Predictive analytics in supply chains: what the crisis revealed
How the pandemic tested investments in demand forecasting and inventory management - and what to take away from it.
Over the past few years many companies have invested in demand forecasting and inventory optimisation using machine learning. The arguments were compelling: reduce excess stock, plan procurement more precisely, reduce stockouts. In normal conditions it worked.
March-April 2020 became a stress test for these systems. The results are mixed.
What did not work
Most forecasting models in supply chains are trained on historical data with a normal level of volatility. They can predict seasonal fluctuations, respond to marketing campaigns, account for holiday spikes.
They were not trained on pandemic patterns - and could not have been. When demand for some goods multiplied several times over a week while demand for others collapsed - the models moved outside their operational range. Forecasts stopped being useful.
Systems that were fully automated - automatic orders based on forecasts with no human oversight - created problems during this period. Some generated excess orders for goods with surging demand that suppliers could not fulfil. Others kept automatically cutting orders for goods whose demand was rising in non-standard ways.
What worked
Companies that used predictive analytics as a decision-support tool rather than an autopilot fared better through the crisis. The model proposes - a person decides. In normal conditions this seems inefficient. In abnormal conditions it turns out to be the right architecture.
Companies with good supply chain visibility - those who knew the inventory status of their suppliers, not just their own - managed risk better. Predictive analytics is not always needed for this: sometimes it is enough to simply see data in real time.
Rapid reprioritisation turned out to matter more than forecast accuracy. The ability to shift focus from one product category to another within a few days depended on the flexibility of operational processes, not on algorithm quality.
What this means for analytics investments
A few conclusions I consider durable:
Forecasting models are a tool, not a substitute for judgment. In normal conditions they reduce cognitive load and improve decisions. In abnormal conditions they are supplementary, and this needs to be accepted.
Automation with low oversight is dangerous in a crisis. Full automation of operational decisions reduces costs in a stable environment. But there is a price: the loss of the ability to intervene quickly when the model is wrong.
Visibility matters more than forecasting. Knowing the current state - inventory, supplier status, actual demand - is more valuable than having a beautiful three-month forecast. Investments in data visibility pay off better than investments in sophisticated forecasting algorithms.
Supplier diversity is not only risk management. Companies with diversified supply chains had more room to manoeuvre. This is a structural question that analytics can illuminate but cannot replace.
A crisis is a bad time for sweeping conclusions. But the observations made now are worth recording.