Machine learning has great potential to improve supply chains. So at my company, Kinaxis, when analysis of data from a major customer revealed that 55% of their lead times were wrong as designed, we began applying machine learning.
Lead times matter because overly optimistic planning assumptions mean supplies are expected to arrive sooner than they actually do. Waiting delays production and on-time customer delivery while building up parts that arrived on time but cannot be used until remaining parts needed arrive. Overly pessimistic planning assumptions mean actual lead times shorter than planned, so some parts arrive early, building up inventory and storage costs, while others are still in transit. If demand is slower than expected, parts accrue in inventory, unused due to obsolete needs.
More accurate planned lead times allow on-time customer orders, minimize inventory, and reduce buffer stocks necessary to ensure production. Predicting lead times is a problem well-suited to machine learning and automation. The planner sets tolerances for variations in lead times, which we use to configure processing rules for what actions to take. Our machine learning models use historical data to predict actual lead times, compare them to designed lead times, and then use the processing rules to improve decisions, leading to more realistic results.
We have taken a similar approach to predicting yield times. The results from these projects can be significant – for one company we were able to save $17 million in late revenues for their North American region over their 6 month planning horizon.
Minor deviations not worth the time to analyze but deemed worthy of a change are automatically accepted by the model, thereby “self-healing” the deviation. Those with a significant enough impact are flagged for manual review. Minor deviations with minimal impact are simply ignored by the processing rules. Planners can focus on decisions that matter most and let math automatically handle those that do not.
Here is a link to a longer version of the article I published in Analytics.