Every port call, berth plan, pilot booking and onward slot starts from one number: when the ship arrives. And the industry's default source for it — the ETA field the vessel broadcasts — is, in a reproducible 2025 study, worse than simply predicting the average arrival time. Not inaccurate: worse than the mean. The number the whole logistics chain plans around is, on honest analysis, nearly worthless. Research demonstration
That's the opening machine learning walked into — and it's a rare case where the ML claim doesn't need inflation, because the baseline it beats is that bad. Maritime-GPT analysis
Key findings
- ML ETA prediction works: Random Forest reached ≈99.9 min MAE on real Eastern-Med AIS, beating shipping-agent ETAs. Research demonstration
- The AIS-reported ETA is a terrible baseline — worse than predicting the mean in a leakage-free workflow. Research demonstration
- Weather fusion is the multiplier — adding marine weather cut error to 2.9–3.4 h where raw AIS alone sat near 14 h average deviation. Research demonstration
- Data leakage is the field's dirty secret — features that wouldn't exist at prediction time inflate published accuracy; honest studies guard against it. Established
- Tree ensembles (RF/XGBoost) beat deep nets on this tabular problem — the headline "AI" is a random forest, not a transformer. Established
Why trees beat transformers here
ETA prediction is a tabular problem — position, speed, distance-to-go, vessel type, historical turnaround, weather — with messy missing values and strong feature interactions. On that shape of data, gradient-boosted and random-forest ensembles consistently outperform deep networks, which want huge clean sequence data. The Eastern-Med study ran the comparison directly: Random Forest at ~99.9 min MAE edged out XGBoost and well past DNNs, KNN and decision trees. Research demonstration
The deeper point the Sines reproducibility study makes: model choice matters less than pipeline honesty. Rigorous voyage-grouped validation (so a ship's own history doesn't leak into its prediction) is what separated real accuracy from inflated accuracy — and it's what revealed the AIS ETA field's embarrassing baseline. Research demonstration
The weather is the feature
A ship's remaining time isn't just distance/speed — it's distance/speed through weather it will meet en route. The Hong Kong study is the cleanest evidence: fusing AIS movement with wave height, wind, swell and vessel parameters cut prediction error to ~3 hours where the raw average ETA-ATA deviation was 13.8 h — nearly a 5× improvement attributable largely to the weather terms. Research demonstration
That reframes the problem: ETA prediction and weather routing aren't separate capabilities — a good ETA model is partly a weather model, which is why the two research streams are converging. And it links directly to fuel: the same speed-power-in-waves relationship that determines ETA determines consumption, so a model that predicts one well usually predicts the other. Maritime-GPT analysis
Maritime-GPT synthesis of the ETA/fuel feature pipeline. Maritime-GPT analysis
Where it still breaks
Anomalous behaviour. Loitering, orders-to-slow-steam, diversion — anything off the learned pattern fails, and the honest studies name it as the key deployment limit. Research demonstration
Data sparsity per route. A model is only as good as the history on that lane and port; thin lanes predict poorly, which is why single-port studies dominate the literature. Maritime-GPT analysis
The uncertainty isn't in the number. A point ETA is a guess with false precision; the useful output is a distribution. Most published work still reports point errors, which is why "AI ETA" can look more certain than it is. Maritime-GPT analysis
What we know / don't / being tested
What we know
- Ensemble ML beats both the AIS ETA field and shipping-agent estimates on real data.
- Weather fusion is the biggest single accuracy lever.
- Pipeline hygiene (no leakage) matters more than model choice.
What we don't know
- Cross-route generalisation — a Sines model on a Shanghai call.
- Real-time continuous updating — most results are static snapshots.
- Downstream value — does a better ETA actually change berth/port outcomes, measured.
Being tested
- Multi-source fusion — AIS + weather + port-call records.
- Anomaly-aware forecasting for loitering/diversion.
- Probabilistic ETA — distributions, not point predictions.
Maritime-GPT analysis
The win isn't a cleverer model — it's an honest pipeline. The most valuable thing the recent literature did wasn't a new architecture; it was proving the default is broken (the AIS ETA field is worse than the mean) and showing that leakage-free evaluation is what reveals it. For operators, the discriminating question for any ETA vendor is: "show me your error measured only on data available at prediction time." Everything else is marketing. Maritime-GPT analysis
Open research problems
- Anomaly-aware models — predicting the ship that isn't behaving normally, not just the one that is.
- Uncertainty-calibrated ETA — distributions a planner can actually schedule against.
- Closed-loop value — does a better ETA measurably improve berth occupancy, demurrage, port emissions.
- Fuel-ETA joint models — one consistent prediction of both arrival and consumption.
Evidence scorecard
| Dimension | Rating | Basis |
|---|---|---|
| Evidence strength | High | Reproducible real-data results with honest baselines. |
| Research maturity | Field trial → deployed | Methods production-ready; integration is the remaining work. |
| Maritime deployment | Deployed | Port/terminal ETA products in operational use. |
| Confidence in conclusions | High | MAE figures and baseline findings from named studies. |
Related tools & reading
Sources
- Enhancing prediction accuracy of vessel arrival times using machine learning — J. Marine Science and Engineering 12(8):1362 (2024)Peer-reviewed
Eastern-Mediterranean AIS: Random Forest MAE ≈ 99.9 min, XGBoost ≈ 105 min — beating shipping-agent ETAs and prior models. - A reproducible workflow for AIS-based ETA forecasting: preprocessing and model selection — Forecasting (reproducible workflow study)Peer-reviewed — honest methods
Port of Sines: RF best — and the AIS-reported ETA field was <em>worse than predicting the mean</em>. Names data leakage as the field's silent inflator. - Weather-enhanced vessel arrival time prediction at Hong Kong Port: multi-source fusion — TU Delft (2025 thesis)Thesis — multi-source
AIS + weather + vessel parameters: TabPFN reached 2.88–3.42 h error vs an average ETA-ATA deviation of 13.8 h — weather fusion is where the gain hides. - Vessel arrival time to port prediction via a stacked ensemble: port-call records + AIS — Transportation Research Part C (2025)Peer-reviewed
Fusing port-call records with AIS movement — the stacked-ensemble direction for terminal-grade arrival prediction.
MAE figures are the cited studies' results on their stated datasets/ports; accessed 5 October 2026. Editorial analysis only.