Nowcasting port trade from orbit: levels lie, changes travel

Vessel-tracking data built a cottage industry of trade nowcasts with a known flaw: AIS transponders can be switched off, spoofed, and gamed, precisely by the actors one most wants to watch. This paper rebuilds the nowcast from sources that do not cooperate with their subjects: synthetic-aperture radar imagery, which sees ships and infrastructure regardless of weather or transponder honesty, fused with nighttime lights and port characteristics. On 64 US ports, the model nowcasts monthly trade value at an out-of-sample R-squared of 0.945, from public satellite data alone, with no vessel-tracking feed anywhere in the pipeline.

The validation design earns more trust than the headline. Training runs January 2016 through May 2022, with testing on the fully unseen June 2022 through December 2024, a genuine temporal holdout spanning the rate cycle. The satellite-only variant, stripped of port characteristics, still reaches 0.863 on value, which isolates how much signal lives in the imagery itself. Trade weight, the physical tonnage rather than the dollar value, comes in at 0.880 for the full model. These are nowcasting numbers on official-statistics targets, the kind customs agencies publish with a lag the satellite does not have.

What the radar actually measures explains why the fusion works. SAR is an active sensor: it illuminates the scene with its own microwaves, which makes it indifferent to clouds, weather, and darkness, the three things that make optical satellite series gappy at just the latitudes and seasons where ports are busiest. Steel returns the signal hard, so berthed hulls, container stacks, and crane lines read as bright structure against water. Nighttime lights then add the activity dimension radar lacks, terminals running late shifts glow, while port characteristics anchor the physical capacity each pixel pattern implies. The fusion is three sensors answering three different questions: what is parked there, how hard is it working, what could it handle.

The levels-versus-changes result is the keeper

The paper’s most useful finding is its most candid one. Push the model outside its training domain, the leave-one-region-out exercise, and absolute levels collapse: Hawaii, held out entirely, posts an R-squared of -1.380 unadjusted, worse than guessing the mean, even while correlation holds at 0.896. Anchor the level once against any known reference and it snaps back to 0.874. Percentage changes need no such rescue: the Monte Carlo exercise recovers true changes at a slope of 0.988 with correlation 0.997, because the port-specific fixed effects that wreck level extrapolation cancel algebraically in differences. The model does not know what Hawaii’s trade level is; it knows almost perfectly how it moved, and for most decisions the movement is the question.

That distinction is the transferable methodology for every alternative-data signal a desk evaluates. Vendors sell levels because levels demo well; analysis mostly needs deltas, the same level-versus-derivative split that made emphasis-drift informative where sentiment levels were noise. A satellite trade product quoted as “port X handled Y tons” deserves the Hawaii test before anyone trusts it off-coverage. The same product quoted as “port X is up 12 percent on trend” rests on the part of the model that survived every robustness exercise thrown at it.

The sanctions case study, and who this is for

The Russia application shows the method doing surveillance-grade work. Applied to Russian ports after February 2022, the nowcast recovers the reorientation the sanctions literature later documented from official sources: European-facing Kaliningrad declining 1.44 percent while the Far Eastern outlets swing positive, De Kastri up 1.58, Sovetskaya Gavan up 0.87, Rostov-on-Don on the Black Sea up 1.15.

The pattern matters more than any single number: a public-data system detected a strategic trade shift contemporaneously, in a region where official statistics had become unreliable by design, using sensors the subjects cannot turn off. The robustness-to-manipulation property is structural, since SAR measures physical activity, berthed hulls, illuminated terminals, rather than self-reported positions. For the surveillance side of finance, sanctions compliance, counterparty exposure in opaque jurisdictions, commodity-flow verification, that property buys more than three points of R-squared, the same instrument-independence logic that makes supervision worth running outside the supervised process.

The boundaries are stated plainly in the paper and worth restating here. Coverage is US ports plus the demonstrated Russia extension; every new region needs the anchoring step before levels mean anything. Monthly cadence suits macro and credit horizons rather than trading ones. And trade value targets are themselves estimates from customs data, making the 0.945 agreement with official statistics rather than with ground truth, a distinction that matters most in the places where the tool is most valuable, in places where official statistics and ground truth have parted ways. The satellite-modality lesson from SAIFIN applies downstream: validation cadence for this class of signal runs on customs-publication lags, which is the price of nowcasting the thing the lag conceals.

The integration playbook for a desk follows the paper’s own discipline. Consume changes rather than levels by default, anchoring any region you need in absolute terms against a single trusted reference month, the step that took Hawaii from minus 1.38 to 0.874. Treat customs publications as the recurring calibration event: each official release back-tests the nowcast and refreshes the anchors, turning the validation lag into a maintenance schedule. And route the output to the horizons that fit monthly cadence, credit surveillance, macro positioning, commodity-balance work, where a one-quarter information lead on official statistics is worth real basis points and a trading desk’s intraday expectations never enter the conversation.

One adjacent use earns a sentence for the fraud-and-surveillance desks specifically: an instrument that measures physical port activity independently of self-reported data is also a verification layer for trade-finance claims, where invoices describe cargo flows someone may have an incentive to invent. Cross-checking documentary claims against orbital observation is the same trust architecture as every independent-reconciliation control in this archive, pointed at a fraud class that costs banks billions annually.

The replication package is public, the inputs are public, the recipe is documented: the moat here is execution and integration rather than access. For a fund building macro or commodity-facing data infrastructure, that is the attractive kind of paper, the one where the question is whether your pipeline team is faster than your competitors’ procurement committee.

Public SAR and nighttime lights nowcast US port trade at 0.945 out of sample, while the deeper result travels further: levels break outside the training domain and changes recover at 0.988, so buy the delta and anchor the level.