What If You Knew a Delivery Would Be Late When You Placed the Order?
On the shop floor, a material shortage rarely stays contained. One late part forces a rescheduled production plan; pulling other jobs forward to keep the line running adds changeovers; and when a customer due date is finally at risk, expedited air or courier freight enters the picture. The root cause is a single component, but the cost spreads across production, logistics, and sales.
Most companies track on-time delivery rates by supplier, but usually as a month-end or quarter-end look back. That number is useful for supplier scorecards, yet it says nothing about whether the order you just released will arrive late. What planners actually need is an answer to a different question: which of today's open orders are at risk?
The Data Behind the Prediction
ERP Purchase Order and Receipt History
The starting point is data already sitting in your ERP.
Engineered Features
Raw fields alone have limited predictive power, so business knowledge has to be turned into features.
Data Integrity Issues
This is where most of the effort goes. If one order arrives in three partial shipments, you need to decide when it counts as received. A common approach is to agree in advance on a rule such as the date most of the required quantity arrived. When due dates are changed mid-stream, keep both the original promise and the revised date; otherwise you cannot separate a supplier's real reliability from a buyer's schedule adjustments.
Model Design and Evaluation
Classification or Regression?
A classification model ("will it be late, yes or no?") is well suited to building a risk list, while a regression model ("how many days late?") is better for feeding production planning. A practical pattern is to start with classification to flag risky orders, then predict a delay band within them (up to 3 days, 4–7 days, 8+ days).
Tree-Based Models Versus a Baseline
For tabular data, tree-based models such as gradient boosting are usually a solid starting point. But they must always be compared against a baseline. The simplest one is "flag any order placed with a supplier that has a high historical late rate." If a model cannot beat that, there is no reason to run something more complex.
Evaluating in Business Terms
Accuracy alone won't tell you much. Consider a plant that issues 1,200 orders a month, 96 of which (8%) arrive late. Suppose you review the top 10% by risk, or 120 orders. If the baseline catches 40 late orders, recall is about 42% and precision about 33%. If the model catches 60 within the same 120, recall rises to 62.5% and precision to 50%. Add how many days before the due date the warning came, and you can judge whether there was actually time to act.
Turning Predictions into Action
A Risk List and a Review Workflow
Deliver predictions as a daily at-risk order list that buyers check each morning. Show the main drivers alongside each risk score, such as "this supplier's late rate is rising" or "rush order," so buyers have a reason to trust the output.
Responses Matched to Risk Level
Feeding Outcomes Back
Record what action the buyer took and what happened next. Orders where an early check prevented a delay must be distinguishable from orders that slipped with no intervention; otherwise the next training cycle will learn a distorted picture.
Where to Start and How POLYGLOTSOFT Helps
You don't need to cover every item on day one. A realistic pilot focuses on high-impact items and the top 20–30 suppliers, running for three months to measure performance against the baseline. Once the value is clear, expand the scope and connect the results to production planning.
POLYGLOTSOFT builds the full chain: cleaning ERP order and receipt data, developing the prediction model, delivering an at-risk order dashboard, and setting up the buyer review workflow. Drawing on our MES and ERP integration experience, we design the system so predictions lead to action on the floor instead of ending up in a report. If recurring late deliveries are costing you, get in touch with POLYGLOTSOFT.