Demand forecasting in pharmaceutical supply chains: Novo Nordisk case study

Abstract

[en] Forecasting is a common use case in the field of Predictive Analytics and one of the key building blocks of any Supply Chain. This relevance is even magnified in the pharmaceutical industry, where a stock-out does not merely carry a monetary impact but might also tragically affect people’s health. In light of the aforementioned, this thesis has a twofold aim. Firstly, improving Sourcing Operations’ forecasting process in terms of both accuracy, standing at 59%, and efficiency, currently a 5-day process. Secondly, helping to shed some light on the univariate-multivariate debate in the forecasting realm. Attaining these goals required uncovering the best methods in the forecasting realm by scouring the existent literature; both univariate and multivariate applications were eventually pursued. With respect to the former, classic techniques such as simple average, Autoregressive Integrated Moving Average and Exponential Smoothing were chosen. These were also combined in an ensemble in order to leverage each model’s strengths and keep each other in balance. Amongst the several multivariate techniques available, the choice fell upon Gaussian Process Regression and its capability to model complex functions by means of kernels. Identifying these complex structures also required combining such kernels, and given the sheer amount of time series, a mechanical greedy search strategy operating as a forward selection method was devised. Results showed how a multivariate approach (68%) outperformed the univariate models (63%), albeit the former (18 hours) was much slower than the latter (30 minutes). Finally, combining the "best of both worlds" enhanced accuracy up to 71%. With respect to the first goal, these outcomes increase accuracy by 12 percentage points and slash forecasting time to less than a day. In terms of the second goal, these results seem to argue in favor of multivariate methods, demonstrating that these perform better since they can leverage external information; yet, the best-of- both-worlds approach also shows the lack of a clear-cut answer on the matter: each class of models might outperform the other under certain conditions.