Supply Chain Root Cause Analysis with Causal AI

As the last year has shown, disturbances to supply chains impact economies heavily through delays in delivery and production and rising prices. Understanding the reasons for supply chain disruptions is therefore critically important. This process of identifying causes for disruptions is known as root cause analysis. In this blog post we will focus on the identification of systematic issues in supply chains through the lens of causality.

The State of Supply Chain Today

Many companies are now facing supply-chain issues due to the many geopolitical events which are changing the landscape of how goods move across the globe. Take as an example a large auto manufacturer — they were first heavily affected by sanctions on China in 2019, they then had to completely redesign their distribution channels after COVID, and are now once again scrambling to respond to the conflict in Ukraine.

For companies like these, there’s a clear challenge: how to learn lessons from the past to prepare for similar scenarios in a data-driven way in the future. Which suppliers should we order from? Where should we establish our next manufacturing plants? How much lead time should we expect from units that compose our final product?

Ultimately, this learning needs to be channeled into guaranteeing “on time, in full” delivery , which is a highly material problem with big ROI. Decisions made today will only materialize as problematic months from now, and by that time it will already be too expensive to fix them. This is why finding the root cause of problems — diagnosing what’s to blame — is so material. Identifying these root causes of problems allows us to fix the future, making decisions to minimize the likelihood that issues will appear again.

Why Causality?

Root cause analysis is traditionally carried out with a mix of subject matter expertise and data analysis. An analyst leverages classical tools (exploratory and statistical analysis) to crunch the data and tries to figure out where the supply chain’s main vulnerabilities lie using his or her knowledge of the process and the issues that were encountered in the past. As the supply chain at hand grows in complexity, the number of potential root causes grows as well and the problems become increasingly hard for the usual data crunching techniques and for human expertise to capture.

Moreover, traditional statistical methods are prone to picking up on spurious correlations. One might find many correlations to an order being delayed — for instance, a specific manufacturing plant may typically be involved in high numbers of delays, the minimum supply of a given item, the time it takes for an item to be ready for shipping (see Figure 1). Finding these correlations to delay doesn’t necessarily paint a faithful picture, and more importantly will not help in making the right decisions to guide investment and prevent disruptions in the future.

Figure 1: All of these variables are correlated, but there are clear causal paths in the data. Most importantly, we can separate true direct causal drivers (staff absences and road accidents) from completely spurious correlations. For instance, the more we hear “All I Want for Christmas is You” on the radio, the more likely we are to experience delays — but is there really a causal relationship there?

Using causal discovery and inference algorithms, one can comb the data to find the true underlying causal structure. With a causal model in hand, we can more easily identify vulnerabilities, be more confident as to where to allocate resources, and better understand their effect.

Causal Discovery for Supply Chain

The most important component for answering root cause questions is a causal graph. A causal graph captures the causal relationships between the different variables in our problem — what influences what and in what direction.

The complexity underlying the supply chain process is characterized by a large number of data attributes for a given order, including the quantity of the item being ordered, its value, the plant that manufactures it, all events logged during the lifetime of the order, the supply levels for the item, the agreed delivery date, the order route, and many further factors.

Causal discovery allows us to identify how these variables are interrelated and ultimately how they relate to the order being delayed or not. Causal discovery from observational data is a process that takes in raw data and human expertise and discovers the underlying causal graph — mapping all cause-and-effect relationships between all the variables in the system.

Specifically, in supply chains, we know that some causal relationships can only exist in a certain direction. For example, the static characteristics of an item can influence the demand for that item but not the other way around, and similarly whether or not an order is delayed can’t backwards-influence any of the upstream variables in the supply chain. Incorporating this knowledge into the causal discovery process allows us to both significantly speed up the graph discovery process, as well as get more explainable results that are aligned with our domain knowledge and understanding.

Figure 2: Causal graph discovery from raw data through extracted features and domain knowledge

causaLab, a module of the causaLens Decision-Making Platform, allows users to streamline the causal graph discovery process. causaLab provides ready access to optimized algorithms for causal discovery, and enables domain experts to input their knowledge and evaluate different causal graphs.

Figure 3: causaLab supports a variety of different methods to discover causal graphs. Data scientists are empowered to run a variety of different experiments, testing different hypotheses, assumptions, and combinations of parameters.

Understanding the Past, Shaping the Future

Having a causal graph can help us go beyond finding and analyzing root causes. The causaLens Platform can extract a Structural Causal Model (SCM) from the causal graph and go beyond just identifying which variables interact with each other but also how they interact functionally.

Different pieces of information come at different times: suppose we just requested a material from a supplier and the ETA is 1 month — if tomorrow we’re told that the destination port is closed, the ETA will be automatically updated in light of this new event. Forecasts are online and event-driven: they update whenever there’s new information available.

Figure 4: Predicting whether an order will be delayed at different stages using the causal graph and underlying SCM.

A natural question that arises when an order turns out to be delayed is “what could we have done to prevent it?”, that can be answered when we have an SCM using counterfactual explanations. Counterfactual explanation finds the best node(s) within the causal graph to nudge in order to reach a target value. This is done by iteratively changing the SCM and evaluating the effect of the intervention throughout the causal graph.Armed with that knowledge, we can use algorithmic recourse — action optimization — to discover the best action plan to reduce the probability of delay in future orders. This process takes into account all possible environmental variables, as well as costs for levers that we can pull — in order to find the plan of interventions with the biggest ROI to prevent disruptions from arising.

Ordering a helicopter is a great way to ensure a fast delivery process, but is it worth the investment?

Figure 5: Examples of counterfactual recourse for specific orders. On the left, the identified counterfactual indicates that increasing the minimum supply in the plant by 10% would have led to the order not being delayed. On the right counterfactual recourse indicates that reducing the number of sub-orders bunched into the order would have led to the order not being delayed.

Once again, causality is fundamental for this problem: every action has costs, and many unplanned consequences. To truly figure out the best action, given the objective of reducing costs or increasing revenues, we need to accurately measure the impact of our action on every single outcome. Focusing on a single outcome and a single KPI leaves us blind to the holistic consequences of our decisions. This is only possible if we have a full causal graph describing the data.

Tying it all together

Our Platform’s decisionOS module enables us to build highly customizable applications and turn the previously explained steps into an interactive application. Business stakeholders can inspect individual orders to understand why they are delayed, as well as ask big-picture questions about what is affecting their overall KPIs, and ultimately get recommendations to minimize the likelihood of future delays.

Figure 6: A dynamic, interactive, enterprise-ready web app powered by Causal AI and built with decisionOS (a decisionApp!) for diagnosing and remedying root causes of delays in the supply chain.

Our case study has focussed on supply chain management, but understanding the root causes of outcomes is fundamental in many business problems and sectors. Set up a call with one of our consultants if you’d like to learn more about how Causal AI can add value to your business.