2026 · Python · private repository
quantnews
A private prototype for organizing and evaluating news and filing hypotheses.
- Finding
- The prototype organizes five text-event sources into a ranked research list and an outcome log.
- Scope
- A private S&P 100 research system with historical score buckets and a walk-forward evaluation workflow.
- Boundary
- The public page does not publish the target, split dates, costs, baseline, or results, so predictive performance cannot be judged publicly.
What exists
Many company-specific catalysts first appear in text disclosures, including news, SEC filings, and earnings calls. This prototype collects those events, classifies them, and records subsequent outcomes for later evaluation.
Evaluation workflow
A daily job collects RSS, SEC EDGAR 8-K items and Form 4 insider trades, earnings-call transcripts, and analyst revisions, and an LLM assigns each one a direction and a model score.
The research workflow groups model scores by event type, direction, and score band, then records the historical outcomes in each group. Explicit regime and event-window rules filter the resulting ranked research list. Tracked theses record an entry, stop, target, and close rationale.
RSS · SEC 8-K · Form 4 · Transcripts · Analyst revisions
- 01
LLM scoring
one signal per event
- 02
Historical bucketing
event type · direction · score band
- 03
Explicit filters
regime · event windows · lockouts
- 04
Ranked research list
entry / stop / target fields
- 05
Tracked theses
recorded, then reviewed on close
↺ closed theses add outcomes to the historical record.
Historical outcome buckets
A raw model score is not treated as a probability. Each (event type, direction, score bucket) is paired with recorded historical outcomes and pulled toward the base rate when the sample is thin. This is exploratory bucketing, not a calibrated probability mapping.
What is not published
The system has a walk-forward evaluation workflow, but the target definition, split dates, costs, baseline, and results stay private. Until a sanitized methods-and-results artifact is public, this is a process demonstration rather than a trading-performance claim.