42nd International Symposium On Forecasting

The International Symposium on Forecasting (ISF) is the leading conference in the field of forecasting. World-leading forecasting researchers, practitioners and students come together to use the combination of lectures, academic sessions, workshops and social programmes for networking, learning and exchange.

Among the speakers are experts from the Center for Applied Research on Supply Chain Services:

Benedikt Sonnleitner: "The value of hierarchically aligned forecasts for staff scheduling"

Hierarchical forecasting provides value for decision-making by improving accuracy and guaranteeing aligned forecasts (i.e., the sum-constraints in the hierarchy are fulfilled). We examine temporal hierarchical forecasting for staff scheduling in cross-docking and offer a threefold contribution: (i) We quantify the value of aligned forecasts for decision-makers on multiple planning levels. (ii) We investigate the value of accuracy improvements for decision-making. (iii) We offer a new prescriptive solution, by integrating freight demand forecasting with staff scheduling for cross-docking. The analysis is performed using real data from a German carrier.

Mo, 11.07.22 | 3:10-3:30 pm | As part of the session on Hierarchical Forecasting

Claudia Ehrig: "The benefit of clustering for cross-learning models on M5 data"

Cross-learning can enhance prediction accuracy and save computational resources in comparison with series-by-series learning. Its benefit is however depending on the similarity of the time series used. Hence, we propose to cluster the data beforehand and then cross-learn one model per cluster. We benchmark different clustering methods with respect to the forecasting accuracy, runtime, and CO2 emissions of a simple MLP trained on those clusters. We use data from the M5 competition.

Mo, 11.07.22 | 4:10-4:30 pm | As part of the session on Machine Learning

Nico Beck: "Benchmarking Historical Consistent Neural Networks"

Historical Neural Networks are a Neural Network class, specifically designed for macroeconomic forecasting. Hans Georg Zimmermann presented those at several workshops at the ISF. Although they have achieved remarkable results in industrial practice (e.g., copper price forecasting), a rigorous benchmark study is currently missing. To address this research gap, we evaluate under which conditions (i.e., forecast horizon, data properties) the model is superior to state of the art methods. We use real world and synthetic data sets.

Di, 12.07.22 | 3:30-3:50 pm | As part of the session on Macroeconomic Forecasting

Julia Schemm: "Using Uncertainty Estimation in Demand Forecasting for Optimizing the Inventory Planning of a German Wholesaler"

Accurate demand forecasts are a major advantage for wholesalers across the board. However, point forecasts are not enough to make optimal ordering decisions because they ignore forecast uncertainty. As most wholesalers try to avoid stock-outs in order to not lose revenue and customers, defining a reorder level or safety stock is crucial. On the other hand, e.g. storage capacity and capital tie-up can be restricting factors. Reliable dynamic uncertainty estimation in demand forecasting can both help prevent stock-outs and at the same time reduce unnecessarily large safety stocks. We apply multiple statistical, machine learning and deep learning Time Series Models (both univariate and multivariate) for forecasting the demand of a German wholesaler in the area of HVAC and plumbing on an article level. Our training and evaluation procedure includes hyper-parameter optimization on a validation set. In addition, we implemented multiple methods from literature for estimating the forecast uncertainty. We evaluate the accuracy of both the point forecast and the uncertainty estimation on historical data of product sales. For deriving the optimal purchasing decision with respect to service level and restrictions like storage capacity, forecasting results serve as input for a mathematical optimization model.

Mi, 13.07.22 | 2:20 – 2:40 pm | As part of the session on Supply Chain