← Back to precision

Methodology & SLA

This page documents the platform's data sources, models, persistence and service commitments. It's the document that B2B buyers' data teams (traders, insurers, freight forwarders) review before signing.

Data sources

Cargo classification

Combination of AIS shipType signals and a textual heuristic (vessel name, declared destination) to assign a cargo class — tuned per port and fleet profile. Edge case: minimal AIS metadata falls back to a generic class.

Voyage detection

ETA model

  1. Naive estimate: distance / SOG where distance is the great circle between current position and port center. Refreshed continuously per active voyage.
  2. Seasonal correction: rolling median of error (predicted − actual) bucketed by arrival hour. Falls back to the global median when the bucket sample is too thin. Recomputed regularly.
  3. Comparison reference: ETA broadcast field extracted from ShipStaticData messages (entered manually by the vessel's crew — often imprecise and late).
  4. Metrics: RMSE and MAE in hours, computed over closed voyages with both predicted ETA AND broadcast ETA available. The broadcast baseline excludes ETAs more than 7 days from actual arrival: these are AIS sentinel values (default dates, never-updated ETAs, placeholders) — not genuine estimates — and would otherwise distort the RMSE. Model and broadcast are compared over exactly the same set of voyages. Updated on each closed voyage.
  5. Model roadmap: congestion integration, tides, weather, vessel-specific historical average speed.

Anomaly detection

v1: absolute dwell-at-anchor thresholds, tuned by cargo class.

Roadmap: thresholds derived from the historical (port, cargo) distribution; route-deviation detection; out-of-zone loitering detection ("dark fleet" signal).

Persistence & lineage

Service commitments (SLA v1)

Platform availability99.5% / month (MVP)
Live positions latency< 30s (P95)
KPIs / voyages latency< 90s (P95)
Webhook delivery1 retry at 60s · 90-day delivery log
History retention7 days in-memory KPIs · unlimited in SQLite (90-day compaction)
BackfillOn contractual request (replay of persisted positions)

Feed transparency & degradation

No AIS provider in the world has 100 % uptime. Our commitment is not to hide degradations but to detect and communicate them in real time.

Sanctions screening — multi-regime coverage

Vessels are reconciled against four official lists, refreshed daily and matched on IMO/MMSI:

Match on 7-digit IMO first (authoritative), MMSI as fallback. Red halo on the map and 🚫 badge in voyage tables.

Maritime chokepoint transit detection

12 zones tracked continuously: Suez, Hormuz, Bab el-Mandeb, Malacca, Singapore, Bosphorus-Dardanelles, Gibraltar, Skagerrak-Kattegat, Dover Strait, Panama, Cape of Good Hope, Magellan.

CO₂ emissions estimation — in-house method

Bottom-up approach derived from the IMO Fourth GHG Study (2020), integrated over the AIS feed with no external dependency.

Compliance

Attributions & licences

  • AIS terrestre : aisstream.io — flux communautaire AIS (terms of use libres pour usage technique dérivé).
  • Météo : Open-Meteo — données sous licence CC-BY 4.0. Source citée à chaque affichage.
  • Imagerie radar : « Contains modified Copernicus Sentinel data », ESA — données ouvertes Copernicus.
  • Cartographie : tuiles © CARTO (CC BY 3.0) sur fond © OpenStreetMap contributors (ODbL 1.0).
  • VIIRS Boat Detection : produit commercial du Payne Institute, Colorado School of Mines (licence requise) — connecteur prêt côté code, activation après contrat. Citation à inclure : Elvidge, C.D., et al., « VIIRS Boat Detection (VBD) ».
  • Données AIS satellite premium(Spire / MarineTraffic / Orbcomm) : aucune utilisation par défaut. Activation contractuelle requise — termes du fournisseur s'appliquent.