Supply chain network and operations planning teams find that despite innate knowledge of their product portfolio, driving efficiency in these operations through digital transformation is highly complex given global distributed networks. The worldwide pandemic has further disrupted institutionalized world models, altered fundamental assumptions, and exacerbated decision making uncertainty. As a data scientist in the consumer products supply chain space, I found myself at the same impasse, asking questions such as:

  • Do we have sufficient representation of exogenous signals to capture the changing landscape?
  • How do we support decision making while establishing a reasonable degree of confidence?

The guiding principles that helped us work through this gridlock is represented below as an idea maze:

data science idea maze

Yes, everyday decisions must still be made, however they don’t have to be made with limited knowledge. To solve this, we focus on:

  • Real-time intelligence and visibility into the current state of the portfolio across the network.
  • A sense check for assumptions (institutional/human models) about the current state of business as various planning and portfolio decisions are being made in real-time.

Where are we? (Sense Check)

An emergent view of supply chain portfolio

Let me talk about how we put this idea maze into practice in one customer situation which unfolded during the 2020 pandemic. Much like one would look at their stock portfolio every day, we felt it was important to equip teams with a real-time view of their product portfolio across the network. During the early phase of the pandemic many planning teams had the impression that demand was unprecedented across the product spectrum and assumed that anything produced would be shipped. However, careful examination revealed a plethora of behaviors and indicated a quick return to baseline demand levels or below.

How and why did this change? (Recognizing the changing landscape)

Redrawing the supply chain map and recognizing portfolio migration

When we looked at the product portfolio distribution in terms of demand regimes for consumer staples, it became clear that most of the portfolio returned to baseline or below baseline and the sustained consumption that planners were expecting did not actually occur.

trend classification

The graph above further emphasizes the need to carefully manage the “precious few” vs. “insignificant many” in terms of managing service levels, inventory holding across the portfolio, taking appropriate actions and the need to manage the long tail of products using AI. It has become increasingly challenging for planning teams to scale so rapidly and make sense of all the shifts, when they are already resource — stretched in these times of unprecedented volatility.

When considering rapid behavioral shifts in consumer demand during this 2020 black swan event, several analogues came to mind. One in particular that has similarities to what we were seeing in consumer products is the migration patterns in American immigration inflows due to displacement, access to education, economic boom, etc., represented below:

immigration inflows by country

We applied a similar perspective to products (SKUs) in a consumer staples portfolio – i.e. consider consumer behavior patterns to be synonymous with countries and demand/consumption patterns similar to population outflows and inflows to thereby understand shifts/waves in behavior. We used various data science methods in order to learn distinct archetypes in demand behaviors (see below) and developed a consumer pattern alphabet that encapsulates factors such as time-series volatility, underlying network factors such as demand supply network imbalance, etc.

These pattern archetypes have helped us quantify overall portfolio shifts where we see that even established products – i.e. the A and B categories that typically constitute 80% of demand are growing in the most volatile way and the C’s and D’s, which are typically lower volumes, are becoming increasingly important due to regime shift to dot.com and eCommerce buying patterns.

Demand pattern archetypes

demand pattern archetypes

While demand patterns driven by external consumer behavior form the heartbeat of a supply chain, how well the internal supply network adapts and synchronizes to these demand dynamics is what managing risk is all about. Looking purely through the demand lens, one does not necessarily see the key differences, so we furthered the immigration analog and applied our tools to capture downstream signals such as inventory levels, days of supply, lead-times (production and transportation), etc.

In this case, clusters (akin to countries) capture specific regimes in the portfolio in terms of demand supply divergence. This helped us partition the portfolio into classes that captured various behaviors on the basis of demand and supply parameters. Depicted below are two classes that are very relevant operationally as they represent dramatically different risks.

how are the relationships developed?Additionally, we were really interested in the time evolution of the above types of classes/clusters to understand the portfolio level risk visually. It becomes very clear one can observe significantly different patterns in products within categories (e.g. Home care category would include products like: dishwasher pods, home cleaning products, laundry detergents, etc.) and these pattern shifts across categories (e.g. Home Care, Fabric Care, Family Care, Personal Care) are so dramatically different. For the two sample classes (represented as class yellow and red) above, the time series view shows the portfolio suddenly moved into high demand risk regimes right around the peak of the pandemic signaled by a sudden increase in the volume of the “orange” class and a rapid reduction in volumes in the “red” class which represented lower supply risk and clues planning teams into signals to watch-out for.

trend classification
Product behavior across categories (e.g. Home Care, Fabric Care, Family Care, Personal Care) during the early stage of the pandemic based on demand supply imbalance measured via multi-variate signals across the network.

What does it all mean?

This work on portfolio-segmentation enabled us to be nimbler at detecting of demand-supply imbalance, as we are able to telescope in and out of various hierarchies and time resolution. In fact, some of the same methods we’re using to assess portfolio changes over time, also help us provide more robust insights into impending demand-supply risks.

The ultimate goal of Noodle.ai’s Athena Insights application is to shine a light on cause and effects and more importantly, drive action to mitigate risks. Demand-supply imbalance primarily drives risk across the supply chain. The above approaches to understanding portfolio shift and evolution enabled us to be much more effective at detecting divergence in demand and supply as pre-signals to impending risks and recommend appropriate measures.

We noticed that unsupervised models using multi-variate time-series uncovered dramatic shifts in behavior that high-level overall views based on a single dimension missed. Data also shows us very different views when looking at absolute numbers versus percent of population as seen in the immigration inflows visualization above. The power of visualization combined with data science cannot be emphasized enough in terms of serving as an intuitive sense-check for planning & operational teams’ understanding of what is truly unfolding on the ground.

Resources:

https://covid19.healthdata.org/united-states-of-america
https://www.bsg.ox.ac.uk/research/research-projects/coronavirus-government-response-tracker
https://oui.doleta.gov/unemploy/claims.asp