A major challenge with using large language models (LLMs) like ChatGPT for historical analysis is look-ahead bias: an LLM trained in 2024 knows what happened in 2020-2022, making it impossible to trust what it tells you about what it would have done in 2020 without contamination from future knowledge.

causalpending

Speaker

Jacob Pizor

Evidence Quote

when you're looking at uh trying uh to assess what the your stimulation would have done in 2020 and you're using an llm that's fitted in 2024 that llm knows um what happened in in 2020 2021 2022 so um how do you really trust what it is that it's telling you

Source

Redefining Value Investing in a Magnificent Seven Dominated World | Jacob PozharnyExcess Returns
Created: 8/11/2026, 7:46:05 AM

My Notes

Loading notes...