How AI is changing market analysis

The temptation when LLMs first arrived was to ask them to predict prices. The good news is, they cannot. The better news is, the questions LLMs are actually good at are the questions desks have been spending the most expensive analyst-hours on.

What they win at today: narrative compression. A research desk used to need three analysts to read 200 wires and synthesise them into a coherent storyline. Haiku 4.5 does it in 90 seconds, every 15 minutes, for the same cost as a couple of coffees per day. The quality is not equal to a senior analyst — it is competitive with a junior, available 24/7, and never tired. That is the regime change.

The second win is filings parsing. 10-Ks are tens of thousands of words of intentional density. An LLM can extract the changed clauses (this year vs last year), surface the new risk factors, and flag the management language that diverges from prior calls. A junior analyst could do this in three hours; the model does it in 30 seconds.

The third win is question answering against a known corpus. Give a model the last 24 months of a company's earnings calls and let a portfolio manager ask "when did management first mention pricing pressure?" The model will find the exact quote with the date. This is grounded retrieval, not prediction, and it is where AI is most reliable today.

The failure modes are not subtle. Models will confidently quote numbers that are off by an order of magnitude. They will conflate two companies with similar names. They will make up sources that look real. Every workflow that depends on them needs a verification step, ideally automated. At RockstarMarkets we constrain Haiku to only cite facts present in the input batch — anything outside that window is forbidden by the system prompt and dropped if it appears.

The trader's job in 2026 is no longer to gather information; it is to validate the AI's synthesis and decide what to do with it. That is a more leveraged role, and arguably a better one.