news-scanner

name	news-scanner
description	Surface the day's news events that actually moved a stock. For each notable headline across a watchlist (or the broader market), render a Bloomberg news tape / Benzinga Pro-style stream with sentiment, novelty, and the post-publish price reaction. Ranked by impact, capped at top N (default 15-20). The 6am sell-side morning-note prep workflow.

news-scanner

You hand over a watchlist and a time window. The skill pulls every news event Massive has on those tickers in the window, derives a sentiment score per ticker per article, measures whether the angle is novel or a re-run, computes the stock's price reaction since publish, ranks events by impact, and emits a stream of the top N.

This is the workflow a sell-side analyst runs at 6am to write the morning note. Twenty headlines that actually moved a stock, with the context (sentiment + novelty + reaction + volume anomaly + divergence flag) to write about them in 30 minutes. Unlike a news terminal or RSS reader, news-scanner ranks events by signal quality (price reaction × volume anomaly × novelty) rather than recency, and ships the methodology with the output.

When to invoke

• An analyst is prepping the morning note and wants the overnight tape ranked by impact
• A PM is asking "what's the news on my book today"
• The user says "scan news on NVDA TSLA AAPL", "what moved overnight", or "any catalyst on my watchlist"
• A trader wants to spot price/news divergence (negative headline, positive reaction = bad news already priced in)

What you need

• A watchlist of tickers (default: NVDA, TSLA, AAPL, SPY, META, NFLX)
• A time window in hours (default: last 24h)
• MASSIVE_API_KEY exported in the environment
• Stocks Basic + Benzinga News add-on minimum

The skill runs at two fidelity tiers.

• Tier A (Benzinga sentiment + minute aggs): Benzinga News add-on returns per-ticker insights[] with a categorical sentiment label ("positive" / "negative" / "neutral") and sentiment_reasoning from Benzinga's own NLP. Stocks Starter or higher gives reliable minute aggregates for the reaction window. This is the default tier.
• Tier B (keyword fallback): Benzinga News add-on missing or the Benzinga insights field is empty. Sentiment falls back to a keyword scorer (positive: beat, raise, partnership, upgrade; negative: cut, miss, lawsuit, downgrade, recall, probe). Reaction calc still works on Stocks Basic but uses 5-minute aggregates instead of 1-minute. Documented in references/sentiment-scoring.md.

What you get back

Two output layers from one analysis.

Layer 1: canonical JSON matching output-schema.json. Per-event fields: ticker, published_at, source, headline, url, sentiment_score (in [-1, +1]), sentiment_source ("benzinga" or "keyword"), novelty_score, novelty_band, reaction_pct_since_publish, reaction_window_label, volume_anomaly_x, divergence_flag, context_line. UIs, alert pipelines, and downstream agents consume this.

Layer 2: rendered stream in Bloomberg news-tape / Benzinga Pro style. Three lines per event, optional ↳ continuation for context. Format rules in references/rendering.md. Compact, scanable, key:value pairs. Claude Code users read this.

How it works

1. For each ticker in the watchlist, pull /v2/reference/news?ticker={t}&published_utc.gte={window_start}&limit=50. Dedupe by article_url across the merged set so a story syndicated across publishers only appears once. See references/news-sources-and-coverage.md.
2. Score sentiment per (ticker, article). Prefer the Benzinga insights entry for that ticker if present (map "positive" → +0.7, "neutral" → 0, "negative" → -0.7); otherwise fall back to a keyword scorer over the title and description. See references/sentiment-scoring.md.
3. Score novelty per (ticker, article). Bucket the last 7 days of articles for the ticker; compute TF-IDF over titles + first-sentence of description; cosine distance to nearest neighbor in the bucket. Distance > 0.6 = high novelty (new angle), 0.3-0.6 = medium, < 0.3 = low (already covered). See references/novelty-detection.md.
4. Compute the price reaction. Pull /v2/aggs/ticker/{ticker}/range/5/minute/... from publish minute through min(publish + 60 minutes, market close). Reaction % = (close at window end / close at publish minute) - 1. Volume anomaly = avg per-minute volume during the window / prior-5-day same-time-of-day average per-minute volume.
5. Flag price/news divergence per references/price-news-divergence.md: positive sentiment + negative reaction = priced in / sell-the-news; negative sentiment + positive reaction = bad news already priced in.
6. Rank by impact = |reaction_pct| × volume_anomaly × novelty_score. See references/impact-ranking.md. Emit top N (default 15).

Foundations used

• massive-api-patterns for REST auth, the best-price fallback chain for spot, and rate-limit handling on the per-ticker news fan-out

Output mode: stream

Stream mode is the format Bloomberg's news tape, Benzinga Pro's feed, and Reuters Eikon use for incoming events. Each event is a self-contained block; the reader scans top to bottom and stops when they see one they want to act on. Inherited from options-flow/references/rendering.md, adapted for news per references/rendering.md.

Endpoints used

• GET /v2/reference/news?ticker={t}&published_utc.gte={iso}&limit=50: Benzinga News. Returns results[] with id, title, description, published_utc, article_url, tickers[], keywords[], publisher.name, and a critical insights[] array with per-ticker sentiment ("positive"/"negative"/"neutral") and sentiment_reasoning.
• GET /v2/aggs/ticker/{ticker}/range/5/minute/{from}/{to}: minute (or 5-minute) aggregates for the underlying. Used to compute reaction % and volume anomaly post-publish.
• GET /v2/snapshot/locale/us/markets/stocks/tickers/{ticker}: spot fallback chain for the "spot at publish" reference when minute aggs are missing or stale.

Doesn't handle (yet)

• Sector / peer reaction. A positive NVDA story usually moves AMD and AVGO too; the skill doesn't surface sympathy plays. Clean v2.
• Wire-service deduplication beyond URL match. Reuters → Bloomberg → CNBC rewrites of the same story have different URLs and different first sentences; TF-IDF catches most but not all. A more rigorous story-clustering pass (LSH or sentence embeddings) is a v2 candidate.
• Real-time WebSocket streaming. v1 is REST-polled. The massive-websockets foundation covers the live-stream pattern for a future variant.
• Insider transactions, SEC filings, and FDA calendar items. These are catalyst-class news that don't ship through the Benzinga News feed; they live on the corporate-actions and reference endpoints. Separate skill.

These are clean PR extensions and welcome contributions.