LLMs are literally prediction machines (that’s how transformers work), and so you’d expect them to be especially good at predicting things, e.g., for financial forecasting. I’ve found this to be generally true.
That said, I’ve found that the latest models (circa summer 2026, e.g., Opus 5, GPT 5.6) struggle with low-probability / tail events, especially positive tail events.
This feels like an important capability for business: to predict which asymmetric bets will be home runs, to predict which investments will yield outsized returns, etc.
These things are fundamentally harder to predict than normal curves, but I can’t help but notice that the models seem to be lacking… optimism on some level. Going to file this away and keep an eye on it.