GeoThesis

Language models like ChatGPT or Claude don't remember anything between conversations. Without good data they also make things up. I wanted to know what changes if you stop asking the model to remember, and build it a proper data pipeline instead.

I picked stocks because the question is easy to test and hard to get right. Geopolitical news moves markets, but even when you read it, it isn't obvious which stocks a story will touch.

Every day, the system checked what was going on in the world and saved it in a database. Then it wrote me a daily briefing: here is what's happening, and here are the stocks that are likely to be affected, and why.

The part I cared about was what happens over days. If the same news keeps coming back, the saved history lets the system notice that, and set it against how the prices moved.

I stopped running it after a while. I didn't have time to improve the data or the predictions, and I never compared what the briefings said with what the stocks did. So I can't tell you whether it worked.

I'd rather say that plainly than guess. It was a small experiment, and it ended before it had an answer.