Most attempts to apply AI to financial markets start from the same assumption: that a sufficiently powerful model, given enough data, can predict where prices will go. This assumption is wrong — not because the models are not powerful enough, but because markets have no ground truth. Prices are not a measurement of reality. They are the aggregate of every participant's attempt to predict every other participant. An AI trained on that data will converge to consensus — which is already priced in.

The evidence is visible in the data. Give a large language model complete fundamental financial data for 500 stocks and ask it to identify which are undervalued. The result underperforms a basic quantitative factor — not because the model lacks intelligence, but because the task itself is structurally unsolvable using the available information. The best systematic investors in the world, with access to the same public data, regularly disagree on individual stock direction. An AI trained on that data cannot do better.

Axiome uses AI for a different set of problems — ones that do have stable, verifiable answers. Does the fundamental context of this stock significantly contradict what the quantitative signal is saying? Does this novel geopolitical event map to specific risk factors in the current portfolio? What systematic patterns exist in this company's regulatory filing language relative to prior periods? These are questions with ground truth. They are also questions that quantitative models, which operate on historical correlations alone, cannot answer.

The result is an investment platform where AI and systematic methods each do what they are genuinely suited for — and neither is asked to do what it cannot. The research below demonstrates the performance difference this distinction produces across fourteen years of data.

The Market Study

Why LLMs fail as market predictors — and where they add value.

Fourteen years. Identical input data. The only variable is methodology.

The Research Note

The limits of content-based prediction.

The same structure, established adversarially outside markets. A pre-registered study finds that the content of a proposed code change cannot predict whether a team will accept it — only the surrounding context can. A consensus-driven process resists prediction whether it is a market or a code review.