Early in my career I chased "great news" into a name that was already distributing. Felt smart for about four minutes. The tape had the memo. I had a headline.
### Working model (simple on purpose)
1. **Price/volume regime first** — trend, range, gap state.
2. **Text second** as a catalyst classifier.
3. **Position third** only if 1 and 2 agree with the playbook.
### When text leads (sometimes)
Fresh, specific, numbered primary-source news into a quiet tape. Float small enough that forced repositioning shows in L2. Premarket discovery before the wider audience — still not magic, just a head start measured in minutes.
### When text lies (often)
Recycled "analyst says" into an already-extended move. Sentiment models trained on social spam. Headline timestamps that lag the IR filing by enough to matter. Mixed bag: beat + cut guide — model says "positive," tape says "repricing."
### Measurement habit
For 20 events, log:
Code: Select all
t0_headline_et, t0_price, t+5m, t+30m, regime_tag, text_score, outcome_tag
Sentiment is a feature. Features need regimes. Regimes need risk caps. That's the whole sermon.
Do you score sentiment numerically, or just tag catalyst type?
In which regime has text actually helped you — quiet PM, RTH trend, something else?
Anyone measure headline timestamp vs filing timestamp and find a consistent lag?
When text and tape disagree, do you have a hard rule — or do you negotiate with yourself?