IPTM signal backtest: null result

The test

For each cross-section month, two complementary statistics:

1. Top-5 vs bottom-5 spread, time-FE-demeaned — rank countries by exposure within the month, take the 5 highest-exposed and 5 lowest-exposed, compute their excess returns relative to the cross-section mean, take the spread.
2. Pooled slope on within-month exposure z-score — z-score exposure across countries within each month, regress the time-FE-demeaned return on this z-score, pooled across all months.

Both are evaluated at 1m / 3m / 6m forward horizons, with bootstrap-by-month 95% CIs (n_boot=2000, seed=42). Falsifiable: if either CI excludes zero at any horizon, IPTM has signal.

Per-quintile excess returns (1m horizon)

Each bar is the average within-month-demeaned 1m forward return for a quintile of country-month observations sorted by exposure. A monotonic increase Q1 → Q5 would mean signal. The bars wobble without trend.

Why this is the right outcome

A noisy positive on a small sample wouldn't replicate. A clean null on a falsifiable, pre-committed test preserves credibility for the qualitative claims of the IPTM register. The relational graph (country exposure / sector exposure / material exposure / theme groupings and the per-action downstream-sector / downstream-country / ETF cross-impact panels) is a structured-data layer with analytical and cross-reference value, not a quantitative trading signal. That positioning is unchanged by this result; if anything it's reinforced — we tested honestly and it didn't overpromise.

Limitations + caveats

• Small sample. 1315 actions, 44 mapped countries, ~6.5 years of post-action data. Power is limited; absence of evidence isn't evidence of absence — but the point estimates are very close to zero, not just noisily positive.
• Selection. Countries in the test are exactly those in the IPTM register — i.e., already “geopolitically interesting.” This test asks whether more exposure predicts more return relative to other geopolitically-interesting countries. A different test (geopolitically-interesting vs not) would ask a different question — likely answered by the broader factor literature already.
• Linearity. The slope test assumes a linear effect of exposure z-score on return. Strongly non-linear effects (e.g., only very high exposure matters) wouldn't show up. Future work could test bucketed effects more granularly.
• Mechanism. Even if signal existed, this test doesn't identify whether it operates through risk-premium compensation or through information-arbitrage. The null result side-steps the mechanism question entirely.