Regime-Adaptive Proxy SVAR with LLM-Derived Narrative Instruments
International Journal of Business and Social Science Studies · 24.09.2026 · 6 (2) · 211-224
DOI: 10.66795/ijbsss73 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Business and Social Science Studies, online erschienen am 24.09.2026. Verfasst von Sophia Patel und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Proxy structural vector autoregressions (SVARs) identify dynamic causal effects with external instruments, yet they usually treat impact effects and instrument relevance as constant over time, and high-quality narrative instruments are expensive to build. This paper develops a regime-adaptive proxy SVAR (RA-PSVAR) in which a large language model (LLM) supplies a narrative instrument and a Markov-switching layer governs both the reduced-form covariance and the informativeness of that instrument. Event descriptions are scored under three analyst personas, and a coherence screen and a dispersion-based confidence weight turn the scores into a signed monthly instrument. The instrument enters the regime model through event-arrival and proxy-equation emissions estimated by an expectation–maximization algorithm, and smoothed regime probabilities weight regime-specific proxy moments that deliver impact vectors and weak-instrument-robust Anderson–Rubin sets for each regime. In seven Monte Carlo designs, static proxy estimates of the calm-regime oil price response miss the truth by up to 78% and constant-impact heteroskedastic identification breaks down, whereas RA-PSVAR lowers the median bias from 1.51 to 0.13, matching an oracle that observes the regimes; narrative emissions cut classification error by 14% when volatility contrasts are mild. In the global oil market over 1975–2025, the LLM instrument co-moves with high-frequency OPEC surprises (correlation 0.49) yet, unlike them, lowers oil production on impact. Its relevance is concentrated in a calm regime covering 86% of months (first-stage F = 9.0), where robust confidence sets are up to 38% narrower than static ones and imply smaller real-activity losses than pooled estimates.
Zitieren
APA
Patel, S., & Keller, A. (2026). Regime-Adaptive Proxy SVAR with LLM-Derived Narrative Instruments. International Journal of Business and Social Science Studies, 6(2), 211-224. https://doi.org/10.66795/ijbsss73
Vancouver
Patel S, Keller A. Regime-Adaptive Proxy SVAR with LLM-Derived Narrative Instruments. International Journal of Business and Social Science Studies. 2026;6(2):211-224. doi:10.66795/ijbsss73
BibTeX
@article{Patel2026,
author = {Patel, Sophia and Keller, Adrian},
title = {Regime-Adaptive Proxy SVAR with LLM-Derived Narrative Instruments},
journal = {International Journal of Business and Social Science Studies},
year = {2026},
volume = {6},
number = {2},
pages = {211--224},
doi = {10.66795/ijbsss73}
}
EUVIMED