India Geopolitical Risk Monitor

Historical attention proxy, 1979–2016

Read this first. This series is a different construct from the live instrument and is never spliced into it, averaged with it, or drawn on the same axis. The live index (2017–) measures phrase-dictionary share of an English article corpus; no such corpus exists before ~2015, so this study measures the monthly share of global event-mentions involving frozen actor-pair filters in GDELT's coded event archive. "IGRM before 2017" is a phrase this project will never print. The registration — filters, anchor episodes, and publication thresholds — was frozen before the first query ran, verifiably from git history.

Two channels publish; two refused. The registered rule: a channel's historical proxy publishes only if its raw-share co-movement with the live instrument over their 36 shared months reaches r ≥ 0.4. The border channels passed decisively — pakistan_west r = 0.893, china_east r = 0.848, two entirely different constructions hearing the same waves. us_trade (r = 0.216) and gulf_energy (r = 0.153) publish no historical series: event coding and policy salience diverge exactly as the registration flagged before any number existed, and that refusal is the finding, not a footnote. Shipping never ran — chokepoint salience has no actor coding.

The two publishing channels

Trailing-10-year percentile of each channel's monthly event-mention share. The dashed vertical line is the GDELT 1.0 → 2.0 generation seam (2015-02); each segment is normalized against its own history and the raw shares are published in the payload so the seam is inspectable.

The nine frozen anchors, graded

Registered ex-ante with a top-decile rule: each anchor's month must rank in its channel's top 10% of the trailing decade, or the miss publishes with the series. Six of nine hit — and each miss is informative: Blue Star (52.3) is the registered construct-edge outcome (the bilateral filter does not hear domestic events — which is evidence it is truly bilateral); 26/11 (89.1) falls 0.9 points short because non-state perpetrators thin state-actor coding; Sumdorong Chu (54.8) sits in the thin wire coverage of 1986.

AnchorMonthChannel Percentile (trailing 10y)Top decile
Loading back_extension.json…