What if? Causal Machine Learning in Supply Chain Risk Management

Mateusz Wyrembek, George Baryannis, Alexandra BrintruparXiv 2024 · DOI · OpenReview

TL;DR

This paper introduces Causal Machine Learning (CML) as a powerful tool for supply chain risk management, moving beyond correlation to understand causation for better intervention planning. By applying CML to a maritime engineering case study, they demonstrate how to identify specific interventions that can effectively mitigate delivery delays, enabling "what-if" scenario planning for improved decision-making.

What Problem Does This Paper Address?

Traditional machine learning (ML) models in supply chain management excel at predicting outcomes (like delivery delays) by identifying correlations in data. However, they often fall short when it comes to recommending interventions to prevent these outcomes. This is because correlation doesn't imply causation, and managers need to understand the "why" behind a delay to implement effective preventative measures. The "black-box" nature of many ML models further exacerbates this, leading to a lack of trust and reluctance to fully integrate them into decision-making. This paper addresses this gap by proposing and demonstrating the use of Causal Machine Learning (CML) to move from prediction to actionable intervention.

Methods

The core of the paper's methodology lies in applying Causal Machine Learning (CML) techniques. Specifically, they utilize a Double/Debiased Machine Learning (DML) framework. DML combines ML with causal inference to estimate intervention effects. The DML framework, based on the Frisch-Waugh-Lovell Theorem, involves a multi-step regression process to isolate the causal effect of a treatment variable from confounding factors.

Key aspects of their approach include:

  • Causal Graph Generation: They use Directed Acyclic Graphs (DAGs) to visually represent causal relationships between variables, informed by domain knowledge and then validated/generated using automated methods like Hill Climbing, Tabu Search, and the PC algorithm. This helps in selecting relevant features and structuring the causal problem.
  • Interactive Regression Model (IRM): For their case study, they employ an IRM, a suitable ML model within the DML framework for binary treatment variables. IRM allows for the estimation of both Average Treatment Effects (ATE) and Conditional Average Treatment Effects (CATE).
  • Policy Trees: To translate CML insights into actionable strategies, they derive deterministic binary treatment policies using classification trees. These trees provide clear "if-then" rules for making intervention decisions based on observable characteristics.

Datasets

The study uses a historical dataset of orders and supplier transactions from the maritime engineering sector, spanning from 2015 to 2022. This dataset includes information on suppliers, projects, part descriptions, quantities, prices, delays, and whether a supplier serves multiple warehouses.

Key Insights and Ideas

  • Causation is Key for Intervention: The fundamental insight is that to effectively manage supply chain risks and plan interventions, understanding causal relationships is paramount, not just correlations.
  • "What-If" Scenario Planning: CML enables "what-if" scenario planning by allowing the estimation of outcomes under different hypothetical interventions. This empowers managers to proactively explore the impact of their decisions.
  • Supplier Network Complexity Matters: The study found that buyers using suppliers who serve multiple warehouses experience longer delivery delays. This suggests that a supplier's network complexity can influence their operational priorities and capacity, leading to delays.
  • Seasonality Impacts Delays: Seasonal variations in order placement significantly affect delivery delays, with some quarters increasing delays and others decreasing them. This highlights the importance of seasonal timing in supply chain planning.
  • CML Complements Traditional ML: The paper advocates for a complementary approach where traditional ML is used for accurate forecasting, and CML is used for effective intervention planning by altering causal parameters.

Key Results

  • ATE for "Multi" Warehouse Suppliers: Buyers using suppliers that serve multiple warehouses experience an average increase of approximately 17 days in delivery delay. This result was statistically significant.
  • Seasonal ATEs:
    • Buying in the second and third quarters increases delay by 8 and 10 days, respectively.
  • Buying in the first and fourth quarters decreases delay by approximately 6 and 19 days, respectively.
  • CATE Insights: Analysis of Conditional Average Treatment Effects (CATEs) revealed that for small order quantities, the effect of interventions is uncertain. For quantities between 100-200, the effect increases, and then plateaus. Seasonal CATEs showed varying trends, with the fourth quarter exhibiting less uncertainty.
  • Policy Trees: The derived policy trees provide concrete recommendations, such as:
  • For "Multi" warehouse suppliers, if the price is low, use single-warehouse suppliers; if high, consider multi-warehouse suppliers, with specific exceptions for certain suppliers.
  • Seasonal policies suggest specific supplier usage patterns for different quarters, with some suppliers recommended exclusively or to be avoided during certain periods.

Limitations and Future Work

  • Case-Study Specificity: The primary limitation is the case-study nature of the research, which limits the generalizability of findings to other sectors. Further research is needed to validate CML's effectiveness across different industrial contexts.
  • Broader SCRM Applications: The authors suggest applying CML to other Supply Chain Risk Management (SCRM) issues like inventory management, predictive maintenance, and quality control, as these areas offer ample opportunities for interventions.
  • Causal Federated Learning: A promising future direction is Causal Federated Learning, which could enable privacy-preserving collaborative data analysis across multiple self-interested agents in a supply chain, identifying systemic inefficiencies and intervention opportunities.

Summary updated

Abstract

The penultimate goal for developing machine learning models in supply chain management is to make optimal interventions. However, most machine learning models identify correlations in data rather than inferring causation, making it difficult to systematically plan for better outcomes. In this article, we propose and evaluate the use of causal machine learning for developing supply chain risk intervention models, and demonstrate its use with a case study in supply chain risk management in the maritime engineering sector. Our findings highlight that causal machine learning enhances decision-making processes by identifying changes that can be achieved under different supply chain interventions, allowing "what-if" scenario planning. We therefore propose different machine learning developmental pathways for for predicting risk, and planning for interventions to minimise risk and outline key steps for supply chain researchers to explore causal machine learning.
BibTeX
@misc{wyrembek2024what,
  author = {Mateusz Wyrembek and George Baryannis and Alexandra Brintrup},
  title = {{What if? Causal Machine Learning in Supply Chain Risk Management}},
  year = {2024},
  eprint = {2408.13556},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.13556},
}
Figure