Interactive Causal Network Map

This viewer shows lead–lag networks among 50 ICIO V1 sectors, estimated from news about the 2026 Iran war in four language corpora (Arabic, Chinese, English, Persian; 58,530 items, January–April 2026). It accompanies the paper Qualitative Shock Mapping for Quantitative Production Network Models: Evidence from Iran-Related Hostilities (Kim & Erokhin, IIASA).

Paper link: forthcoming.

Method

1. Collect. News items on the hostilities are gathered per language stratum with publication date, source, and risk metadata.
2. Map to sectors. Each article receives a probability vector over the 50 ICIO V1 sectors (lexical TF-IDF classifier; a zero-shot LLM classifier is used for robustness). Direct industry evidence takes priority; purely political items map mainly to public administration and defence.
3. Aggregate daily. Article vectors are summed per day and language, weighted by negativity and war-relatedness, then transformed to within-day shares; low-information (near-uniform) articles are excluded.
4. Test lead–lag pairs. All 2,450 ordered sector pairs are tested with Toda–Yamamoto lag-augmented Granger tests on the daily series.
5. Select edges. Pairs are kept if they pass Benjamini–Hochberg false-discovery-rate control at 5%; each edge is annotated with its lead (1–7 days) and the lagged Pearson correlation.
6. Draw. Sectors sit on a fixed ring (V1 1 at the top, clockwise), sized by total exposure and coloured by sector group; edges are shaded by lead and dashed when the correlation is negative.

Data and code

Derived indicators, edge tables, and scripts are in the replication package (data CC BY 4.0, code MIT), archived on Zenodo: 10.5281/zenodo.21257558. Raw article text is not redistributed.