Earnings releases · semantic extraction
AI reads earnings releases. What does the model learn?
A local Qwen3-4B LLM installation analyses 6,493 official earnings releases filed with the SEC. No future share-price information is included in the model prompt.
Current status
The analysis runs on the research computer.
Each completed release is saved immediately. The public counter is updated approximately every twenty minutes, and processing can resume after an interruption.
- Processed
- Loading
- Model
- Qwen3-4B LLM
- Current phase
- —
- Latest stock
- —
- Time remaining
- —
- Last updated
- —
Study design
What is recorded for each release?
Financial developments
Revenue, earnings, margins, demand, costs and explicit comparisons with a previous period.
Guidance and expectations
Upgrades, downgrades, withdrawals and the certainty with which management discusses the outlook.
Risks and uncertainty
Material headwinds, restructuring, one-off items and operational constraints.
Balance sheet and distributions
Investment, cash position, share repurchases, dividends and changes in capital allocation.
Verifiable quotations
Each non-neutral score is accompanied by a short, verbatim supporting excerpt from the source release.
Confidence and ambiguity
The model flags passages that are insufficient or contradictory and therefore do not support reliable extraction.
Two phases
Develop first, then test independently.
Original dataset
607 releases from the existing research sample. This phase allows comparison with the earlier TF-IDF study.
Replication dataset
5,886 releases from new companies and time periods. This is the decisive test of generalisability.
An important distinction
Progress is live. Conclusions are not.
No interim effects are published as findings during processing. Evaluation takes place only once both phases are complete and the prespecified final test has been carried out.