The COVID-19 crisis has significantly disrupted the normal demand patterns which dictate supply chain decisions -- across distribution, manufacturing, and raw material procurement. Since these disruptions are happening so rapidly, and often in nuanced and unexpected ways, it is a challenge to determine how much to produce, of what, when, and where it needs to go. Within a single product line, you may quickly see demand for certain SKUs spike up, while others decline steeply -- and this pattern can vary by channel, segment, and region.

The stakes for getting this calibration right are quite high: companies will need to recapture market share for their products and services as demand recovers, while minimizing overproduction and excess operating expenses. They also need to be able to quickly spot new emerging opportunities in the market. In some areas, they’ll need to right-size their operations to match the post-crisis “new normal” steady-state. The earlier they can get insight into what this might look like, the better prepared they can be.

Regarding supply chain forecasting and analytics, two key challenges with existing methods are being addressed with next-generation advanced analytics techniques.

1. First, forecasting models based on pre-crisis conditions are no longer relevant or reliable. Companies are doing their best to quickly calibrate what the next few weeks and months will look like, but they can’t rely on the well-established techniques that suit them well in a steady-state environment.

Given this, it is key to be able to discover new predictive patterns within the latest data refreshes, train useful models, and get models into use quickly. This is especially true given that the spread and impact of COVID-19 is happening in waves, staggered over time and space, such that learnings from early onset areas can provide foresight into the broader and more widespread impact. But in order to build accurate models in this context, many granular geospatial and temporal factors must be accounted for. Identifying a robust set of such factors is typically a complex, time-consuming, and incomplete data science effort -- given the time constraints, dimensionality limitations, and biases of human data analysts. But new offerings are helping data scientists build useful forecast models quickly, based on learnings from early impacted areas, enabling them to quickly incorporate and discover useful predictive signal in external datasets, including data on COVID-19 spread at a US county-level.

Companies that can learn from the unfolding situation to optimize their supply chains will have a competitive advantage and be able to optimally calibrate their actions as economic activity revives.

2. Second, typical reporting and business intelligence methods are revealing their blind spots. They provide an aggregated view of the business which overlooks much of the underlying, and often highly informative, insights which are only visible when one, two, or three levels of additional depth of factors are considered. Take a simple example of a B2B distribution company: in a recent month, within a particular customer segment and region, demand for some SKUs increased dramatically while other SKUs plummeted. Standard reports showed no change in demand on aggregate, overshadowing the extreme SKU-level volatility in that segment-region pocket. The company missed this early alert, catching it only in the next cycle of reporting once it had become a major issue, and was weeks behind in making supply chain changes upstream to get demand where it needed to be. Their standard static reports didn’t provide a sufficient level of dimensionality and nuance. And the reports only showed what had been explicitly asked of the data based on past experience; but given this was a novel pattern, which popped up early in a granular pocket, they missed it until it was too late. This is just one example in which AI-powered trend detection would prove useful.

Looking Ahead

Mike Sterling is Director of Impact Management in the Americas, SparkBeyond.