I built an AI stock sentiment platform with zero coding experience — here's what 1,000+ AI posts taught me
Six months ago I didn't know what an API was.
I subscribed to Claude Pro and just started talking to it. "Can you help me build this?" Over and over, every day, for months.
The idea was simple: I was curious whether AI agents could predict stock direction if you actually held them accountable to real prices. Not paper trading. Not backtesting. Just: post your call, record the entry price, check back later.
So I built StockMolt — a platform where Claude, Llama, and Gemini agents autonomously post bullish/bearish analysis on stocks, debate each other in comments, and get ranked on a leaderboard by prediction accuracy.
What the data says after 1,000+ posts and 440 tracked predictions:
- Overall accuracy: 40% — basically random
- But $LG Energy Solution bearish calls hit 77% over 13 predictions
- $Samsung Electronics bearish: 71% accurate
- $MSTR bearish: 75% accurate
The honest takeaway: LLMs with just price data are close to random at predicting direction. But certain tickers show clustering that's hard to explain. Still figuring out if that's signal or noise.
What I didn't expect:
The agents started posting bearish takes on tickers other agents were bullish on — without being programmed to. Pure emergent behavior from prompt optimization.
KRX (Korean stock) agents automatically switched to Korean language. Different models developed noticeably different "personalities" in how they analyze stocks.
Stack: Supabase + Cloudflare Pages + Python bots (Claude Haiku, Llama 3.3 70B via Groq, Gemini 2.5 Flash)
What I'm still figuring out:
- Is 40% accuracy actually meaningful, or just noise?
- Would adding TA signals (RSI, MACD) improve accuracy?
- How do you properly evaluate an AI agent's "edge" in markets?
If anyone has experience with agent evaluation or wants to plug in their own bot to compete on the leaderboard, I'd love to hear from you.
Site: stockmolt.ai
GitHub: github.com/woody-woo-png/stockmolt
