Applying Causal Inference to Inventory Management: Demand Forecasting and Strategy Implementation

The presentation introduces the application of causal inference on sequential data to warehouse management, outlining four main parts: an abstract view of inventory management, a review of related technical concepts, how causal inference assists demand forecasting, and how it supports strategy implementation.

1. Abstract Understanding of Inventory Management – The goal of supply‑chain management is to deliver high‑value service while controlling costs. Accurate demand prediction based on extensive historical data and real‑time inventory observations is crucial for optimal replenishment, especially when data come from multiple scenarios or products.

2. Review of Related Technical Concepts – Time‑series data exhibit temporal dependencies; common assumptions include weak stationarity and autocorrelation. Causal analysis is framed within Structural Causal Models (SCM), enabling the extraction of causal graphs, intervention, and counterfactual reasoning. Methods for learning such graphs include constraint‑based, score‑based, model‑specific, and Granger‑causality approaches.

3. Causal Inference for Demand Forecasting – By learning dynamic Bayesian networks, multi‑step forecasts can be generated with interpretable features. Techniques such as ARIMA, VAR, VAR‑LiNGAM, PCMCI, and deep learning models (RNN, LSTM) are compared; causal‑enhanced methods reduce spurious edges and improve coefficient accuracy, leading to better predictive performance.

4. Causal Inference for Strategy Implementation – Inventory control is modeled as a Markov Decision Process. A generative adversarial causal network augments observed data with counterfactual samples, enabling efficient policy learning even under environment shifts (e.g., varying equipment lengths). Online replanning reduces computation time compared with offline methods, adapting quickly to demand or supply changes.

The integrated platform provides a web API that automates data ingestion, statistical testing, graph learning, and downstream forecasting or attribution, demonstrating strong performance on real‑world datasets such as energy demand, price promotion, and order forecasting. The solution has already been deployed internally and is expanding to frontline business scenarios.

Conclusion – Causal inference offers interpretable, accurate, and adaptable tools for both demand prediction and replenishment policy design, helping enterprises achieve sustainable profitability in inventory management.