84.8% of Vessel Cargo Estimates Are Zero. The Method That Produces Zero for Every Single Visit Has a Higher Confidence Score Than the Method That Reads Draft Data.

The Setup

In the last 90 days, 110,372 vessel visits entered the tracking layer. 96.5% of them — 106,532 visits — carry a hydrostatic cargo estimate. That coverage figure is accurate. What it conceals is the shape of those estimates.

Of the 106,532 visits with a cargo estimate, 90,323 (84.8%) carry a value of zero metric tons. Not null. Not missing. Zero.

That distinction matters for compliance workflows: a null record triggers a data-quality flag. A zero triggers a cargo check. These are two different downstream outcomes.

The Chain

Seven hydrostatic methods are active. Two produce zero for essentially every visit they touch:

• no_draft_data: 17,707 visits. Zero cargo in all 17,707 cases. Confidence score: 0.762.
• max_draft: 13,258 visits. Zero cargo in 13,257 of them. Confidence score: 0.799.

Together they account for 30,965 visits with a confident zero estimate — not because the vessel was verified empty, but because no draft differential was available.

The method designed to actually measure cargo change from draft data — draft_velocity_filtered — covers 27,901 visits and produces a nonzero figure in only 3,540 of them. Its confidence score: 0.720.

The method that guarantees zero has a confidence score 5.8 points higher than the method that at least attempts to measure cargo movement. The best-performing method is trim_saline_corrected: 3,485 visits, 95.3% nonzero, average confidence 0.833. It is the least-used method in the set by roughly 8×.

The Implication

When a downstream system sees hydrostatic_confidence = medium and estimated_cargo_hydrostatic = 0, it cannot distinguish between a genuinely empty vessel and one where draft data simply was not available. Both states look identical in the output field.

This matters for sanctions screening, cargo-weight risk models, and port capacity forecasting. A vessel flagged for STS transfer activity or dark-event correlation will often carry a cargo estimate. If that estimate is zero via no_draft_data, risk analysts may read the vessel as low-tonnage or in ballast. The compliance question never reaches cargo weight.

Of the 16,209 visits that do produce a nonzero estimate, the average cargo is 16,432 MT. The maximum is 254,640 MT. The range of what is being assigned zero is enormous.

What to Watch

The 96.5% coverage figure is a lagging signal. The meaningful metrics are nonzero estimate rate (15.2%) and confidence score calibration. If no_draft_data confidence scores drop relative to trim_saline_corrected, that indicates the confidence system is being recalibrated toward actual cargo output. It has not moved yet.

The fallback_tons_per_meter method covers 44,154 visits — 41.5% of all estimates — and produces zero in 78.8% of them. When it does produce a nonzero figure, that figure is inferred from a proxy rule, not derived from a physical draft observation. 9,340 nonzero fallback estimates sit in the dataset alongside 34,814 zeros from the same method.

Limitations

This analysis covers estimated_cargo_hydrostatic, hydrostatic_method, and hydrostatic_confidence_score in vessel_visits for the last 90 days. It does not distinguish confirmed-empty vessels (arriving in ballast with genuine zero cargo) from no-data-available zeros. That is the core limitation — and the core problem. The data layer cannot separate them either.

Data: 110,372 vessel visits, arrived_at > 90 days. Overwatch production data as of 2026-07-31.