AI Stock News Predictor
Can AI predict the stock market from the news?
The system behind my master's thesis (grade 1.0): six language models read financial news in three analyst roles and forecast five indices. Then I measured whether it actually works.
- Thesis grade
- 1.0
- Directional accuracy, chance level 50%
- 35.0%
- Cells: 6 models × 3 roles × 5 indices
- 90
- Random portfolios as benchmark
- 10,000
- Role
- Master's thesis · Karlsruhe University of Applied Sciences
- Period
- 2026
ai-stock-news-predictor.vercel.app
ai-stock-news-predictor.vercel.app
ai-stock-news-predictor.vercel.app
ai-stock-news-predictor.vercel.app
ai-stock-news-predictor.vercel.app
ai-stock-news-predictor.vercel.app
The problem
Plenty of people claim LLMs can read markets from the news. I wanted to measure it: across several models, roles and markets, without task-specific retraining and against 10,000 random portfolios.
What I built
- A Python pipeline that collects financial news and has six models from two providers (GPT-4o, GPT-4o mini, o3-mini, Claude Sonnet, Opus, Haiku) forecast every day.
- A full factorial design: 6 models × 3 analyst roles (conservative, neutral, aggressive) × 5 indices (S&P 500, DAX, MSCI World, MSCI EM, Nikkei 225) = 90 cells.
- Evaluation with a binomial test and a trading simulation with realistic costs, against buy-and-hold and 10,000 random portfolios.
- A bilingual results dashboard in React and Recharts, plus the colloquium deck.
Results
- A clear answer for anyone betting on AI stock tips: 35.0% directional accuracy, below the 50% chance level.
- Every number traceable in a public dashboard, from headline to forecast.
- Thesis graded 1.0.
Tech stack
Stack
- Python
- OpenAI
- Anthropic
- React
- Vite
- Recharts
- Convex
- Vercel