Probabilistic Machine Learning for Reducing Root Causes of Food Waste
When most businesses talk about machine learning (ML) for S&OP processes, they are using deterministic ML. Predicting future demand. Generating an optimised distribution plan. Calculating what price will maximise gross margin. Many of these are done with deterministic ML.
A common challenge with deterministic ML is that it is a black box. Users cannot easily understand how the answer is reached and therefore can struggle to trust the outputs. Deterministic ML also removes a lot of useful information in reaching an exact answer. For example, when predicting an exact monthly demand for an item, information on how stable or not the demand for that item has been is lost. For a supply planner using the demand forecast to create resupply plans, knowing whether they should be planning for potential spikes or troughs in demand is valuable information. So how can this useful information be captured and used without leaving operations teams with information overload?
Machine learning is incredibly effective at crunching huge quantities of data to understand trends in data over time. Instead of using machine learning to predict the exact demand for a SKU next month, probabilistic ML calculates a band, or range, that the demand is likely to fall within. This is calculated based on the historical demand for the product, including the seasonality, growth/decline trend and the level of volatility in demand.
Probabilistic ML is not just for demand forecasting though. It can also be used to understand distributions of factors including order lead time, promotion impact/uplift and capacity on a particular network element/node. By combining probabilities of different scenarios across the supply chain, it is possible to get data-driven insights into the probability of stocking out, or the probability of having expiring unsold inventory, based on the current supply plan, or a hypothetical what-if scenario.
With the ability to calculate the probability of stockout or waste based on a what-if scenario, probabilistic ML can be used for a scenario planning tool to help planners understand supply risks based on different actions. This can be used to either optimise for total supply cost, or to hit specific availability targets at the lowest cost.
To hit an availability target of 95%, a supply plan will need to be able to cope with demand at the top end of the 95% probability band (see chart above). Whereas to optimise for supply cost, the probability and impact of different supply costs (waste, stockout, inventory) are calculated for the planned scenario. This gives the planner the second of the charts below to see how supply cost could be reduced by changing a scenario parameter.
As shown with the Optimum Order points on the charts above, it is a small step from probabilistic scenarios to prescribing optimum decisions. As well as prescribing the optimum order date, based on either meeting availability targets or minimising total supply costs, probabilistic scenarios can be applied across the supply chain for safety stock targets, contract quantities and more, across both raw materials and finished goods.
A recent engagement with a Top 10 global food company started with using probabilistic ML to create dynamic safety stock targets (SSTs). These SSTs automatically adjusted to changes in demand, volatility and supply reliability to optimise safety stocks, based on achieving a minimum availability target. The same probabilistic ML models were then used to predict the stockout probability of each raw material at each factory, based on the current supply plan. By combining these stockout probabilities with the daily production value and the feasibility and cost of alternative supply sources, we created a risk score for each material and factory pair.
These risk scores were used within a control tower to identify where the current supply plan was too risky. In these instances, the control tower was then used to assign an action to reduce the risk score to an acceptable level. These actions included spot purchasing more raw material, moving inventory in from another factory or altering production plans to reduce demand for that material in the short term.
The problem with deterministic demand forecasts is that there is either a risk (demand >= supply), or there is not. This makes it challenging to identify where current supply or operations plans carry a high element of risk, and almost impossible to define what an acceptable level or risk is. By leveraging probabilistic ML, we enable supply planners to identify the most important risks (high probability of stocking out and/or high impact of a stockout and a non-trivial stockout probability), weeks or even months before they would normally.
While the dynamic, probabilistic, safety stock targets are effective at reducing the frequency of availability issues, probabilistic risk tracking in a control tower slashes the cost of addressing the remaining issues, unlocking lower cost mitigation actions with earlier identification.
Purpose AI is an AI and Analytics Consultancy delivering data-driven solutions for complex Supply Chain and Commercial decision making. We work across the Consumer Products and Retail Industries and have worked with many Top 20 CPGs and Retailers in Europe and North America. Our USP is that we have developed industrialised solutions to common industry problems, reducing the time and costs to deliver value. But unlike SaaS providers, there is no licensing costs, we customise our solutions to your requirements and embed a strong business process to ensure they deliver value.