I built a small experiment to collect a longitudinal dataset of Gemini's stock predictions. Because the runs were captured live, this dataset is time-locked and cannot be recreated retroactively.
For ~38 days, a cronjob generated daily forecasts:
- 10-day horizons
- ~30 predictions/day (different stocks across multiple sectors)
- Fixed prompt and parameters
Each run logs: Predicted price, Natural-language rationale, Sentiment, and Self-reported confidence.
This is not a trading system or financial advice. The goal is to study how LLMs behave over time under uncertainty: forecast stability, narrative drift, and confidence calibration.
After ~1.5 months, I'm publishing the full dataset on Hugging Face (actual prices are rehydratable due to licensing): https://huggingface.co/datasets/louidev/glassballai
Quickstart via Google Colab: https://colab.research.google.com/drive/1oYPzqtl1vki-pAAECcv...
I also built a simple MVP to explore the data interactively: Main site: https://glassballai.com
Browse all runs: https://glassballai.com/dashboard
Prompts and setup are all contained in the dataset. The setup is also documented here: https://glassballai.com/changelog
Feedback and critique welcome!