How to Predict Late Deliveries Before They Happen

You predict late deliveries by watching how a supplier behaves before the delivery date, not after it. Late or changed confirmations, slipping dates, slower replies and missing shipping notices appear days to weeks before a delivery fails. Measured per supplier and compared with its own history, and with how it behaves toward other customers, these signals give an early warning your on-time delivery report never can.

Why is on-time delivery a lagging indicator?

Every purchasing team measures on-time-in-full. Most review it monthly, per supplier, and discuss the worst performers in the quarterly supplier meeting. It is a useful number for that conversation. It is useless for the one that matters more: which of this month's open orders is going to be late.

OTIF records what happened. By the time a line counts as late, the goods are not there, the production planner has already found out, and the options are the expensive ones: expedited freight, a broker, or a rescheduled build. One machine builder we interviewed put a figure on it: a study of their shortages came to roughly 50,000 to 60,000 euro a year in direct cost, plus around ten air freights, and that in a year when demand was low enough that safety stock absorbed most of the damage. The same supplier performance in a peak year would have meant daily shortage calls.

Their own words describe the problem better than any metric: "we look at the report and suddenly see that a supplier is doing really badly." Suddenly is the issue. The supplier did not become unreliable on the day the report was run. The deterioration started weeks earlier, in a place the report does not look.

That place is the order flow itself: the confirmations, changes, replies and shipping notices that pass between buyer and supplier every day. They are the supplier's behaviour, recorded in real time, and behaviour changes before results do.

What are the leading indicators of a late delivery?

A supplier under strain changes its behaviour in small, measurable ways before it misses a date. Seven signals carry most of the information. Each is only meaningful against the supplier's own baseline: a supplier that always confirms on day three is not slipping when it confirms on day three.

Two features of this table matter. First, none of the signals requires new data. All of them sit in the order flow a buyer already runs, in the ERP, in the portal or in the inbox. The cost is in capturing them consistently, which is what structured channels and AI intake are for. Second, the signals compound. One late confirmation is noise. A late confirmation, a pushed date and a slow reply on the same supplier within two weeks is a pattern, and the pattern is what predicts.

In our interviews, buyers recognised every signal in the table from experience. What they did not have was the measurement. The experienced buyer who "just knows" that a supplier is in trouble is reading these signals by hand, across a few dozen suppliers. The method breaks at a few hundred, and it retires with the buyer.

The black box between confirmation and receipt

There is a second kind of late delivery the signals above do not catch: the order that was confirmed on time and then does not arrive. For suppliers in China, India or Turkey, and even for Italian or Spanish suppliers shipping by road, the gap between "confirmed" and "received" is weeks, and most buyers see nothing in it. As one supply chain manager told us: "the data is probably somewhere, but we can't see it. We see it when it's here."

This is a transit problem, and it has its own leading indicators:

• Should be in transit, is not. If the confirmed date minus the normal transit time has passed and no dispatch notice or tracking event exists, the goods have not left. That is knowable three to six weeks before the delivery date on a sea shipment.

• Departed late. A dispatch notice later than the confirmed date minus transit time means the delivery date is already lost, before the goods have reached a port.

• Container or lane disruption. Port congestion, a vessel delay or a road closure on a lane you use affects every order on it. These are public signals, and matching them to open orders by route is a data exercise, not a forecasting one.

The practical fix is to ask for two more events per order from distant suppliers: dispatched and, where available, a tracking reference. Many suppliers already send these by email; they are just not captured. Once they are, the "black box" becomes two checkpoints, and a missing checkpoint is itself an alert.

What a network sees that one buyer cannot

Every signal so far comes from your own orders with a supplier. That has a limit: you only see the supplier's behaviour toward you, and you only see it when you order. A supplier you buy from four times a year gives you four data points. A supplier whose factory is in trouble will show it to all of its customers at once, and you will be the last to know if you happen not to be ordering that month.

This is where a shared order network changes the picture. Tradecloud carries the order flow between around 30,000 connected companies in European industry. When a supplier starts confirming later, pushing dates or going quiet, the pattern appears across its customers before it appears clearly in any single one of them. Anonymised and aggregated, that is an early warning that no single-company tool can produce: not because the tool is worse, but because the data does not exist inside one company.

Three things a network signal adds to your own:

1. Earlier detection. A supplier slipping for several customers at once is a supplier-level problem, visible weeks before your own next order would reveal it.

2. A baseline you did not have to build. How does a supplier's response time compare with similar suppliers? Is a three-day confirmation normal for this category or slow? Your own history cannot answer that; a network can.

3. Coverage of suppliers you rarely use. The long tail of occasional suppliers, where your own data is thinnest, is exactly where network data is most useful.

The condition is consent. Network signals only work if customers agree to contribute their anonymised order behaviour in exchange for the shared warning, and if the operator is explicit about what is shared and what is not. Buyers we have asked are willing, on two terms: nothing traceable to their company leaves the network, and they get earlier, better signals in return. That is the deal that turns one company's order history into protection for all of them.

From signal to action: what to do with a prediction

A prediction is only worth having if it changes what someone does today. Three design choices make the difference between an alert buyers act on and one they learn to ignore.

Score the order line, not the supplier. A supplier risk score is interesting; a line-level probability of late delivery is actionable, because it attaches to a part, a date and a production order. Combine the supplier's behavioural signals with the line's own facts: lead time remaining, stock cover, criticality of the item, and whether a dispatch notice is overdue. The output is a short list: these twelve lines, from these four suppliers, are likely to be late, and these three of them will cause a shortage.

Show the evidence and the confidence. A buyer will not expedite a shipment because a model said so. They will if they can see that the supplier confirmed six days late, pushed the date once already and has not answered the chaser, and that two other customers in the network saw the same pattern this month. State how confident the prediction is, and what it is based on. In procurement, a confident prediction with no evidence is worse than no prediction: it either gets ignored or it triggers an expensive action for nothing.

Attach a next step. Each predicted miss should come with the action that fits its lead time: chase now with a specific question, ask for a partial delivery, pull a safety stock release forward, or start sourcing an alternative. The earlier the signal, the cheaper the action. A pushed confirmation six weeks out is a phone call. The same information at goods receipt is air freight.

Automation fits in here. The chasing, the request for a dispatch notice, the partial-delivery question can all be sent by an agent as soon as a signal fires, with the buyer deciding only on the cases where the answer is unsatisfactory. That turns prediction into prevention without adding to the buyer's day.

How to start with your own data

You do not need a network or a model to begin. Four steps, with data you already have, show whether leading indicators work for your supply base.

1. Backtest one quarter. Take every order line that was delivered late last quarter. For each, look at what the supplier did before the due date: when it confirmed, whether it changed the date, how it answered chasers, whether a dispatch notice came. Count how many late lines showed at least one signal in the table above. In our experience the share is high, which is the whole argument.

2. Build a baseline per supplier. Median confirmation delay, median confirmed-versus-requested gap, median response time. Thirty orders is enough for a usable baseline; fewer than ten and you will need network data to fill in.

3. Flag deviations from the baseline, not absolute values. A supplier confirming on day five is not late if its baseline is day five. A supplier moving from day two to day five is the one to watch.

4. Run it alongside the buyers for a month. Each week, compare the flagged lines with what the buyers were already worried about. Expect two outcomes: lines the buyers had already spotted (the model is right but not yet useful) and lines they had not (this is where the value is). Track how many of the second group became late.

The prerequisite for all four steps is capturing the order flow consistently, including the confirmations that arrive by email and the dispatch notices that currently go unread. Without that, the signals exist but the data does not. That is why structured supplier channels and an intake that reads email are the foundation for prediction, not a separate project.

If you want to see what this looks like on your own orders, the backtest in step one takes an afternoon with an order export, and it answers the only question that matters: could we have seen it coming?

Frequently asked questions

What is a leading indicator of supplier delivery performance? A measurable change in supplier behaviour that appears before a delivery fails: later confirmations than usual, confirmed dates later than requested, date changes after confirmation, unsolicited split deliveries, slower replies to chasers, and missing dispatch notices. Each is read against the supplier's own baseline.

Why is OTIF not enough to manage supplier risk? On-time-in-full measures what already happened. By the time a line is late, the planner knows and the cheap options are gone. OTIF is useful for supplier reviews, not for deciding which open orders to act on this week.

How far in advance can late deliveries be predicted? It depends on the signal. Confirmation delays and pushed dates typically show two to eight weeks ahead; a missing dispatch notice shows days ahead. Signals on several items from one supplier at once point to a supplier-level problem and tend to appear earliest.

Can we predict late deliveries without new software? You can backtest with an order export: take last quarter's late lines and check which signals preceded them. Running it continuously needs the order flow captured consistently, including email confirmations and dispatch notices, which is where tooling comes in.

What does network data add to our own supplier data? Earlier detection, because a supplier in trouble shows it to all its customers at once; a baseline for what is normal in a supplier category; and coverage of suppliers you order from rarely. It depends on customers consenting to contribute anonymised order behaviour.

What should a buyer do with a prediction? Act according to the lead time: chase with a specific question, ask for a partial delivery, release safety stock, or start sourcing an alternative. The earlier the signal, the cheaper the action. Routine chasing can be automated so buyers decide only on the unsatisfactory answers.

About the author

Tonnis de Boer is CEO of Tradecloud, the supply chain platform connecting around 30,000 industrial companies across Europe, and leads Syracus AI, Tradecloud's AI product line. He has spent over a decade on the operational side of industrial procurement and supplier collaboration.

Syracus AI Critical Items scores delivery risk per part from your own order flow and from anonymised signals across the Tradecloud network, and tells you what to do about it. Supply Chain Ops acts on those signals with your suppliers. Both are in development; apply for the closed beta, read more about Critical Items, or see how order confirmation tolerances feed the same data.