In financial markets, news significantly impacts stock prices. Despite the widely postulated "Efficient Market Hypothesis," empirical evidence consistently reveals market inefficiencies, particularly when processing complex textual information. Previous research addressing these inefficiencies has predominantly employed dictionary-based methods, sentiment analysis, topic modeling, and more recently, vector-based models such as BERT. However, these approaches often lack a comprehensive understanding of textual nuances and typically neglect firm-specific economic shocks, relying excessively on headlines rather than full-text analysis. This paper addresses these limitations by leveraging Large Language Models (LLMs) to provide a comprehensive, firm-specific analysis of complete news articles. Using a dataset of Spanish business news from DowJones Newswires during a period of heightened uncertainty (June 2020 to September 2021), we apply LLMs guided by a structured news-parsing schema. This schema systematically identifies firms affected by news articles and classifies the implied economic shocks by type, magnitude, and direction. Our findings demonstrate that traditional vector embedding-based clustering methods (e.g., KMeans) yield unstable article distributions over sequential data splits, resulting in short-lived trading signals and negligible out-of-sample profitability. In contrast, the LLM-based methodology produces stable and economically meaningful clusters, generating robust and persistent trading signals. The resulting trading strategy effectively identifies winners and losers, consistently anticipating market trends by comprehending the economic implications of firm-specific shocks. Moreover, the profitability of this approach remains robust across various hyperparameter choices, including holding period lengths and the number of selected clusters. Overall, our results highlight the superiority of LLM-based analysis in capturing nuanced, economically relevant information from financial narratives, offering a promising avenue for predicting market reactions to firm-specific news during volatile periods.