Domain C - Korea Trade-Cycle Nowcast (Calibration Check)

Domain C — Results: Korea Trade-Cycle Nowcast (Calibration Check)

Computed 2026-08-05 against hypothesis.md (commit fa8eef8, pre-registered before any correlation was computed). Reproducible via python -m src.domains.domain_c_korea_nowcast.compute.

H1 — Korea/Taiwan semiconductor export co-movement: passes the calibration check

The pipeline reproduces the well-established relationship. Contemporaneous (lag 0) correlation between Korea's and Taiwan's YoY semiconductor export growth is r = 0.63, already above the pre-registered 0.5 threshold. The best fit across the tested range is at lag −1 (Taiwan leading Korea by one month), r = 0.69.

Falsification criterion (max correlation ≤ 0.5) was not met — H1 is not falsified, and per the calibration-check framing in the hypothesis, this result is read as validating the pipeline's method, not just as an economic finding. Two independently-constructed aggregations (Korea summed across partners at 6-digit HS level; Taiwan summed across partners and 11-digit national codes) produce the expected strong, positive, roughly-contemporaneous co-movement that trade commentary has long attributed to these two economies. If this hadn't reproduced, the hypothesis explicitly called for treating that as a likely pipeline bug first — it didn't come to that.

The correlation is fairly flat and strong across lags −3 through 0 (0.62 to 0.69) and decays toward the positive-lag end (0.50 to 0.48 at +2/+3) — consistent with a real, roughly-contemporaneous shared demand cycle rather than one series mechanically leading the other by a fixed number of months. The slight edge at lag −1 (Taiwan leading by one month) is suggestive but not sharply distinguished from lag 0 — both exceed the threshold comfortably and the difference (0.69 vs 0.63) is not large relative to the noise in a 30-40 month sample.

What this domain could not test, stated plainly

• No intra-month nowcast. Korea's 10-day/20-day provisional releases are confirmed absent from FactIQ (reconfirmed directly before building anything). Every number here is monthly-frequency, a genuine reduction in scope from what "nowcast" implies.
• No downstream global-demand comparator. This result shows two upstream exporters moving together — it does not show either of them "leading the global cycle" against a demand-side benchmark (ISM/PMI new orders, aggregate world trade volume). No such series exists in FactIQ (ISM/PMI absent entirely; Census's m3 dataset is listed but returned zero populated series; a US HS-8542 import aggregate was a plausible candidate but the query timed out and wasn't retried). The result should be read as "Korea and Taiwan chip exports move together," not as "Korea's exports lead the world economy" — the latter claim is outside what was actually tested here.

Antithesis checks

• Composition effects, unresolved. HS 8542 blends memory (DRAM/NAND) and logic/processor chips, which have historically different cycle timing and price behavior. The strong correlation found here is on the blended aggregate; it is not known whether it holds, strengthens, or weakens within either sub-category. Not decomposed in this pass.
• Note in support of this being a supply/demand-cycle signal and not pure noise: the shape of both raw series independently matches well-documented real-world events (2022 highs, the 2023 chip-market downturn trough, 2024 recovery, a 2025–2026 surge consistent with AI/HBM memory demand) — that both series show the same real-world shape independently is itself informal corroboration before any correlation was even computed.
• Lunar New Year timing. Both economies observe Lunar New Year holidays that shift working days between January and February year-to-year. Not separately isolated in this pass — a legitimate follow-up would re-run the correlation excluding Jan/Feb to check whether the relationship survives without that shared calendar effect.
• Real vs. nominal. Both series are trade value, not volume. Given 2023's well-known memory-chip price collapse and 2025–2026's price surge, some of this co-movement plausibly reflects a shared price environment (both countries selling similar commodity memory products) rather than purely a shared demand-volume cycle. Not decomposable with value-only series; a genuine limitation, not resolved here.
• Data completeness. Both series show the same three gaps (2022-08, 2023-09, 2024-10) out of a 54-month window that neither of two separate SQL pulls returned data for. A definitive COUNT(*) verification query timed out and was not retried. That both countries are missing the same months is itself notable — worth checking whether this reflects a genuine FactIQ ingestion gap common to both schemas, rather than two independent coincidences, in a future pass.
• Mirror-statistic / base-year checks: not applicable, as declared in advance (same-side comparison, not a bilateral flow or index series).
• FactIQ vintage caveat. Every value used here is FactIQ's current, revised print (Phase 1 finding), not what KCS/Taiwan's ITA originally published for that month.

Bottom line

• The calibration check passes. The pipeline reproduces a real, well-documented relationship (Korea/Taiwan semiconductor export co-movement) with a contemporaneous correlation of 0.63 and a slightly stronger one-month-lead fit of 0.69 — good grounds for confidence in this project's general SQL-aggregation and YoY/ correlation methodology, which the other domains also rely on.
• The domain could not deliver what its name promises — no intra-month nowcast (data doesn't exist) and no tested claim about leading the broader global cycle (no demand-side comparator was available or completable this pass). What it delivers instead — a validated two-exporter co-movement check — is real and useful, just narrower than "Korea trade-cycle nowcast" suggests.
• A natural next step, not done here: retry the US HS-8542 import aggregate with an exact pre-enumerated series-ID list instead of a broad LIKE join, to get the downstream demand leg this domain is currently missing.

Pre-registered hypothesis

Domain C — Korea Trade-Cycle Nowcast (Calibration Check)

Status: PRE-REGISTERED, 2026-08-05. No cross-correlation or lag analysis has been computed at the time this file is committed — only series existence, aggregation structure, and a raw monthly total pull for two series (metadata plus construction, not a computed result). Per project principle, this hypothesis will not be edited after results are computed; if it fails, the failure is recorded, not the hypothesis.

Framing — this domain is explicitly a calibration check

Per the original spec: Korea's export-leads-the-cycle relationship is well-established in real-world trade commentary, which makes it the best test of the pipeline itself, not just of the data. If this pipeline cannot reproduce a relationship this well documented, that is a warning about the pipeline's method, not merely an inconclusive economic finding. This framing is carried through to H1's falsification criterion below.

Operational finding, reconfirmed before building anything

Korea's 10-day and 20-day provisional export releases remain confirmed absent from FactIQ (Phase 0 finding, re-checked here via search_series across export-related terms — every Korea trade series returned is Monthly frequency; no dekad-release series exists). This means the "nowcast" in this domain's name can only ever mean early monthly detection of a turn, not the intra-month nowcast the original spec envisioned. Stated here plainly, not worked around.

Data sources (confirmed present as of 2026-08-05)

Korea semiconductor exports — FactIQ korea_trade schema: No single series carries Korea's total semiconductor export value; it must be aggregated in SQL across every partner country at the CN/HS 6-digit level for HS chapter 8542 (electronic integrated circuits), flow = X (export), excluding _qty (weight) series. Confirmed constructible and non-trivial: a test pull for 2022-01 through 2026-06 (54 months) returned real values ranging from ~$5.3B/month (early 2023, the real-world chip downturn) to ~$33.6B/month (June 2026, consistent with the real-world AI/HBM memory demand surge) — a plausible, complete 54-of-54-month series, not sparse.

Taiwan semiconductor exports — FactIQ taiwan_trade schema, as the comparator: Same aggregation logic, but Taiwan's HS detail only exists at the 11-digit national level (hs_level = '11', no 6-digit rollup), so the aggregation sums across all 11-digit codes under 8542 and all partners. Confirmed constructible: the same 2022-01–2026-06 test pull returned real values from ~$21.6B/month (early 2024) to ~$51.0B/month (May 2026) — the same broad shape as Korea's series (2022 high, 2023 trough, 2024 recovery, 2025–2026 surge), which is itself a first, qualitative version of the exact relationship H1 formally tests.

A genuine "global cycle" downstream-demand comparator was attempted and not completed. No US ISM/PMI New Orders series exists anywhere in FactIQ (searched directly — absent, consistent with it being a private survey most government warehouses don't carry). Census's m3 (Manufacturers' Shipments, Inventories, and Orders) dataset is listed in search_datasets but returned zero series on direct query — listed but unpopulated. A US Census HS-trade aggregate (us_census_hs, HS 8542, "Total For All Countries" partner code) does exist and would have been a reasonable downstream-demand proxy, but the aggregation query timed out (30s limit) when joining across the full HS 8542 code range. This domain proceeds without a downstream global-demand comparator — H1 tests co-movement between two upstream exporters, not a lead against final demand. Stated as a real scope limitation, not silently dropped; see Caveats.

H1 — Korea/Taiwan semiconductor export co-movement (the calibration check)

Korea and Taiwan are the world's two dominant semiconductor exporters, both selling into the same global electronics/AI-hardware demand cycle. Their monthly export value growth rates should be strongly positively correlated at or near zero lag — this is the well-established relationship the original spec calls out.

Method (deterministic, in compute.py): 1. Compute year-over-year % change for both series (same-month-prior-year lookup, as in Domain A's yoy() — not a naive row-offset). 2. Cross-correlate the two YoY series at lags −3 to +3 months (Taiwan shifted relative to Korea; negative lag = Taiwan leads, positive lag = Korea leads). 3. Report the correlation at every tested lag and identify the maximum.

Falsification, framed per this domain's calibration-check purpose: if the maximum correlation across lags −3..+3 does not exceed 0.5, H1 is falsified — and per the framing above, that outcome should be read first as a possible pipeline problem (wrong HS aggregation, wrong date alignment, a real methodological bug) and only second as a genuine economic finding, precisely because the underlying relationship is so well documented elsewhere that its absence here is more likely to indict the method than the world.

Antithesis — required counter-checks before any result is called a finding

• Composition effects. HS 8542 blends memory chips (DRAM/NAND), logic/processors, and other integrated circuit types with different demand drivers and cycle timing (memory is famously more cyclical than logic). A correlation computed on the blended aggregate could mask offsetting sub-cycles. Not decomposed in this pass — flagged, not resolved.
• Chinese/Lunar New Year timing. Both Korea (Seollal) and Taiwan observe Lunar New Year holidays that shift the number of working days between January and February from year to year, exactly the kind of distortion the original spec calls out for Asian Jan/Feb data. Any correlation driven disproportionately by Jan/Feb co-movement should be checked separately from the rest of the year before being read as proof of a demand-cycle relationship rather than a shared-calendar artifact.
• Real vs. nominal. Both series are trade value in current USD, not volume — memory-chip prices are notoriously volatile (the 2023 price crash, the 2025–2026 AI-driven price surge), so some of any observed co-movement could reflect a shared price cycle (both countries selling the same commodity-like memory chips into the same price environment) rather than a shared demand-volume cycle. This is a real ambiguity, not resolved by a value-only pipeline pass.
• Base-year changes: not applicable — HS trade value is not an index-linked series.
• Mirror-statistic check: not applicable — this is a same-side (export-to-export) comparison between two countries, not a bilateral flow with a counterparty ledger.
• FactIQ vintage caveat. Every value used here is FactIQ's current, revised print (Phase 1 finding, reconfirmed by FactIQ directly), not what KCS/Taiwan's ITA originally published for that month.

Caveats (known before results; will be restated in the results writeup)

• No genuine intra-month nowcast is possible — Korea's 10-day/20-day releases are confirmed absent. This domain can only ever detect a turn at monthly frequency, a real reduction in scope from the original spec's framing.
• No downstream global-demand comparator (ISM/PMI absent; Census m3 listed but empty; the one viable candidate, US HS-8542 import totals, hit a query timeout) — H1 tests two upstream exporters against each other, not against final demand. A future pass could retry the US Census aggregate with a narrower, pre-enumerated list of exact 10-digit series IDs instead of a broad LIKE join.
• HS 8542 is a blended aggregate (memory + logic + other ICs) — no sub-category decomposition in this pass.
• Value, not volume — see Real vs. nominal above.
• FactIQ's trade data is not vintage-aware — see FactIQ vintage caveat above.

Independent personal project by Ian Bergin. Not affiliated with, endorsed by or sponsored by FactIQ, which is used as one of several data sources.