The Setup

In 60 days, 41,377 vessel visits were logged. Eighty-three point three percent — 34,469 — carry an estimated_cargo_hydrostatic value. The coverage headline looks solid. The method breakdown does not.

Of those 34,469 visits with estimates:

10,726 (25.9%) were computed under the no_draft_data method. That value is 0 tons. Not missing, not uncertain — zero. These are not cargo estimates; they are null-equivalent placeholders with a tag.
15,125 (36.6%) used fallback_tons_per_meter — a population model derived from vessel class statistics, not from the vessel's measured draft during this visit.
6,393 (15.5%) used draft_velocity_filtered — a real draft-change signal with medium confidence.
514 (1.24%) used trim_saline_corrected — the only method that accounts for both trim and port salinity.
6,908 (16.7%) have no estimate at all.

The Chain

The method choice is not neutral. For bulk carriers — where cargo is the primary throughput signal — trim_saline_corrected produces an average of 5,557 tons per visit. fallback_tons_per_meter produces an average of 7,641 tons for the same vessel class. That is a 38% gap, and it runs in the direction a fallback model would not be expected to go: the population model overstates cargo relative to the saline-corrected measurement.

The reason is selection bias. The 454 bulk carrier visits that qualify for trim_saline_corrected are vessels with complete draft data at arrival and departure in ports with known water density. They are typically mid-tier vessels with moderate loads. The fallback fires on the rest, including lighter-loaded vessels where the population average overestimates what is actually aboard.

Tankers take a different path. 6,114 tanker visits use draft_velocity_filtered at medium confidence, averaging 1,427 tons. That figure reflects draft change — load transferred at berth — not departure displacement. For tankers that load partially or transfer cargo at sea, the signal is structurally incomplete.

The Implication

Any analytics layer that aggregates estimated_cargo_hydrostatic for throughput forecasting is working with a composite signal that behaves differently depending on which method fired. A port dominated by no_draft_data visits will show systematically near-zero throughput. A port dominated by fallback_tons_per_meter bulk calls will show cargo estimates that skew 38% above measurement. These are not noise — they are systematic biases by port and vessel type.

Screening workflows that use cargo estimate as a risk filter — looking for anomalously light vessels as a sanctions or smuggling signal, for example — will fire differently in ports where measurement is possible versus ports where fallback is the rule. A heavy fallback rate is indistinguishable from a genuine light-load signal unless the method tag is checked.

What to Watch

trim_saline_corrected coverage at 1.24% is not an infrastructure ceiling — it is a data availability limit. The method fires when both arrival and departure draft readings are present alongside port water density data. The port_water_density_cells table currently covers 104 cells at 35 ports. Expanding that coverage would directly lift the quality tier for the most accurate method.

The 10,726 no_draft_data visits are the most tractable gap. A subset of those vessels self-report draft in AIS messages at hourly intervals; identifying which visits have usable AIS-reported draft would pull the no_draft rate down without altering the method stack.

Limitations

The 38% overestimate gap between fallback_tons_per_meter and trim_saline_corrected is computed on overlapping but non-identical bulk carrier populations. The comparison is informative but not a controlled experiment — the fallback fires precisely on vessels where measurement is unavailable, which may have systematically different load profiles. The actual population-level gap could be smaller or larger. The draft_velocity_filtered average for tankers reflects cargo exchanged at berth, not total cargo aboard; comparing it to full hydrostatic estimates is not equivalent. Method confidence scores (0.720 for medium, 0.850 for high) are model-assigned thresholds, not empirically validated accuracy figures.

Data as of 2026-06-09. Sources: vessel_visits (60-day window, n=41,377), vessel_hydrostatic_profiles, port_water_density_cells.