LLM-Based Bayesian Forecasting Update Method

sirbayes · x · 2026-07-14

The author explains a Bayesian forecasting approach: using an LLM to estimate the log-likelihood ratio \(\lambdat\) at each step, then applying recursive/online updates to the log-odds. A tempering parameter \(\alpha\) is introduced to mitigate overconfidence.

They note that on forecast bench, this method significantly underperforms their heuristic BLF ("direct") approach; however, for questions lacking a crowd prior, it can sometimes make the agent "less wrong."

Related event: Bayesian Prediction Updating Method Using LLMs(3 posts)→

Original post →

More from Research

Research channel →