Gemini-Powered Stock Analysis: Parsing Financial News for Trading Decisions (pgaleone.eu)

A developer built a production Go service that ingests Italian finance RSS feeds, extracts article bodies, and uses Google’s Gemini (Vertex AI GenAI) to output structured buy/sell/hold recommendations for detected tickers. The system follows a producer‑consumer architecture (Stream + AIAnalysisService) that fetches feeds concurrently, parses site‑specific HTML to isolate article text, sends a low‑temperature (0.1) JSON-only prompt to Gemini, and routes recommendations into long/short trading channels after normalizing ticker variants (e.g., STLAM vs STLA.MI). It integrates with a broker API and trading strategies (eg. trend following) to automatically act on AI signals. This project highlights practical tradeoffs for putting LLMs into high‑velocity, high‑stakes workflows: brittle but token‑efficient HTML extractors versus the robustness and higher token cost of feeding full pages to the model; explicit prompt engineering (strict JSON, multilingual handling, ticker format instructions) to reduce hallucinations; and operational concerns like timeouts, retries, concurrency, caching, and batching to control latency and Vertex AI costs. Planned enhancements—sentiment scoring, multi‑model comparisons, backtesting, and portfolio‑aware logic—underscore the broader AI/ML implication: productionizing LLMs requires not just model calls but robust data extraction, strict output contracts, cost-aware design, and rigorous validation before automating trades.