Fin-LSE

AI powered LSE stock intelligence where we can screen stocks, detect chart patterns automatically, and get LLM enriched news with per stock impact mapping, all with explainable scoring.

Comment

Inspiration

Traditional stock screeners force traders into rigid filter UIs with dropdowns and sliders. We wanted to build something that feels like talking to a financial analyst , explainable results instantly. The London Stock Exchange, with its 350 listed companies, was the perfect sized universe to prove this out.

What it does

Fin-LSE is a full-stack fintech platform that delivers explainable trade insights for LSE-listed stocks through three intelligent modes:

  • Natural-Language Screener : Ask questions like "Show me stocks with ROE above 20% and PE below 15" and a LangGraph multi-agent system parses your intent, builds SQL, queries a live fundamentals database, and returns ranked results with a human-readable summary.
  • Technical Pattern Scanner :Automatically detects Golden Cross, Death Cross, RSI Oversold/Overbought, Near Breakout, Volume Surges, and MACD crossovers across all LSE stocks every 30 minutes. Users can also describe custom strategies in plain English.
  • Market Intel Feed : Ingests RSS feeds from BBC Business, CNBC, Google News, MarketWatch, and Yahoo Finance, deduplicates articles into unique stories, and uses GPT-4o-mini to extract sentiment, entities, and per-stock impact mappings.

Every stock also gets a 0–100 Opportunity Score and 0–100 Risk Score, broken down into named, auditable drivers (e.g., "Strong ROE +12", "High Volatility −8") so traders understand why a stock scores the way it does.

How we built it

The backend is a FastAPI application orchestrating multiple LangGraph state-machine agents:

  • A Supervisor agent classifies user intent and routes to specialist sub-graphs.
  • A Screener agent uses GPT-4o-mini with Pydantic structured output to convert natural language into validated SQL filters.
  • A Market Feed pipeline (fetch → deduplicate → LLM-enrich → persist) runs on a 1-hour auto-ingest loop.
  • A Scoring engine combines Yahoo Finance fundamentals and 6-month price data (RSI, MACD, moving averages, volatility) into weighted composite scores.

All data is stored in SQLite (WAL mode) for zero-config deployment. The frontend is a Next.js 16 app with React 19, Tailwind CSS 4, and Recharts for interactive OHLCV charts. Firebase Authentication handles Google sign-in.

Challenges we ran into

  • LLM output reliability : Getting GPT-4o-mini to consistently produce valid SQL filter structures required careful prompt engineering and Pydantic structured output with retry logic.
  • RSS deduplication : Different news sources often cover the same story with different headlines. We built a similarity-matching pipeline to group articles into unified stories before LLM enrichment.
  • Rate limiting : Yahoo Finance throttles requests heavily. We had to implement batching, delays, and caching to scan ~350 tickers without getting blocked.
  • Real-time feel : Balancing background data freshness (pattern scans every 30 min, news ingest every hour) with API response speed required careful async scheduling.

Accomplishments that we're proud of

  • A unified prompt bar that auto-detects whether you're asking a screening question, describing a technical pattern, or searching for news — no mode switching needed.
  • Fully explainable scores : every number has a traceable driver, making the system auditable rather than a black box.
  • The entire stack runs locally with just Python, Node.js, and an OpenAI API key : no cloud infrastructure needed.

What we learned

  • LangGraph's state-machine approach to multi-agent systems is far more controllable than chaining LLM calls : each node has a clear contract and errors are recoverable.
  • Structured output (Pydantic models as LLM output schemas) dramatically reduces parsing failures compared to free-form JSON extraction.
  • Building for explainability from day one (named drivers with impact values) makes the product fundamentally more trustworthy.

What's next for Fin-LSE

  • Real-time WebSocket streaming for live price updates and pattern alerts.
  • Portfolio tracking : let users build watchlists and get personalized scoring alerts.
  • Expand to more exchanges : the architecture is exchange-agnostic; adding NYSE/NASDAQ is a configuration change.
  • Fine-tuned models : replace GPT-4o-mini with fine-tuned open-source models for lower latency and cost.

Built With

Share this project:

Updates