TradingNews, Financial news API with dedup, urgency tags, and sentiment

Hey HN,

I built trading bots for a year. The trading part was easy. The news pipeline almost killed me.

Before my model could see a single headline, I had to: scrape 12 sources, parse inconsistent formats, deduplicate (one Fed decision = 8 near-identical articles from different publishers), classify urgency (rate hike vs. new VP hire), run sentiment, and normalize into clean JSON. Seven steps, each with its own failure modes.

I spent more time maintaining that pipeline than building strategies. So I turned it into an API.

curl "https://api.tradingnews.press/v1/news?urgency=breaking&limit=5" \ -H "X-API-Key: $KEY"

→ { "articles": [{ "content": "FOMC: Fed raises rates 25bps to 5.75%", "urgency": "breaking", "sentiment": -0.42, "received_at": "2026-04-07T14:00:00.034Z" }] }

What makes it different from raw news feeds:

1. Deduplication. When the Fed announces a rate decision, Reuters, AP, Bloomberg, CNBC, WSJ, MarketWatch, and FT all publish within 30 seconds. That's 7 articles for one event. We merge them into one. Three-layer approach: URL dedup → embedding similarity clustering within a time window → entity + event type matching. Layer 2 catches ~80%, layer 3 gets the rest.

2. Urgency classification. Every article is tagged breaking / flash / regular. Trained on ~50K labeled financial headlines, ~94% precision on "breaking." One query parameter eliminates 90% of noise:

GET /v1/news?urgency=breaking

3. Sentiment scoring. Per-article score from -1.0 to +1.0, from a financial NLP model (not a general LLM). 84% agreement with FinBERT. Computed in the same processing pass — no added latency.

4. Speed. Sub-200ms average, 34ms on breaking news like FOMC decisions. Most financial news APIs I tested deliver in 1-5 seconds.

Pricing: $20/mo (REST, 1 req/min) / $50/mo (REST + WebSocket streaming, 30 req/min). 3-day free trial.

What this is NOT:

- Not a news reader or dashboard — API only, JSON in/out - Not a market data provider — no quotes or OHLCV, use Polygon for that - Not enterprise-grade yet — built for solo traders and small quant teams - No historical archive yet (most requested feature, on the roadmap)

Stack: Express/TypeScript, custom NLP pipeline, embedding model for dedup, Postgres, Redis, dedicated infra (not serverless — latency matters).

Docs: https://docs.tradingnews.press Sign up: https://tradingnews.press

Happy to answer questions about the dedup architecture, urgency classifier, or anything else.