47% of Vessel Visits Do Not Have Draft Data. The Cargo Estimate You Are Reading Is a Fallback.
If you build trade-flow models from observation-based cargo estimates, the methodology coverage matters more than the methodology accuracy. Here is the actual distribution.
The Setup
Of 3,161 vessel visits with hydrostatic cargo estimates in the last 90 days, 1,473 visits (46.6%) had no usable draft data at all and were resolved by class-prior fallback. Another 1,028 (32.5%) used the next tier — a fallback tons_per_meter formula that assumes a uniform deadweight-per-immersion ratio across vessel class. Just 2 visits (0.06%) used the gold-standard vessel_curve method, which integrates the actual hydrostatic curve for that specific hull.
The accuracy hierarchy that researchers cite — vessel_curve >> max_draft > draft_velocity_filtered > fallback_tons_per_meter >> no_draft_data — is real. But the coverage hierarchy is inverted: the most accurate method covered 0.06% of visits, while the least accurate covered nearly half.
The Chain
A hydrostatic cargo estimate is built from three inputs: vessel hydrostatic profile (a per-hull table of draft-to-deadweight points), AIS-broadcast draft, and port water density. When all three are available, the estimator reads off the ship exact loaded tonnage by interpolating the curve. When only two are available, it falls back to a class-mean linear approximation (fallback_tons_per_meter). When draft data is missing or noisy, it falls back further to max_draft (the deepest reading from the visit) or class priors (no_draft_data plus a fallback assumption that the vessel is at typical-loaded condition).
The reason vessel_curve is rare is not methodology — it is data. Vessel hydrostatic profiles are non-public for the vast majority of the merchant fleet. They live with classification societies, owners, and charterers, not in any public registry. Without them, the estimator has no curve to interpolate. Hence: 2 visits.
The Implication
A trade-flow model built on aggregated hydrostatic estimates is not measuring what its users think it is measuring. It is measuring a class-mean approximation across 95%+ of visits, with the actual hull-specific signal arriving for less than 1 in 1,000 cases. That is not necessarily a problem — class-mean estimates are usable for fleet-aggregate flows, especially in commodity-flow models that already integrate over thousands of voyages. But it does mean single-voyage trade-flow predictions made from hydrostatic estimates carry a wider uncertainty band than the methodology label suggests.
For analysts using this data: when a single-vessel report says "the MV X loaded 87,243 metric tons," the underlying number is much more often "the average bulk carrier of class Y loaded ~9,000 tons per meter of draft change, and X moved 9.7m." That is a useful quantity, but it is not a survey reading. Treat it as such.
What to Watch
- Coverage of
vessel_curveover the next quarter. A jump from 2 visits to 200 visits would mean the hydrostatic profile registry has gained meaningful breadth — that is the milestone to look for. - Confidence distribution of
fallback_tons_per_meter. This is the workhorse method; if itsmediumshare grows relative tohigh, the fallback class priors are getting noisier (more out-of-distribution vessel classes entering the dataset). no_draft_dataas a percentage of total visits. This tracks AIS data hygiene. If it drops, draft-broadcast compliance is improving in your covered fleet. If it rises, vessels are increasingly suppressing the field.
Limitations
The 3,161 visits over 90 days does not represent the global merchant fleet — it represents the subset of port calls that the Overwatch pipeline observed cleanly enough to attempt an estimate. Visits at low-AIS-coverage ports are systematically under-represented. The method-classification logic itself can mis-tag a borderline case (a draft profile with only two valid readings can be classified max_draft or draft_velocity_filtered depending on how the points are spaced), so the boundary between medium and high confidence is fuzzier than a strict ordinal would suggest. Finally, hydrostatic estimates are inherently observational — they cannot distinguish between cargo, ballast water, and bunker loaded into double-bottom tanks.
Data current as of 2026-05-01. Source: vessel_visits.estimated_cargo_hydrostatic over 90 days. See cargo-estimation skill for method definitions.