Can News Predict the Market? Limits of Zero-Shot Financial NLP and the Role of Explainable AI

Problem
This work addresses the unresolved question of whether financial news can reliably predict short-term stock movements using zero-shot natural language processing (NLP) techniques. Despite advancements in large language models, the authors note a gap in the literature regarding the effectiveness of these models in the financial domain without domain-specific training. The paper is a preprint and has not undergone peer review.

Method
The authors propose a structured pipeline that integrates zero-shot natural language inference with temporal aggregation. This approach explicitly models recency and event-dependent impact horizons to synthesize information from multiple articles. The framework includes a multi-layered explainability component that connects predictions to token-level, article-level, and aggregate evidence, generating grounded natural language rationales for the predictions. The models evaluated include various zero-shot NLP architectures, although specific architectures and training compute details are not disclosed.

Results
The findings indicate that zero-shot approaches consistently underperform against simple baselines across multiple models and prediction horizons. Notably, the models exhibit particularly weak performance in predicting negative stock movements. The authors report that the best-performing zero-shot model achieved an accuracy of only 45% on predicting stock price increases, while a naive baseline based on historical price trends achieved 50%. The explainability framework, however, successfully identified trustworthy predictions, suggesting that while accuracy is limited, the ability to discern reliable signals is a valuable outcome.

Limitations
The authors acknowledge that the zero-shot models fail to capture the complexities of financial sentiment and its relationship with stock price dynamics, indicating deeper structural limitations in the approach. They also note that the reliance on explainability does not compensate for the lack of predictive accuracy. Additionally, the study does not explore the potential of fine-tuning models on domain-specific data, which could yield different results.

Why it matters
This research underscores the limitations of zero-shot financial NLP in practical applications, emphasizing the need for decision-support systems that prioritize transparency and uncertainty awareness. The findings advocate for a shift in focus from mere predictive accuracy to the interpretability of predictions, which is crucial in high-stakes financial environments. This work contributes to the ongoing discourse on the role of explainable AI in finance, as highlighted in related literature (see arXiv). The implications of this study may influence future research directions, particularly in developing models that balance predictive performance with explainability.