Related instruments, different constructs. Two exact India benchmarks, with the uncomfortable months preserved.
IGRM retains its established name, but it measures press salience: the share of monitored coverage matching five frozen India-relevant dictionaries. It does not measure the probability or severity of geopolitical events. It is distinct from the Caldara-Iacoviello Geopolitical Risk (GPR) family.
Registered before the first comparison statistic · 7 August 2026
The primary test compares month-over-month changes in IGRM's monthly mean daily
composite with Iacoviello and Tong's monthly India_all series. The analysis
script, exact source and IGRM inputs, missing-month rule, bootstrap, decision text and
forbidden claims were frozen in a public commit before the result was calculated.
Frozen decision sentence
The two measures share a common component but are far from redundant.
| Registered statistic | Estimate | Role |
|---|---|---|
| Spearman, monthly changes | 0.256 | Inferential primary; 95% CI 0.050 to 0.407 |
| Spearman, levels | 0.252 | Descriptive only |
| Pearson, monthly changes | 0.298 | Descriptive comparability check |
| Pearson, levels | 0.351 | Descriptive comparability check |
| Spearman, changes; 12-month blocks | 0.256 | Robustness 95% CI 0.082 to 0.364 |
The sample runs from July 2017 through July 2026. A month enters mechanically if IGRM
has at least 15 valid daily composite observations and India_all is present.
Changes never bridge an excluded month. No logs, seasonal adjustment, detrending,
smoothing, winsorising or outlier removal are permitted. Raw AI-GPR values are not
redistributed. The 12-month interval rests on fewer than nine effective blocks; it is a
registered robustness check, not a precision gain.
Receipts: public registration commit · protocol · frozen script · verbatim aggregate output · run log and disclosed environment failure · largest rank divergences.
This separate benchmark compares monthly mean IGRM percentile scores with
GPRC_IND, the Caldara-Iacoviello country-specific India series. The two
pipelines share no corpus, dictionary or normalisation. Across 110 shared months, the
composite Pearson correlation is 0.232 in monthly changes and 0.484 in levels. Changes are
the sterner test because persistent series can correlate in levels through a shared trend.
The result shows limited co-movement between related measures. It is not proof that either
instrument is superior.
| IGRM series | Months | Pearson r, monthly changes | Pearson r, levels |
|---|---|---|---|
| Composite | 110 | 0.232 | 0.484 |
| US Trade | 110 | 0.276 | 0.315 |
| Gulf & Energy | 110 | 0.060 | 0.304 |
| Shipping & Chokepoints | 110 | 0.098 | 0.272 |
| Pakistan / Western Border | 110 | 0.117 | 0.229 |
| China / Eastern Border | 110 | -0.020 | -0.085 |
Live aggregate: gpr_comparison.json. The GPR input is fetched for the calculation and is not redistributed by IGRM.
The AI-GPR primary above is Spearman rho; the older country-GPR table is Pearson r.
The like-for-like Pearson change correlations are 0.298 for AI-GPR and 0.232 for
GPRC_IND, still across different samples and constructs. They are not scores
on one contest.
| IGRM | Country GPR | AI-GPR India_all | |
|---|---|---|---|
| Construct | Attention to five India-relevant geopolitical channels | Country-specific geopolitical risk coverage involving India | Article-level geopolitical risk intensity involving India |
| Frequency used here | Daily, aggregated to monthly means | Monthly GPRC_IND | Monthly India_all |
| Corpus | Worldwide English-language coverage monitored by GDELT | Independent newspaper-archive corpus | New York Times, Washington Post and Chicago Tribune |
| Output | Five channel percentiles plus a composite | One India country series | India all, initiator, respondent and spillover series |
| Interpretation | How much matching coverage appeared | Adverse geopolitical events and associated risks in news coverage | Model-scored geopolitical-risk intensity |
IGRM does not claim to beat GPR or AI-GPR. The registered result is positive but modest, which is consistent with related instruments measuring different constructs in different corpora. The only permitted discussion frames were registered before calculation: corpus, construct and anchoring. Claims such as “outperforms”, “more accurate”, “validates” or “confirms IGRM measures risk” are forbidden under every outcome.
The Divergence Register keeps the five largest AI-GPR rank gaps and the sharpest physical-flow gaps permanently visible. That is the stronger test of the project's integrity: awkward disagreements publish with the aggregate result.