Data provenance
The data behind the AI research.
The source, time boundaries and role of each dataset are defined in advance. This makes clear what information the model could have known and which data are reserved exclusively for subsequent evaluation.
Official earnings releases
SEC documents linked to a specific publication date and a publicly listed company.
Releases from 30 companies
The original study used to develop the extraction process, variables and models and test them chronologically.
Releases in the replication set
5,886 releases from a preselected universe of 300 companies; the number of releases is not the number of companies.
Sources
Three strictly separated layers of information.
Company documents
Official SEC 8-K earnings releases and their text exhibits. Qwen3-4B LLM receives only selected passages from these sources.
Price data
Daily prices, volume, volatility and momentum are used for baselines and for returns after publication.
Market and sector
Broad market and sector benchmarks adjust price reactions for movements that are not company-specific.
Variable register
What is recorded for each earnings release?
| Group | Examples | Role |
|---|---|---|
| Identification | Ticker, SEC accession, acceptance timestamp | Unique linkage and timestamp verification |
| Financial developments | Revenue, earnings, margin, demand | Scores from −2 to +2 |
| Outlook | Guidance, management confidence | Explicit increases, decreases or uncertainty |
| Risk | Restructuring, customer, supply-chain and legal issues | Material changes in headwinds |
| Capital | Liquidity, one-off items, dividends and share repurchases | Balance sheet and capital allocation |
| Quality | Evidence quotations, confidence, ambiguity | Traceability checks |
| Outcome | Abnormal returns after 1, 5 and 20 days | For evaluation only, never included in the Qwen prompt |
Qwen3-4B LLM is not shown future price information. Text extraction and return evaluation remain technically separate.
Chronological splits
Development without inspecting the final test.
Training
Historical observations used to estimate model coefficients.
Validation
A later period used to choose models, thresholds and settings.
Holdout
The untouched time period reserved for final evaluation.