Can AI predict lead times and freight delays for fashion shipments?
Machine learning can narrow the range of likely arrival dates, but it cannot remove shocks such as rerouting, tariffs or port disruption.
KEY TAKEAWAYS Summary by the editors
- AI can predict lead times and freight delays within a range by learning from past shipments, but it performs worst exactly when conditions change, such as rerouting or new tariffs.
- UNCTAD reports that Suez Canal tonnage transits were about 70% below the 2023 average as of early May 2025, and that ships which once crossed the Red Sea in days now sail for weeks around the Cape of Good Hope.
- The Shanghai Containerized Freight Index averaged 2,496 points in 2024, up 149% from 2023, and UNCTAD describes freight rate volatility as becoming the norm.
- Fashion lead time has two parts, production and transport, and a model for one does not cover the other; buyer behaviour such as late order approval is often a larger source of delay than the voyage.
- Planners should use predicted ranges and buffers, track prediction error over time, and keep fallback rules for periods when the model's history no longer applies.
Can AI predict lead times and freight delays?
Yes, within limits. Machine learning models can estimate the likely range of production and transit times from historical shipments, and they can flag shipments that are drifting late. They cannot reliably predict events that have no precedent in their training data, so the output is best treated as a probability range, not a promise.
This analysis separates the parts of fashion lead time that models handle well from those they do not.
What makes up a fashion lead time?
A garment's lead time starts well before the container sails. It includes design and sampling, fabric and trim sourcing, approvals, production, quality checks, inland transport, ocean or air freight, customs clearance and delivery to the warehouse. Prediction models usually address one or two of these steps.
Better Buying's framework recognises that part of the problem sits with the buyer. Its Management of the Purchasing Process category assesses whether suppliers are given production lead time and whether that time is maintained through product development and launch. A late approval by a brand shortens the time available to the factory, whatever a freight model says about the sea leg.
What has been happening to freight in recent years?
UNCTAD's Review of Maritime Transport 2025 gives a useful background for why historical averages mislead. It reports that Suez Canal tonnage transits were about 70% below the 2023 average as of early May 2025, that the average voyage haul rose from 4,831 miles in 2018 to 5,245 miles in 2024, and that seaborne ton-miles grew 5.9% in 2024 against 2.2% growth in volume.
Costs moved too. The Shanghai Containerized Freight Index averaged 2,496 points in 2024, up 149% from 2023. Container spot rates approached the peaks of the pandemic period by mid-2024 and eased by the end of the year, but stayed well above pre-crisis levels. UNCTAD states that freight rate volatility is becoming the norm, and that container rates became more volatile in 2025 with new tariffs and geopolitical risks. It also notes that average port waiting times rose between December 2023 and March 2024, from 5.2 to 6.4 hours in developed economies.
Where does AI help, and where does it not?
| Question | Where models help | Where they struggle |
|---|---|---|
| When will a booked container arrive? | Learning typical transit and dwell patterns by lane and carrier | Sudden rerouting or port congestion not seen in the data |
| Will production finish on time? | Spotting factories and styles that often slip | Late approvals and fabric delays that are not recorded |
| Which orders are at risk? | Flagging drift against plan early | Explaining why, if reason codes are poor |
| How much buffer is needed? | Quantifying the spread of past delays | Periods with a structural break, such as new tariffs |
| Air or sea? | Comparing cost and time trade-offs from past cases | Rapidly changing rates and capacity |
What data does a delay prediction model need?
- Planned and actual dates for each milestone, from order confirmation to warehouse receipt, with reasons for variances.
- Shipment details such as lane, carrier, port pair, mode and container type.
- Supplier and style attributes that affect production time.
- External signals including port waiting times, schedule changes and rate indices, where reliable sources are available.
- Customs clearance outcomes, since holds and queries add time that is easy to overlook.
In many fashion companies the limiting factor is milestone data entered late or not at all. A model trained on inaccurate actual dates will learn the inaccuracy.
How should planners use predicted lead times?
- Use ranges, such as a likely and a late-case date, instead of single dates.
- Measure prediction error by lane and supplier each month and publish it internally.
- Set rules for when a model is overridden, such as a declared rerouting or a new tariff.
- Link predictions to actions, for example a trigger to review air freight or to inform wholesale customers early.
- Retrain or recalibrate when the error pattern shifts, instead of waiting for a seasonal review.
The last point matters most. Models built on a period of stable routes may be systematically optimistic after a structural change. Monitoring error is therefore as important as building the model.
A further consideration is how predictions reach the people who act on them. A delay forecast that stays inside a logistics system does little for a wholesale team that must tell customers when orders will land, or for an e-commerce team that plans launches. The output is most useful when it feeds the commitments the business makes, for instance by widening the delivery window shown to customers when the predicted range is wide.
Teams also need a view on which decisions actually have options. Re-routing, switching mode or splitting a shipment may be possible for some orders and not for others. Prediction is only worth the investment where a cheaper or faster alternative exists and someone has authority to choose it. Where no alternative exists, the benefit is limited to earlier warning, which still has value for customer communication and cash planning.
Finally, model quality should be reviewed against simple baselines. A rule that adds a fixed buffer to each lane may perform nearly as well as a complex model in stable periods. If a machine learning model cannot beat the simple rule by a meaningful margin on the shipments that matter, the extra complexity is hard to justify.
What are the risks of relying on delay prediction?
The main risks are over-trust, data gaps and cost. Over-trust leads to thin buffers justified by a model that was trained on calmer conditions. Data gaps arise when forwarders, carriers and suppliers report milestones inconsistently. Cost matters because data engineering and monitoring often exceed the licence fee for a prediction tool.
McKinsey's State of Fashion 2026 names tariffs as the top hurdle cited by executives in its survey and says that 46% of executives expect conditions to worsen in 2026, up from 39% the previous year. In that environment, a prediction model is one input to planning, alongside scenario buffers, supplier dialogue and an agreed process for communicating delays to customers.
Frequently asked questions
How accurate is AI at predicting freight delays?
Accuracy depends on lane, data quality and stability. Models do reasonably well where routes and carriers behave consistently, and poorly after sudden changes such as rerouting. Teams should measure error by lane and publish it, rather than rely on a vendor's average figure.
Why have freight transit times become harder to predict?
UNCTAD reports longer voyages around the Cape of Good Hope, lower Suez transits and volatile freight rates. Historical data from before these changes can mislead a model trained on it.
What data do I need to predict fashion lead times?
Planned and actual milestone dates, shipment lane and mode, supplier and style attributes, and reasons for delays. Milestone data that is entered late or inaccurately is the most common weakness.
Is production or transport the bigger cause of delay?
It varies, but late approvals and unstable orders from the buyer can reduce production time before transport begins. Better Buying's framework treats lead time given to suppliers as a distinct purchasing practice for this reason.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
