sentiment-news

name	sentiment-news
description	Computes ticker/sector news sentiment for DSE stocks and separates rumour-driven items from fundamentals-driven news, since the Dhaka Stock Exchange is rumour-heavy. Use when the user asks about news sentiment, "what's the news saying about GP", headline tone, rumour vs real news, or wants a sentiment score to feed signal-synthesizer.
license	Apache-2.0
compatibility	Prompt-first Agent Skill. Usable with the script absent. Script is Python 3.8+ stdlib-only, no network.
metadata	{"author":"stock-buddy","version":"0.2.0","prd_refs":["PRD-001:REQ-027"],"mode":["momentum","investment"]}

Sentiment & News

Prompt-first skill. Score the headlines by the rules below. scripts/sentiment.py is OPTIONAL (lexicon scorer). The key discipline: separate rumour from fundamentals — DSE is rumour-heavy, so rumours can never drive a strong score on their own.

Role & objective

You produce an aggregate news sentiment score (−1..+1) for a ticker, explicitly down-weighting rumours and surfacing fundamentals-driven items.

When to use

"What's the news saying about GP?", "headline tone", "rumour vs real news", "sentiment score". Feed the score into signal-synthesizer.

Inputs you need

news — array of {headline, source, category}. Categories that count as fundamental: earnings, regulatory, corporate_action, macro. Sources/terms that mark rumour: social/forum/unconfirmed/rumour/speculation/alleged. If news missing → error; if empty → error.

Method (follow in order)

1. Per-headline tone — count positive vs negative keywords; tone = (pos−neg)/(pos+neg), clamp −1..+1 (0 if no keywords).
2. Classify & weight each item: rumour → ×0.3; fundamental → ×1.0; general → ×0.7.
3. Aggregate = weighted average of tones.
4. Rumour discipline: if there are rumours but zero fundamental items (rumour_dominated), clamp the score to [−0.4, +0.4] and flag it.
5. Flag thin_coverage if fewer than 2 items.

Scoring rubric

Rating: score ≥ 0.2 → positive · ≤ −0.2 → negative · else neutral.

Confidence = clamp(0.4 + 0.1·min(fundamental_count,4) + 0.05·min(item_count,4) − 0.2 (if rumour_dominated) − 0.15 (if thin_coverage), 0.15, 0.9).

Output (emit this Thinking Card)

{ "skill": "sentiment-news", "ticker": "..", "mode": "both", "as_of": "..",
  "score": 0.0, "confidence": 0.0, "rating": "positive|negative|neutral",
  "key_metrics": { "item_count": 0, "fundamental_count": 0, "rumour_count": 0,
    "avg_fundamental_sentiment": 0.0 },
  "reasoning": ["[FUNDAMENTAL|RUMOUR|general] headline -> tone"], "flags": ["..."],
  "disclaimer": "Educational analysis only. Not financial advice." }

DSE pitfalls

• Rumour-heavy market: a wave of social/forum buzz must never produce a strong signal — enforce the ±0.4 clamp when no fundamental item is present.
• Thin coverage: one headline is not sentiment — flag and lower confidence.
• Prefer disclosure/earnings/regulatory items; treat anonymous tips as noise.

Optional precision helper

python3 scripts/sentiment.py --input data.json --pretty

Returns the per-item classification and the weighted aggregate.

Worked example

3 items: earnings "record profit, dividend" (+1.0, fundamental ×1.0), regulatory "approval" (+1.0, ×1.0), forum "rumour of merger" (+0.5, ×0.3) → aggregate ≈ +0.93 → positive, confidence ≈ 0.75 (2 fundamental items).

References

Lexicon & rules: references/SENTIMENT.md. Output is educational analysis only, never financial advice.