LLaMA extracts multi-dimensional financial signals beyond sentiment

The paper's actual contribution is narrower than the summary suggests: it shows LLMs can label financial news with structured metadata (event category, impact scope, time horizon) that correlates with stock movement independent of sentiment. This is not sentiment replacement; it's structured annotation on top of sentiment, and the predictive gains come from having multiple independent variables rather than a single polarity score.

This connects directly to the TCA-SIR paper from late July, which argues that systems should extract abstract problem-solving principles rather than surface-level features. Here, researchers are doing exactly that for finance: moving past 'positive or negative' to 'what type of event, how broad, how fast.' The related work on linguistic bias in LLMs (the World Englishes studies from the same week) also matters as a caution: if the model encodes financial reporting conventions from dominant markets, it may miss or misclassify signals from emerging-market news, limiting the generalizability of these extracted features.

If the team publishes out-of-sample backtests on 2026 news-stock pairs (data after the 41,618 training set), and the structured features maintain predictive power without retraining, that confirms the extraction generalizes. If instead performance drops sharply on unseen time periods, the correlations were likely artifacts of the training window rather than structural properties of financial language.