This notebook demonstrates how Statistics can be used to extract quantitative answers from the qualitative outputs of large language models (LLMs).
I used alpaca api to get historiocal data and bencmarked ARIMA and LLMP for 6 months daily data whole preicting the next 5 days.
Inspired by the NeurIPS 2024 paper LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language, this project explores how we can formalize and analyze the implicit uncertainty and beliefs encoded in LLM predictions.
By conditioning LLMs on contextual text and using Bayesian inference, we can model a predictive distribution over numeric targets — effectively bridging the gap between free-form natural language and structured statistical modeling.
This work is inspired by:
- LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language
James Requeima, John Bronskill, Dami Choi, Richard E. Turner, David Duvenaud
This is simple prelimary work that I spent a single evening on, so it needs some polishing (like accounting for stock splits)