Container Delay Prediction: Which Data Signals Improve ETA Accuracy?

The signals that improve container delay prediction change with the milestone being forecast and the time remaining before it. Vessel position and voyage progress help estimate arrival near the destination port. Berth availability and terminal workload become more informative for discharge. Transshipment connections determine whether the container stays on its planned itinerary, while release events matter when the target is cargo availability or gate-out. A signal adds value only when it is available before the prediction and improves performance against a credible baseline.

This makes ETA accuracy a question of both physical movement and information timing. A detailed terminal event received after an operational decision offers little forecasting value at that earlier point. Conversely, an incomplete but timely indication of a missed connection can justify a substantial ETA revision.

Define the arrival event before comparing accuracy

“Container arrival” can refer to entry into port limits, arrival at anchorage, vessel berthing, container discharge, cargo availability, or terminal gate-out. These events have different causes of delay. A vessel can reach the destination area on schedule while its containers remain unavailable because of berth queues, yard operations, documentation, or inspection.

A defensible prediction specification identifies the target event, location, timestamp convention, and forecast issue time. It also defines what constitutes delay: deviation from the original booking schedule, the latest carrier estimate, or an operational deadline. These references are not interchangeable. Measuring against a repeatedly revised schedule can conceal deterioration relative to the original commitment.

For a discharge forecast, the remaining journey can be separated into sailing time, waiting for a berth, and time from berthing to container discharge. Cargo availability adds relevant release and terminal processing conditions. Some activities overlap, so simply adding average customs time to average terminal dwell can double-count elapsed time.

Model the event that supports the intended decision. A highly accurate port-arrival estimate does not establish accuracy for warehouse delivery or cargo availability.

Where each signal earns its place

No universal ranking of input data applies across every route and forecast horizon. The useful comparison is whether a signal explains the next unresolved constraint. The following distinctions help separate informative inputs from data that merely looks detailed.

Vessel movement establishes progress, not container certainty

Automatic Identification System (AIS) observations support estimates of location, speed, heading, and movement history. Their value depends on correct vessel identification and reliable shipment-to-voyage linkage. If a container is rolled to another departure, accurate tracking of the originally booked vessel produces an accurate forecast for the wrong movement.

Remaining distance should reflect the plausible sailing route rather than a straight line between coordinates. Recent speed also needs context: slow movement near an anchorage is different from sustained slow steaming offshore. Extending either speed mechanically across the remaining journey creates unrealistic forecasts.

Position freshness should enter the model explicitly. When observations stop, confidence should fall as the gap grows. Repeatedly carrying forward the last position without recording its age makes stale information appear current.

Congestion needs a destination-specific interpretation

A count of vessels near a port is an imperfect measure of the queue relevant to a particular container ship. The count can include vessels using different terminals, ships awaiting instructions, or traffic unrelated to container operations. More useful features connect waiting activity to the destination terminal, vessel class, service pattern, and recent berth turnover.

Queue length alone also misses processing capacity. A growing queue with rapid berth turnover differs from a smaller queue facing disrupted operations. Recent waiting-time distributions and departure activity provide context, although estimates become unstable when only a few comparable calls are available.

Terminal dwell requires equally careful interpretation. Long discharge-to-gate-out intervals can reflect customs holds, collection scheduling, or storage behavior. Treating the full interval as terminal congestion incorrectly attributes every delay to handling capacity. For cargo availability, release-state information is more relevant than the eventual collection time.

Connections create discontinuous delays

Transshipment often produces a step change in ETA. A small inbound delay can consume the connection buffer and move the container to a later sailing. Until that threshold is crossed, the same delay might have little effect on destination arrival.

Useful features include expected inbound discharge, onward loading cutoff, terminal transfer requirements, planned departure, and evidence that the container was actually loaded. Connection feasibility must allow for handling and transfer time; comparing vessel arrival directly with onward departure overstates the usable buffer.

A connection timeline is particularly useful here: inbound arrival, discharge, transfer, loading cutoff, and onward departure expose where the apparent buffer disappears. The forecast should also consider the replacement service. Missing a connection does not imply a fixed delay when onward sailing frequency and available capacity differ.

Weather and release events require the right target

Weather features become useful when matched to the expected route and time window. A storm near the destination today is weak evidence for a vessel arriving much later. Forecast conditions along the planned passage, together with relevant port restrictions, offer a closer causal connection. Historical evaluation must use weather forecasts available at the prediction time rather than subsequently observed conditions.

Customs status has a different relationship to ETA. It directly informs release-related milestones, but it does not ordinarily explain offshore vessel progress. A pre-arrival clearance event can reduce uncertainty about cargo availability; an inspection hold changes the remaining release process. Absence of an event should not automatically be interpreted as absence of clearance, particularly where reporting is intermittent.

Freshness and coverage determine usable signal quality

Every event should retain both its occurrence time and its first availability to the prediction system. A discharge event reported late belongs to the physical timeline at discharge, but it was unavailable to forecasts issued before receipt. Backtests that ignore this distinction give the model information it could not have used.

Three time fields are useful: event occurrence, source publication, and system ingestion. Where publication time is unavailable, ingestion time provides a conservative boundary for historical replay. Corrections should remain versioned so that a later revised event does not silently replace the earlier information state.

Missingness also carries structure. One terminal may report frequently while another publishes only selected milestones. Feeding both into a model without source and freshness indicators can make reporting practices look like differences in operational performance. Coverage should be measured across shipment stages, locations, and actual prediction opportunities.

There is also an availability tradeoff: rich terminal data arriving shortly before discharge might improve a near-term forecast while contributing nothing to an earlier inventory decision. Signal value should therefore be reported by forecast horizon, rather than as one blended accuracy figure.

Prove incremental value without leaking the outcome

Begin with baselines that represent information already available: the current carrier ETA, a route-and-service historical estimate, and a simple voyage-progress forecast where appropriate. Preserve the carrier estimate as it existed at each forecast issue time. Comparing a model against an outdated schedule while allowing the model fresh operational data exaggerates the gain.

Use chronological evaluation: train on earlier shipments and test on later ones. Randomly splitting event rows risks placing updates from the same shipment in both sets. Containers on the same vessel call also share much of their delay exposure, so grouping related movements helps prevent an apparently large test set from overstating independent evidence.

For each signal family, compare the baseline, the model with that family added, and the full model with it removed. These comparisons reveal whether the signal contributes useful information or largely duplicates another source. Evaluate the same shipment population when possible; otherwise, improved scores may reflect easier routes with better data coverage.

Common leakage paths include final voyage assignments substituted for earlier booking records, completed dwell durations used before departure, and congestion summaries calculated using observations after the forecast. Even a valid-looking feature such as “remaining connection time” leaks information if it uses an actual onward departure that was unknown at issuance.

Measure errors that change operational decisions

Mean absolute error describes typical timestamp accuracy, but it hides error direction and severe misses. Report signed error with an explicit convention—for example, predicted time minus actual time—alongside a tail-error measure. Persistent early predictions have different consequences from conservative late predictions, even when absolute errors are equal.

Results should be separated by target milestone, route, forecast horizon, direct versus transshipment movement, and data availability. Measure delay detection against the agreed schedule reference as well. A model can improve average ETA error while still failing to identify shipments that cross an important lateness threshold.

Prediction intervals add useful information when their coverage is tested. If an interval is intended to contain a stated share of outcomes, compare that claim with observed coverage and interval width. Very wide intervals achieve coverage without supporting precise planning. Narrow intervals that regularly miss delayed shipments create unjustified confidence.

Event-triggered forecasts need a consistent evaluation schedule. Scoring one system only after informative updates, while scoring another throughout the journey, creates an unequal comparison. Replay forecasts at shared decision times or clearly separate fixed-time and event-triggered results.

The final deployment test is a chronological replay using the data actually available at each issue time, including missing feeds and delayed reports. Retain the previous forecast when a source fails, or apply a defined fallback, and measure that behavior too. An ETA model earns operational trust when its reported accuracy survives those ordinary information gaps.