For ~38 days, a cronjob generated daily forecasts:<\/p>\n
\u2022\u2060 \u206030 stocks across different sectors<\/p>\n
\u2022\u2060 \u206010-day horizons<\/p>\n
\u2022\u2060 \u2060~30 predictions\/day (different stocks across multiple sectors)<\/p>\n
\u2022\u2060 \u2060Fixed prompt and parameters<\/p>\n
Each run logs:<\/p>\n
\u2022\u2060 \u2060Predicted price<\/p>\n
\u2022\u2060 \u2060Natural-language rationale<\/p>\n
\u2022\u2060 \u2060Sentiment<\/p>\n
\u2022\u2060 \u2060Self-reported confidence<\/p>\n
Because the runs were captured live, this dataset is time-locked and can\u2019t be recreated retroactively.<\/p>\n
### Platform<\/p>\n
I built a simple MVP to explore the data interactively:<\/p>\n
https:\/\/glassballai.com<\/a><\/p>\n https:\/\/glassballai.com\/results<\/a><\/p>\n You can browse and crawl all recorded runs here <\/p>\n https:\/\/glassballai.com\/dashboard<\/a><\/p>\n .<\/p>\n ### Goal<\/p>\n This is not a trading system or financial advice.<\/p>\n The goal is to study how LLMs behave over time under uncertainty:<\/p>\n forecast stability, narrative drift, confidence calibration, and prompt-conditioned bias.<\/p>\n ### Dataset<\/p>\n After ~1.5 months, I\u2019m publishing the full dataset on Hugging Face.<\/p>\n It includes forecasts, rationales, sentiment, and confidence.<\/p>\n (Actual prices are rehydratable due to licensing.)<\/p>\n https:\/\/huggingface.co\/datasets\/louidev\/glassballai<\/a><\/p>\n ###Stats:<\/p>\n Stocks with most trend matches: ADBE (29\/38), ISRG (28\/39), LULU (28\/39)<\/p>\n Stocks with most trend misses: AMGN (31\/38), TXN (28\/38), PEP (28\/39)<\/p>\n Feedback and critique welcome.<\/p>\n<\/div>\n submitted by \/u\/aufgeblobt <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
