SenFlow
Overview
n8n pipeline that gathers forex news and YouTube market coverage for ten instruments, filters it for relevance, has LLM analysts label each item Bullish, Bearish, Neutral or Mixed, and writes per-instrument sentiment reports.
- Year
- 2025
- Status
- Ongoing
- Role
- Builder
- Kind
- Forex Sentiment Automation
- Domain
- Markets
- Terminal command
/senflow
What it does
- Covers XAU/USD and nine currency pairs, from AUD/USD to USD/JPY
- Two-tier LLM analysis: a junior pass lists bullish and bearish factors, a senior pass sets the label, confidence and conviction
- 553 sentiment reports between July and December 2025, plus 29 weekly COT and 26 daily calendar briefings
- Every run logged per trading session (Asia, London, New York, After Hours) for validation
From headlines to a sentiment report
SenFlow is an n8n pipeline that reads the news for ten instruments, XAU/USD and nine currency pairs, and writes a sentiment report for each one. Between July and December 2025 it produced 553 reports.
As the builder, I work across the whole pipeline: collection, relevance filtering, the two LLM analyst passes, report generation and the run logs.
- AUD/USD
- EUR/JPY
- EUR/USD
- GBP/JPY
- GBP/USD
- NZD/USD
- USD/CAD
- USD/CHF
- USD/JPY
- XAU/USD
Read more: Self-hosted automation stack
One run for one instrument
The path a news cycle takes through the n8n workflows, from headline to report.
Collect
Headlines and snippets are recorded with a timestamp; YouTube market-news transcripts are pulled alongside.
Relevance filter
An LLM marks each headline RELEVANT or IRRELEVANT for the instrument; one EUR/USD run screened 99 articles.
Full text
Relevant articles are scraped, cleaned and summarised.
Junior analyst
Bullish and bearish factors, each with a strength and a source type.
Senior analyst
Final label, confidence, conviction and reasoning.
Report
Gemini 2.5 Pro writes the report: label counts, weighted confidence, agreement index, dominant sentiment and a market narrative.
One report: EUR/USD, 31 October 2025
42 articles and 95 YouTube items were labelled. Both leaned bullish (21 and 45); the report's dominant sentiment was Bullish at 80% average confidence, with a 48% agreement index and Low strength.
- Bullish
- Bearish
- Neutral
- Mixed
Counts from the report's executive summary.
Show the numbers
| Bullish | Bearish | Neutral | Mixed | Total | |
|---|---|---|---|---|---|
| Articles | 21 | 10 | 2 | 9 | 42 |
| YouTube | 45 | 30 | 7 | 13 | 95 |
Two analysts, one label
Each relevant article goes through two LLM passes. The junior analyst lists bullish and bearish factors, each with a strength (High, Medium or Low) and a source type such as Economic Data, Fed Policy Action or Geopolitical Tension. The senior analyst weighs those factors and sets the final label, Bullish, Bearish, Neutral or Mixed, with a confidence, a conviction level and its reasoning.
YouTube market-news transcripts get the same treatment, so each report compares what the articles say with what market TV says.
Briefings around the reports
Two side streams give the reports context. A weekly COT analysis turns the CFTC's Commitments of Traders data into net positions, weekly changes, longs and shorts per trader category for gold, silver, the US Dollar Index and the major currencies (29 reports). A daily forex calendar lists the day's events from Forex Factory with impact, forecast and previous values, and explains what each one measures, why it matters and how the market tends to react (26 briefings).
Built to be checked
Every run is logged as JSONL by date and trading session (Asia, London, New York, After Hours): the inputs, both analysts' outputs and the YouTube predictions. A validator compares predicted labels with reference labels and writes a confusion matrix, an accuracy figure and the misclassified items.
No accuracy is quoted here: the validation samples logged so far are small.
Read more: Related reading · forex robots paper (2024)
References
- Self-hosted automation stack →
The n8n infrastructure the workflows run on.
- Related reading · forex robots paper (2024) →
Its limitations note that the trading robots it tested do not read market sentiment.
- Automation experience →
Personal automation work and sentiment workflow development.
Stack
- Backend & APIs
- Webhooks
- AI & vision
- LLM APIsGemini 2.5 Pro
- Automation
- n8n