Semiconductor Trade Diversion Risk: Singapore and Hong Kong

Hong Kong's imports of semiconductor manufacturing equipment grew 10x in three years — and 94% of what it re-exports now goes to China, up from 76% before US export controls tightened. A live, reproducible risk screen over real UN Comtrade trade data, built end-to-end: data pipeline, auditable SQL scoring model, and an analyst memo that survives its own rule-out checks.

Read the full memo →

The result

Tier	Count	HS code x Hub
High	0	—
Elevated	2	HKG / semiconductor mfg. equipment, SGP / processors & controllers
Watch	8	—
Low	2	—

Both "Elevated" flags looked identical on the mechanical screen. Driver analysis pulled them apart: Hong Kong's flag holds up (no domestic fab industry exists to explain the volume, and the re-export concentration on China keeps climbing every year). Singapore's flag doesn't (its growth tracks a real, independently documented $30B+ fab buildout — GlobalFoundries, UMC, Micron all expanding there — and its China re-export share is falling, not rising). Reporting a finding that doesn't survive scrutiny, right alongside the one that does, is the point: a screen that flags everything is worthless.

What this demonstrates

• Data engineering under real-world constraints — a resumable fetch pipeline (retry-with-backoff, per-call caching) against a public API that rate-limits aggressively, built after discovering live that the "obvious" paid-API path in the original project plan was unnecessary.
• Auditable analytical logic — the entire three-signal risk model lives in one readable SQL file (build_risk_score.sql), not a black-box model. Anyone can check the math against the conclusion.
• Judgment under ambiguity — HS trade codes get revised, agency API slugs don't match their public names, government data formats aren't documented accurately. Each of these got caught and fixed against a live source, not assumed from memory (see Notes from the build).
• Knowing when a finding doesn't hold up — the Singapore flag is reported and then explicitly ruled out with evidence, not quietly dropped. Same standard applied to a real, load-bearing data-integrity check on the US Census cross-validation, which turned out to confirm the pipeline rather than provide independent corroboration — and the memo says so plainly.

How it works

Three-signal risk score per HS-code/hub combination, run against live UN Comtrade bilateral trade data and a Federal Register-sourced timeline of BIS/EAR export control actions since October 2022:

1. Rate-of-growth gap — corridor import growth vs. the hub's own overall trade growth, pre- vs. post-controls.
2. Absorption gap — hub imports minus re-exports to non-China destinations minus an estimated domestic-consumption baseline. What's left over is the unexplained volume.
3. Event proximity — does growth cluster right after a control date, or trend smoothly regardless of it.

Two signals aligning tiers a row "Elevated"; all three tiers it "High." Full logic, with the reasoning behind every threshold, in src/build_risk_score.sql.

Data sources — all live, none synthetic

• UN Comtrade — HS6-level bilateral trade, 2017-2025, pulled from the free public endpoint (no account needed — see fetch_comtrade.py).
• US Census Bureau — pipeline cross-check on the US-reported side (fetch_census.py).
• Federal Register — BIS/EAR rule effective dates and Entity List actions, pulled live (fetch_control_dates.py).
• HS6 codes verified against the current UN Stats Classification Registry, not assumed from the original project brief (two of the originally-specified codes turned out not to exist under current nomenclature).

Repository structure

supply-chain-diversion-risk/
  data/
    raw/                      <- downloaded Comtrade, Census, and Federal Register pulls
    processed/                <- tidy long-format trade table, event timeline table
  src/
    config.py                 <- HS code list, corridor definitions, API keys (gitignored)
    fetch_comtrade.py         <- pulls bilateral series from Comtrade (free preview endpoint)
    fetch_census.py           <- pulls US export cross-check data (requires CENSUS_API_KEY)
    fetch_control_dates.py    <- pulls BIS/EAR rule effective dates from Federal Register
    build_risk_score.sql      <- diversion risk scoring logic
    drivers.py                <- driver analysis and chart generation
  analysis/
    findings.md               <- the full memo
    charts/
  requirements.txt
  Makefile

Running it

make all

Reproduces the full pipeline from a clean clone — fetch, parse, build the event timeline, score, and driver analysis with charts. No cloud, no external services beyond the public sources above.

Notes from the build

A few things that only showed up by checking live sources instead of trusting assumptions, kept here because they're representative of how the rest of the project was built:

• The original plan assumed a paid Comtrade API key was required. It wasn't — the free public endpoint returns the same data; the paid key only matters at higher volume.
• Two of the originally-specified HS codes (854242, 854243) don't exist under current WCO nomenclature. Caught by checking the UN Stats registry directly rather than trusting the code list as given.
• The Federal Register API's slug for the Bureau of Industry and Security is industry-and-security-bureau, not the more intuitive bureau-of-industry-and-security — found by querying the agency list, not guessing.
• The Census trade API silently rejects a plausible-looking "YYYY-MM:YYYY-MM" date-range parameter with a 400 error; a bare year works. Caught by testing against the live endpoint rather than assuming the request would work.
• A flow-direction bug in the first draft of the trade-pull script (pulling "US imports from Singapore" instead of "US exports to Singapore") was caught and fixed before it reached the scoring stage — verified with a synthetic dry run of the SQL logic before ever pointing it at real data.

Limitations

This is a screening methodology, not a finding of wrongdoing. Aggregate customs data cannot see through a single re-export step — a shipment that transits Hong Kong or Singapore and is re-labeled before continuing to China wouldn't appear in this data at all, so the method structurally understates diversion rather than overstates it. HS classification is self-reported and inconsistent across countries. The Census cross-check confirms this pipeline parses the source data correctly, not that a second, independent agency corroborates the finding — a genuinely independent check (Singapore's own SingStat, or Hong Kong's own trade statistics) remains open. Full discussion, including what would and wouldn't change the headline finding, in analysis/findings.md.