In their framework, fuzzy clustering is used to identify clusters of analogous sales profiles. A predictive model – Multiple Linear Regression (MLR) – is then trained on each cluster. They then assign a new product to a cluster, after which it follows that specific model for prediction. In verifying their methods, we provide proof that the use of the distance measure from Frigui and Krishnapuram (1999) does not minimise the fuzzy clustering cost function. Therefore, the original Gustafson-Kessel algorithm is implemented in this research (Gustafson & Kessel, 1979). Actual sales from a Dutch e-commerce retailer is then predicted with MLR. Forecasts are attempted to be improved with Multivariate Adaptive Regression Splines (MARS), logtransformations and a weighted forecast. The latter incorporates the fuzzy clustering approach, in which a new product is assigned to a weighted combination of all identified clusters. In terms of Root Mean Squared Error (RMSE) of out-of-sample forecasts, it is shown that analogue-based forecasting with MLR beats a simple naive benchmark. However, a first indication on MARS show no substantial improvement compared to MLR, while it does provide a great extra computational burden. Moreover, clustering performance seems to limit model performance, resulting in poor overall forecasting performance.