Project Goldstein
Geopolitical risk is a variance input — so test variance, not direction. Twelve chokepoints, mapped to liquid ETF proxies, validated 2022–2026.
ProblemGeopolitical risk is normally reported as a country rating, an analyst note, or a headline index — outputs that cannot be backtested or compared across events. Most attempts also try to predict direction, which is exactly where event data is weakest.
SolutionGoldstein maps 12 global chokepoints to liquid ETF proxies and tests one claim: does event-level conflict data from GDELT and ACLED predict a rise in the linked instrument's variance, not its return? The Geopolitical Risk Premium Score (GRPS) combines event-based instability, a sector volatility premium versus rolling benchmarks, and VIX z-score conditioning, then reduces to a three-regime label — STABLE / ELEVATED / CRITICAL.
Implementation / progressBuilt as a validation framework with a hard 252-day per-region warm-up, deliberate proxy isolation, and out-of-sample testing. Current output: 323 validated threshold-crossing events over a 21-day forward window, a 64.4% hit rate for realised vol above the 75th percentile, a published 29.2% average false-positive rate, and Spearman ρ from 0.08 to 0.39 (all p<0.05). The data pipeline, fetchers, quality checks, and backtest framework are open-source; the scoring engine is private.
| Region | ETF Proxy | ρ (geo→var) | p-value | Rationale |
|---|---|---|---|---|
| Taiwan Strait | SOXX | 0.393 | <0.001 | Semiconductors — TSMC supply-chain risk |
| Strait of Hormuz | USO | 0.378 | <0.001 | Crude oil — chokepoint transit premium |
| Korean Peninsula | EWJ | 0.339 | <0.001 | Japanese equities — escalation spillover |
| Sahel / W. Africa | GDX | 0.314 | <0.001 | Gold miners — resource conflict / coup premium |
| India–Pakistan | INDA | 0.298 | <0.001 | Indian equities — subcontinent escalation |
| Venezuela | ILF | 0.277 | <0.001 | LatAm equities — oil-state political risk |
| South China Sea | EWH | 0.251 | <0.001 | Hong Kong equities — trade exposure |
| Panama Canal | IYT | 0.249 | <0.001 | Transport — global shipping dependency |
| Red Sea / Suez | IYT | 0.225 | <0.001 | Transport — Suez transit disruption |
| Eastern Europe | XME | 0.176 | <0.001 | Metals & mining — commodity shock channel |
| Russia / Arctic | XOP | 0.114 | <0.001 | Oil & gas E&P — Arctic / sanctions exposure |
| Middle East | XLE | 0.083 | 0.017 | Broad energy — diffuse exposure attenuates signal |
Variance, not returns
Across 12 regions, ρ vs. 21-day realised vol ranges 0.08–0.39 (all significant). No return-prediction relationship held — the variance channel is the one that works.
Out-of-sample signal
323 threshold crossings validated on a 21-day forward window. The signal measures a structural regime, not an inefficiency, so it does not decay on publication.
Proxy isolation
Each of the 12 proxies is chosen to isolate region-specific exposure. Early iterations reused XLE and produced mathematically identical signals.
Regime classification
The continuous 0–100 score reduces to STABLE / ELEVATED / CRITICAL. The label is the summary output; the score is the audit trail.
VIX conditioning
Adding a VIX z-score conditioning variable cut false positives by roughly 8 percentage points versus the unconditioned model.
Warm-up exclusion
The first 252 trading days per region are excluded from all validation — no exceptions. Published results reflect only the post-warm-up window.