AI-Driven BIST100 Financial Analysis System

A comprehensive stock analysis platform using LightGBM and LLMs for signal generation and news sentiment analysis.

This project was a comprehensive freelance undertaking to build an automated financial analysis system targeting the BIST100 stock exchange. The goal was to create a “Smart Advisor” that not only predicts stock movements based on technical indicators but also understands the market context through news analysis.

Key Achievements

Although the project was developed for a startup initiative, the technical outcomes demonstrated significant predictive power:

  • High-Performance Modeling: Developed a LightGBM model that achieved a Weighted F1-Score of ~0.67.
  • Risk/Reward Success: In backtests, the model demonstrated a 0.97 Sharpe Ratio with approximately 230% return, significantly outperforming the benchmark.
  • Intelligent News Analysis: Integrated DeepSeek LLM to analyze financial news in multiple languages (TR, EN, DE, KO). The system automatically summarized news and assessed its potential impact on specific stocks.

Technical Architecture

The system was built with a microservices approach to ensure scalability:

  1. Backend: Developed using FastAPI for high-performance, asynchronous API endpoints.
  2. Database: Used PostgreSQL for robust transactional data storage.
  3. Deployment: Containerized the entire application using Docker for consistent deployment environments.
  4. MLOps: Implemented comprehensive documentation including Model Cards and data dictionaries to ensure reproducibility.

Core Features

  • Automated Buy/Hold/Sell Signals: Based on technical analysis and ML predictions.
  • Customizable News Feeds: Users can manage their own news sources for personalized insights.
  • Chatbot Interface: A multilingual chatbot integrated with the backend for user interaction.
The project showcases the integration of classical Machine Learning (LightGBM) with modern Generative AI (LLMs) in a FinTech context.