apify-financial-news

Financial News Intelligence

Discover and extract financial news for portfolio companies via Apify Actors. Two modes:

• Single company — news scan for one company.
• Portfolio scan — same pipeline run across multiple companies.

33 Tier 1 sources organized in 4 categories (Global / Pan-European / Institutional / CEE Local). Tier 2 = broad Google News fallback for unverified domains.

Estimated cost: $0.10–0.50 per single company, $1–5 per full portfolio scan.

Prerequisites

• Apify access — preferred: apify CLI (npm install -g apify-cli && apify login); fallback: Apify MCP connector (call-actor tool). CLI is faster and preferred when both are available.
• Python 3 + pip install readability-lxml lxml (for the extract_and_clean.py post-processor)
• Companies data at ${CLAUDE_PLUGIN_ROOT}/data/companies.json

${CLAUDE_PLUGIN_ROOT} is the plugin's root directory (where .claude-plugin/ lives). It is resolved automatically by Claude Code when the plugin is installed, or set to the --plugin-dir path during development.

Workflow checklist

Copy this and tick boxes as you progress:

Task Progress:
- [ ] Step 0: Verify prerequisites — try `apify --version && apify info`; if unavailable, check for `call-actor` MCP tool; if neither, tell user to install apify CLI or Apify MCP connector. Also verify: `python3 -c "from readability import Document; print('OK')"` (install: `pip install readability-lxml lxml`)
- [ ] Step 1: Build queries (look up company in data/companies.json or construct manually)
- [ ] Step 2: Discovery — pick 8-12 sources by region, run 2-phase Google News
- [ ] Step 3: Dedup + route (Tier 1 = whitelisted domain → verified extractor; Tier 2 = broad → rag-web-browser)
- [ ] Step 4: Extract & clean (run extractor, then extract_and_clean.py)
- [ ] Step 5: Output Tier 1 + Tier 2 tables to user

Constraints

Allowed Apify Actors (exhaustive — do NOT use others)

Do NOT use any other actor. Do NOT use WebSearch, WebFetch, or browser tools.

Extractor routing (mandatory — do NOT substitute)

Full per-source config: reference/SOURCE_CONFIGS.md [blocked]. Machine-readable: data/sources.json [blocked].

Pipeline

Step 1: Build queries

Look up company in ${CLAUDE_PLUGIN_ROOT}/data/companies.json. Key fields under queries: gnews_en, gnews_cz, bloomberg.

For non-portfolio companies, construct manually: quoted full legal name + ticker OR variant + geographic qualifier.

Query rules:

• Use "InPost SA" not InPost (quoted full names avoid false positives — see reference/EUROPEAN_COMPANIES_GUIDE.md [blocked])
• Per-source site: operator: site:bloomberg.com "InPost SA" OR "INPST"
• FT tip: use site:ft.com/content/ (bare ft.com returns stock-data pages)
• Valid timeframes: "1h", "1d", "7d", "1y", "all" (NOT "30d")
• Tickers < 4 chars: always pair with full company name
• ALWAYS set decodeUrls: true on Google News input

Step 2: Discovery

Do NOT search all 33 sources. Pick 8–12 based on company region.

Regional priorities

Two-phase Google News strategy

Phase 1 — Targeted (per priority source, with site:):

apify

Phase 2 — Broad (always run, no site: operator). Classify results by domain in Step 3 — whitelisted → Tier 1, others → Tier 2.

CEE local-language discovery: For CEE companies, run additional queries with region_language set to CZ:cs, PL:pl, HU:hu, BG:bg, SK:sk. See region_language field per source in data/sources.json [blocked].

EC Press Corner — direct REST API (no Actor)

Parse IP_XX_NNN reference IDs from Google News titles to construct API calls.

RSS / Headline discovery (optional)

For sources with RSS, you can supplement GNews with louvre/rss-news-aggregator (max 10 feeds per run; split into batches). For 6 sources, rodrigo_pacelli/headline-news-scraper works (CNBC, SCMP, Nikkei, Caixin, Zawya tag-pages-only, Handelsblatt). RSS/headline output is unfiltered — filter client-side by company name in title/description. Full feed list: reference/PIPELINE_DETAIL.md [blocked].

Step 3: Dedup & route

1. Collect URLs from all discovery runs.
2. Classify by domain: whitelist (33 Tier 1 sources) → Tier 1; other → Tier 2.
3. Filter non-article URLs: /quote/, /stock/, /sitemap, /author/, /tag/, /key-metrics/, /newsletters/, /topic/, /profile/, redirectUrl=.
4. Filter by company name/ticker in title.
5. Deduplicate (URL normalize).
6. Route Tier 1 URLs to verified extractor per routing table.
7. Route Tier 2 URLs to apify/rag-web-browser.

Low-coverage fallback (< 3 articles): broaden timeframe to "1y", run Phase 2 if skipped, try local-language queries.

Step 4: Extract & clean

Run the extractor per the routing table. For rag-web-browser calls, use outputFormats: ["html"].

After extraction, run extract_and_clean.py on the dataset to strip nav/menus/footers via readability-lxml. The script auto-detects format: HTML cleaned via readability-lxml, already-clean output (Bloomberg scraper, workhard3000) passes through.

The --json flag returns run metadata including defaultDatasetId. Errors produce non-zero exit code so the pipeline fails fast.

Step 5: Output

Two tables to the user — Tier 1 (verified) and Tier 2 (broad). Per article: source, title, author, date, char count, URL.

Macro context

For country-level economic context, use these alongside news scans:

• ING Think (think.ing.com) — daily CEE FX/rates via apify/rag-web-browser (~11K chars). Best free open-access CEE macro source.
• IMF (imf.org) — Article IV concluding statements via apify/rag-web-browser (~22K chars).
• ECB (ecb.europa.eu) — already in Tier 1.
• Central banks (ČNB, NBP, MNB, BNB, NBS) — direct URL extraction; see reference/MACRO_SOURCES.md [blocked].

Critical gotchas

• ALWAYS decodeUrls: true in Google News (encoded redirects break ALL extractors).
• "InPost" is ambiguous — matches "post-Maduro". Use "InPost SA" for precision.
• RSS max 10 feeds per run — split into batches.
• Zawya headline-scraper returns tag pages, NOT articles — exclude from headline runs.
• Reuters needs RESIDENTIAL proxy — without it, returns 386 chars.
• Reuters: rag-web-browser returns 0 chars on ~60% of URLs — use workhard3000 only.
• Morningstar .co.uk URLs fail with workhard3000 — use .com URLs for extraction.
• WSJ livecoverage pages fail — skip URLs matching wsj.com/livecoverage/.
• FT: rag-web-browser returns 16 chars ('Client Challenge') — use workhard3000.
• Forbes: workhard3000 returns 0 chars — use rag-web-browser.
• Barron's / MarketWatch cost $0.25/page — use selectively for high-value articles.
• Caixin URLs must be complete — truncated URLs fail extraction.
• EC Press Corner is an Angular SPA — rag-web-browser returns 0 chars. Use REST API.

Full failure-mode catalog: reference/PIPELINE_DETAIL.md [blocked], reference/SOURCE_CONFIGS.md [blocked].

Reference

• data/sources.json [blocked] — machine-readable: 33 sources × discovery × extractor × cost × gotchas
• reference/SOURCE_CONFIGS.md [blocked] — per-source curl examples, run IDs, output samples
• reference/PIPELINE_DETAIL.md [blocked] — RSS feeds, headline sources, warnings, failed paths, source quick-reference table
• reference/MACRO_SOURCES.md [blocked] — ING Think, IMF, central banks
• reference/EUROPEAN_COMPANIES_GUIDE.md [blocked] — European company handling tips
• reference/MORNING_NOTE_DETAIL.md [blocked] — HTML output spec (archived; out of scope for this skill)
• reference/scripts/extract_and_clean.py [blocked], reference/scripts/clean_article.py [blocked] — readability-lxml post-processors