War Risk: Time Series and Cross-sectional Evidence from the Stock and Bond Markets

Abstract

We employ a semi-supervised topic model to extract the rare disaster risks and economic narratives from 7,000,000 NYT articles over 160 years. Our approach addresses the look-ahead bias and changes in semantics. War positively predicts market return in- and out-of-sample, while the economic narratives only predict in-sample. The predictability of War increases over time and is robust when extracted from WSJ. War as a solo factor prices characteristics-sorted portfolios with a negative risk premium and outperform some multifactor benchmarks when pricing machine learning-based nonlinear portfolios with an R2 of 54%. Our study lends support to the time-varying disaster risk model.

Keywords: War, disaster risk, topic modeling, narratives, War factors

JEL Classification: G00, G1