A recent arXiv preprint asks a deceptively simple question: which inputs actually drive a language model's financial forecast? To answer it, the researchers constructed synthetic earnings announcements that vary three factors independently.

The first factor is 20-day momentum, meaning the stock's recent price trend. The second is commentary tone that carries no real information—sentiment that should not affect fundamentals. The third is whether the announced actual EPS beats a manipulated earnings threshold, creating a controlled earnings surprise.

By reversing momentum and manipulating the surprise threshold, the authors can separate correlation from causation in the model's predictions. The abstract does not yet report which factor dominates, but the experimental design offers a way to probe the reasoning of LLMs in financial settings.