[PAST EVENT] [Via Zoom] Xiaodan Zhu, Computer Science - oral exam
In supply chain planning, manufacturers need to obtain accurate demand forecasts of multiple products based on limited historical records. The problem is about sparse high-dimensional time series forecasting, which is common in many fields. However, due to the unique characteristics of the pharmaceutical (pharma) supply chain, this problem is even more critique and challenging for pharma demand forecasting. Pharma demand is affected by many factors that require a large amount of data and more sophisticated models to capture. However, a significant portion of the drug demand records is zero. Besides, due to changes in market situations, distant historical data may not be of much help in recent forecasts. Therefore, limited data prevent us from pursuing more advanced models. According to our experiments, the temporal variables of different product demands have a weak dependency, which makes standard multivariate time series models fail to work in our settings.
To overcome the problems in demand forecasting, we develop a cross-series machine learning framework that uses different time series to train a machine learning model jointly. The use of cross-series training can increase the sample size. Still, at the same time, the learning of global time series may affect the forecasting accuracy of an individual time series. To balance the tradeoff between sample size and sample quality, we build separate models for subgroups of time series based on different grouping schemes. In addition, we introduce non-demand features with substantial contemporaneous impacts. Specifically, we develop a novel demand forecasting framework that leverages machine learning algorithms (known for detecting patterns) and makes use of historical demand and downstream inventory data across many different pharma products, as well as information of supply chain structure and relevant domain knowledge to improve the overall forecasting accuracy.
We test the benefit of our proposed framework with various modeling possibilities on two large datasets from major pharma manufacturers. With numerous design considerations, we conduct extensive numerical experiments for different machine learning algorithms, different grouping schemes, different levels of data aggregation based on supply chain structure, different time lags of historical demand as well as inventory data. The results not only validate the superiority of our framework but also provide significant empirical value and insights of downstream inventory information in the context of demand forecasting. Based on the robust performance, our framework can be applied to other industries and provides practical guidelines in reality.
Xiaodan Zhu has been working on his Ph.D. degree in the Department of Computer Science at William & Mary since Fall 2015. He is working with Dr. Zhenming Liu in the fields of time series prediction and trajectory inference. Xiaodan Zhu received his M.E. degree from the University of Electronic Science and Technology of China in 2015, a B.E. degree from Jiangnan University in 2012.