louidev releases GlassBallAI, a time-locked dataset of Gemini forecasts for studying LLM behavior

louidev has released the GlassBallAI dataset on Hugging Face, capturing live forecasting behavior from Google's Gemini models before outcomes are known. The dataset contains over 3,655 rows of data collected between February 17 and May 19, 2026, covering Gemini 2.5 Flash, 2.5 Pro, 2.5 Flash Lite, and 3.0 Flash Preview.

• Each entry includes a 10-trading-day lookahead for price direction, sentiment, confidence, and full reasoning traces.
• The data is time-locked to prevent retrofitting answers, with global Expected Calibration Error (ECE) measured at 0.217.
• Early findings indicate a confidence/accuracy inversion where accuracy drops to approximately 28% in the highest-confidence bin (0.8–0.9).
• A companion dashboard and methodology writeup are.

The dataset is intended to facilitate research into LLM calibration and hallucination by providing grounded, real-time forecasting data rather than resolved historical records.