@Wyliesau i read from a guy who actually did something like this. he trained some (pre-llm days) models on posts from stock pundits, and distilled them in to ticker-indicator tuple. then he backtested each of the pundits to a stock to make a profile of which pundits tended to be accurate about particular ticker-indicator tuples, and then that was the basis of his investment strategy.

literally just learn which people are statistically correct and then gamble on them