A Step Towards Demand Sensing: Employing EDI 852 Product Activity Data in Demand Forecasting

Jente Emiel Van Belle, Wouter Verbeke

Research output: Chapter in Book/Report/Conference proceedingMeeting abstract (Book)Research

204 Downloads (Pure)


This paper deals with short-term demand forecasting of medicines in a US drugs factory based on historical sales and EDI 852
product activity data. Traditionally, demand forecasting relies on statistical methods such as ARIMA, and smoothing methods
such as exponential smoothing, to extrapolate the series of historical sales. Although these methods produce reasonably
accurate forecasts in many cases, companies are looking for an increasingly higher level of forecast accuracy to further enhance
the efficiency of their supply chains. With more and more diverse data becoming available at a higher velocity, it becomes
possible to employ non-traditional data sources and generate forecasts for increasingly shorter time periods. From this point of
view, in recent years, the concept of demand sensing emerged. However, in the scientific literature and empirical studies, the
concept has gained only very little attention to date. Essentially, demand sensing comes down to leveraging a combination of
near real-time (i.e. orders), downstream (ideally point-of-sale) and external data to improve the accuracy of short-term forecasts.
In their purest form, however, the traditional models do not allow for the inclusion of covariates. In this paper a particular type
of downstream data, EDI 852 product activity data (a data standard used for exchanging product related information between
the supplier and the (end)customer), is employed in producing short-term weekly demand forecasts obtained from linear,
dynamic (i.e. with ARIMA errors) and lasso regression and an ETS-X, artificial neural network and support vector regression
model. In order to produce the forecasts, three years of historical weekly factory sales data and weekly wholesaler sales, ending
inventory, on order quantities and receipts for the same time period are used. As a benchmark, also forecasts from an ARIMA
and ETS model are produced. To evaluate the forecast accuracy, we adopt a sliding window approach and measure
out-of-sample RMSE, MAE, MAPE and MASE, so that both comparison between the methods as well as comparison across
different series is facilitated. Our results clearly indicate that forecast accuracy can effectively be improved by leveraging
downstream data from wholesalers. The extent of improvement over traditional forecasting techniques varies between the
forecast items included in the study, as well as between the various models considered. Although linear models can already
produce considerable gains in accuracy levels, nonlinear methods tend to outperform.
Original languageEnglish
Title of host publicationThird Conference on Business Analytics in Finance and Industry
PublisherFacultad de Ciencias Físicas y Matemáticas, Universidad de Chile
Number of pages1
ISBN (Electronic)0719-8981
Publication statusPublished - Jan 2018
EventBAFI 2018: Business Analytics in Finance and Industry - Santiago, Chile
Duration: 17 Jan 201819 Jan 2018
Conference number: 3


ConferenceBAFI 2018
Abbreviated titleBAFI
Internet address

Cite this