Local Gemma/Qwen News Triage Prompts (No Cloud Leak)

Description: Leverage quantitative data and automated agent pipelines. Discuss utilizing localized large language models like Gemma or Qwen for market sentiment analysis, news scraping, and algorithmic edge.
Post Reply
NYC88
Posts: 50
Joined: Sat Sep 05, 2026 7:19 am

Local Gemma/Qwen News Triage Prompts (No Cloud Leak)

Post by NYC88 »

I use local models to **sort** headlines, not to invent entries. Cloud LLMs + live positions + paste-everything workflow is how research notes wander into someone else's training set. Call me paranoid. I've been paranoid for twenty years about a lot of things that later got names.

Triage is classification. You don't need poetry. You need: ticker, event type, hard numbers, and a ruthless "this is rumor" flag.

### Use case

Rank a pack of premarket headlines into: actionable / noise / needs primary source. Extract ticker, event type, time, hard numbers. Refuse price targets and "guaranteed" direction. If the model starts coaching trades, your prompt is wrong.

### System prompt (local)

Code: Select all

You are a trading news triage assistant for US equities.
Output ONLY valid markdown with these sections:
- Ticker
- EventType (earnings|guidance|offering|FDA|lawsuit|macro|rumor|other)
- TimeET
- HardFacts (bullets, numbers only if present in text)
- Unknowns (what is missing)
- Actionability (LOW|MED|HIGH) with one-line reason
Rules:
- Do not invent numbers.
- Do not give trade directions.
- If rumor language appears, mark EventType=rumor and Actionability=LOW.
### User prompt template

Code: Select all

Headline pack (paste raw):
"""
{headlines}
"""
Watchlist filter (only these tickers matter): {TICKERS}
Session: premarket | RTH | AH
### Operator rules (mine)

1. Model suggestion ≠ order. Ever.
2. HIGH actionability still needs float/SI/context. The model doesn't know your book.
3. Keep it offline if notes include account size or broker identity.
4. Re-run on the **primary source** link before you size. Headlines lie by omission.

Can't run a big local box? Quantized Qwen/Gemma builds are enough for triage. You're sorting mail, not writing a novel.

What local model/size are you actually running for headline triage?
Anyone tried this and found the model inventing EPS numbers anyway — what guardrail fixed it?
Do you filter to watchlist tickers first, or triage the whole firehose?
Primary-source step — do you force a click-through before any size, or is that aspirational?
Post Reply