Abstract

Inventory management systems support firms in planning for an uncertain future by using demand forecasts and optimization models to make restocking decisions. Recent work on the "data-driven" newsvendor found that incorporating machine learning (ML) can improve the success of inventory management by accounting for demand-driving information. However, ML methods are infamously hard to interpret, which may hinder their acceptance. To ameliorate this, we show how to apply an interpretable attention-based architecture, the Temporal Fusion Transformer (TFT), to the data-driven newsvendor problem. Our approach replicates and extends the original TFT time series forecasting method to the inventory management domain. We evaluate our method on two real-world retail datasets, each covering 260 perishable food items, and provide domain-specific benchmarks. The computational study illustrates TFT’s interpretable predictions and their comparatively high accuracy. Our work aims to lay the groundwork for further design science research on transparency in human-AI collaboration in this domain.Especially the commonly used fully connected neural networks are infamous for being "black boxes". To ameliorate this, we show how a time-series-specific and attention-based architecture, the Temporal Fusion Transformer, can be applied to the data-driven newsvendor. We demonstrate the application on two real-world multivariate retail datasets, each containing sales data for 260 perishable food items. This computational study illustrates the resulting interpretable predictions and their comparatively high accuracy.

Download

Share

COinS