Methodology · v1
How TremorWatch builds the data
TremorWatch monitors public risk feeds, deduplicates and geocodes events, summarizes each in EN/KO/JA with a frontier LLM, and maps them to chokepoint and commodity exposure. This page documents the pipeline so analysts can audit methodology before relying on the output.
1. Sources
We ingest from five public feeds. ACLED is licensed-restricted; we keep the raw events for internal analytics but exclude them from any paid dataset bundle. The remaining four are open-license.
| Source | Cadence | License |
|---|---|---|
| GDACS (Global Disaster Alert) | 15-min RSS | Open |
| GDELT | Hourly BigQuery | Open |
| NASA FIRMS (active fire NRT) | 6-hour | Open |
| WHO DON (Disease Outbreak News) | 6-hour JSON | Open |
| ACLED (conflict events) | Weekly REST | Restricted — kept in raw store, excluded from licensed exports |
2. Deduplication
Each ingested event has a stable external_id (source-prefixed) and a geocoded point. We deduplicate within a 100-km / 24-hour window using PostGIS ST_DWithin. Conflicts pick the earliest occurred-at and the most-detailed description, preserving both external_ids.
3. AI summarization (EN canonical)
Each event passes through a Claude-Sonnet-4 prompt producing a 2–3 sentence summary (ai_summary) and a severity grade (critical / high / medium / low). EN is the canonical column; KO/JA are derived translations. All AI output is cached in the database and never regenerated for the same event.
A refusal-pattern garbage filter drops translations starting with "I can't", "申し訳", "죄송" etc. Failed translations are stored as NULL, not as garbage.
4. Chokepoint and commodity mapping
Events within ~250 km of a curated chokepoint (Suez, Strait of Hormuz, Panama, Bab el-Mandeb, Bosphorus, Taiwan Strait, etc., 17 total) are tagged with a chokepoint-exposure score (0–100). Commodity associations come from keyword + entity-recognition over the AI summary, mapped to a curated list of 13 commodity tickers.
Known limitation: chokepoint mapping coverage is sparse on the current MVP (~0.5% of non-ACLED events have a chokepoint association). We are evaluating whether this is a structural artifact (most events occur outside chokepoint radii) or a pipeline gap; the result will be reflected in v1.1 export.
5. Severity scale
We use four severity tiers (critical / high / medium / low) defined by source-reported magnitudes (e.g., GDACS earthquake intensity, FIRMS fire confidence, WHO outbreak risk) plus override rules from the AI summary when the source label is missing or stale.
6. Output formats
The licensed export produces three tables (events / chokepoints / entities) in Apache Parquet (Snappy) with a 10% CSV sample. Array columns are JSON-string encoded for BigQuery / Athena compatibility. Schema and column list: /data.
7. Update cadence and versioning
The source_dataset column tags every row with the snapshot version (e.g., 2026-05-08). The licensed package gives 12 months of weekly snapshots from the purchase date. A change-log per snapshot is published alongside the file.
8. What we deliberately do not do
- No financial impact estimation. We provide event exposure; translating that into expected loss requires actuarial models we do not run.
- No user-private data. All sources are public. No PII.
- No real-time streaming. Updates are weekly batch. Customers needing real-time should use the original source feeds directly.
9. License and attribution
Licensed exports include a copy of the license and the SPDX-style attribution line: "Contains data from GDACS, GDELT, FIRMS, WHO, derived by TremorWatch (Volt AI)."
Questions or audit requests: hello@tremorwatch.com.
Methodology in plain language
TremorWatch evaluates supply-chain risk by combining public event signals with structured context about countries, commodities, chokepoints, and topics. The system is designed to highlight where disruption pressure may be increasing, not to declare that a specific company or shipment has failed.
Inputs are organized around severity, recency, geography, and relevance. A recent high-severity event near a critical logistics route receives more attention than an older low-severity event with no clear supply-chain connection. Commodity and country relationships provide additional context so readers can see why one event may be more important than another.
Human judgment remains necessary. The public scores and briefings should be used as a triage layer before deeper research. Analysts should validate signals with official notices, local reporting, supplier communication, logistics data, and market information before making procurement, investment, or operational decisions.