Gemma 4 spots earnings-call signal

The useful part here is not the alpha claim, it’s the workflow: LLMs can turn messy qualitative filings into structured features fast enough to be practical, but only if you treat every “signal” like a hypothesis and stress-test it against known factors.

• –The real signal was modest but cleaner: CFOs getting vague in outlook language lined up with ~1.8% underperformance over 5 days and low factor overlap.
• –The ghost signal is a reminder that tone models are especially prone to regime contamination; sector beta can masquerade as linguistic confidence very easily.
• –Local inference is the right setup for this kind of work because earnings calls are exactly the sort of proprietary, compliance-sensitive corpus you do not want leaving your environment.
• –Fine-tuning on 800 labeled transcripts is enough to find candidate features, but not enough to trust them without factor regressions, sector controls, and walk-forward validation.
• –The Q&A section is the more interesting next target because it is less rehearsed and more likely to expose real uncertainty than polished prepared remarks.