Portrait of Dr. Avinash Kori

Dr. Avinash Kori

Research Associate, Imperial College London
Co-founder & Chief Scientist, Blurgs AI

Portrait source

Prediction is not intervention: a causal view of port-call optimisation

2 minute read

In brief

  • ETA prediction answers what is likely; causal modelling asks what would change if we acted.
  • The useful state joins AIS, weather, berth/service events, cargo messages and reporting quality.
  • The decision target is the effect of feasible interventions on waiting, fuel, congestion and emissions—not another delay score.

A recurrent inefficiency in port operations occurs when a vessel arrives before the berth, terminal, tug, cargo operation or clearance process is ready. The resulting waiting time may increase fuel use and emissions without improving service completion time.

Predicting this delay is useful, but prediction alone does not identify which feasible intervention would change the outcome.

The causal question is sharper:

Would this vessel have waited less—and burned less fuel—if it had slowed earlier, been assigned differently, or received better operational information?

From events to a causal state

Port calls are multimodal processes. I propose to connect:

  • AIS-derived arrival, anchorage, berth and departure states;
  • weather, tide and metocean context;
  • berth, pilotage, tug and terminal-service events;
  • cargo/load-discharge messages and vessel attributes;
  • clearance/reporting timestamps and missing-data indicators.

Standards such as DCSA port-call events and IMO Just-In-Time guidance provide a common operational language. The research layer is to turn those timestamps and states into variables, treatments, outcomes and assumptions suitable for causal analysis.

Discover mechanisms, then test interventions

The modelling pipeline is:

multimodal state → expert-constrained causal graph → structural mechanisms → intervention effects.

Data-driven causal discovery should not operate unconstrained. Port processes contain known temporal orderings and operational impossibilities. Experts should be able to reject edges, add missing mechanisms and flag reporting artefacts.

The resulting model must survive more than graph fit. I propose to test stability across ports and time, sensitivity to unobserved confounding, expert agreement and interventional consistency.

Decision quantities

Candidate treatments can include:

  • Just-In-Time speed advice;
  • berth reassignment or sequencing;
  • improved cargo/reporting timeliness;
  • arrival smoothing under congestion.

Outcomes can include waiting time, fuel/CO₂ proxies, berth delay, departure delay and downstream congestion.

Average effects are not enough. The same intervention can help one vessel class or congestion regime and hurt another. Conditional treatment effects are therefore the practical unit: what works, for whom, and under which port/weather/cargo state?

Why this is different from optimisation on predictions

A predictive optimiser can exploit correlations that fail under intervention. A causal decision model has to state which variables are actionable, which are context, and which assumptions make the effect identifiable.

That enables questions such as:

  • would slowing earlier reduce total fuel without increasing berth delay?
  • did weather cause the delay, or merely coincide with a terminal bottleneck?
  • would local berth optimisation shift congestion downstream?
  • which missing/reporting timestamp has the greatest operational consequence?

The goal is not a universal causal dashboard. It is a small set of stakeholder-relevant interventions whose effects can be tested through simulation, retrospective cases and field trials where feasible.

Impact

A credible causal layer changes maritime analytics from delay reporting to decision evidence: root-cause attribution, counterfactual comparison and treatment-effect estimation at vessel, port and corridor level.

That is the bridge I see between causal AI and maritime operations: not adding causal terminology to an ETA model, but making operational interventions explicit and testable.

References and standards