Markets do not always efficiently incorporate news, particularly when information is complex or ambiguous. Traditional text analysis methods fail to capture the economic structure of information and its firm-specific implications. We propose a novel methodology that guides LLMs to systematically identify and classify firm-specific economic shocks in news articles according to their type, magnitude, and direction. This economically-informed classification allows for a more nuanced understanding of how markets process complex information. Using a simple trading strategy, we demonstrate that our LLM-based classification significantly outperforms a benchmark based on clustering vector embeddings, generating consistent profits out-of-sample while maintaining transparent and durable trading signals. The results suggest that LLMs, when properly guided by economic frameworks, can effectively identify persistent patterns in how markets react to different types of firm-specific news. Our findings contribute to understanding market efficiency and information processing, while offering a promising new tool for analyzing financial narratives.

The EM principle implies the moments underlying the information matrix test for multivariate Markov switching autoregressive models with covariate-dependent transition probabilities are the smoothed values of the moments we would test were the latent Markov chain observed. Thus, we identify components related to the heteroskedasticity, skewness and kurtosis of the multivariate regression residuals for each of the regimes, the neglected multivariate heteroskedasticity of the generalised residuals for each of the columns of the transition matrix, and a final component that assesses the conditional independence of these generalised residuals and the regression residuals, their squares and cross-products given the observed variables.