991. Automatic Model selection for improving the forecasting accuracy in Manufacturing Demand
Invited abstract in session WB-47: Empirically Driven OR in Retail, stream Retail Operations.
Wednesday, 10:30-12:00Room: Parkinson B08
Authors (first author is the speaker)
| 1. | Giacomo Gaggero
|
| Department of Economics, University of Genoa | |
| 2. | Pier Giuseppe Giribone
|
| Department of Economics, University of Genoa |
Abstract
This study explores a rigorous approach to forecasting the future sales volumes for one of the most
important Italian companies in the automotive components sector. The methodology investigated in
this research is based on analyzing the behavior of historical sales data for over 250 different
products, leveraging the classification method proposed by Syntetos and Boylan. This method
relies on the Average Demand Interval and the Coefficient of Variation to categorize time series
into four classes: lumpy, erratic, intermittent, and smooth.
Once the products are classified, the most suitable forecasting model is applied to each time
series. For highly intermittent demand patterns—characterized by frequent zeros and unexpected
peaks (lumpy)—the Teunter, Syntetos, and Babai (TSB) model is employed, which is a refined
version of the well-known Croston method designed specifically for intermittent demand
forecasting. For the other categories (erratic, intermittent, and smooth), the SARIMA model is used
when the time series is stationary, while a recurrent neural network (RNN) is applied if the series is
non-stationary.
In this way, we do not apply a generic forecasting method to all products indiscriminately, but rather select the most suitable model for each product, thereby achieving more
reliable demand predictions.
Keywords
- Forecasting
- Manufacturing
- Inventory
Status: accepted
Back to the list of papers