Frontier-LLM measurement of protectionist intensity in official WTO documents: Evidence from emerging economies
International Economics, cilt.187, 2026 (ESCI, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 187
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.inteco.2026.100735
- Dergi Adı: International Economics
- Derginin Tarandığı İndeksler: Emerging Sources Citation Index (ESCI), Scopus, EconLit
- Anahtar Kelimeler: Bilateral trade, Emerging economies, Gravity model, Large language models, PPML, Trade policy measurement
- İstanbul Ticaret Üniversitesi Adresli: Evet
Özet
We construct quarterly indicators of protectionist intensity in official trade-policy text for 13 emerging economies by using Claude Opus 4.7 to 5000 sentences from WTO Trade Policy Reviews and SPS/TBT notifications over 2012Q1 to 2025Q2. Claude achieves Cohen's kappa of 0.920 against 255 hand-coded sentences, above the 0.80 threshold of Ludwig et al. (2025); an open-weight Llama 3.1 8B classifier serves as a cross-check. In a PPML gravity model over 3920 bilateral observations, the indicator carries the expected negative sign in the baseline (−0.029, p = 0.005) and in the horse race against the Caldara-Iacoviello newspaper benchmark, applied MFN tariff and Global Trade Alert Red intervention count (−0.039, p < 0.001); the tariff enters negatively (p = 0.049) while the GTA count is insignificant. The result is driven by the protectionist component and is stable across pre- and post-COVID subsamples, import and export margins, and excluding the largest pair. Cross-classifier comparison with Llama shows convergence on TPR text (81%) and divergence on SPS/TBT notification metadata; under Llama the headline coefficient flips sign on SPS/TBT-heavy specifications, a methodological finding about classifier choice. The indicator does not improve out-of-sample forecasting and is positioned as an in-sample identification tool.