A systematic review of how studies combine financial signals, design AI agents and report the evaluation of their results.
Before building an AI system that fuses financial evidence, we wanted to know how the field already tests such systems.
The review asks three questions: how studies combine fundamental, technical and textual signals; how they design and coordinate AI agents; and how they report the evaluation of their predictions.
A pre-specified search of journal articles indexed in Scopus and Web of Science, 2015 to 2026.
A language-model pipeline screened and coded titles and abstracts, checked against human coding of 198 abstracts.
Fifty selected studies were read in full to test and explain what the abstract-level coding shows.
journal studies reviewed, 727 of them empirical.
empirical studies state no evaluation protocol in their abstract.
combine walk-forward testing, an economic measure and transaction costs.
walk-forward reporting rose sharply, in every research question.
Fusing all three kinds of evidence, fundamental, technical and textual, is rare.
Fusion and agentic studies report walk-forward designs less often than studies focused on evaluation methods.
The recent rise in walk-forward reporting suggests stronger standards are starting to take hold.
Figures describe what studies report in their titles and abstracts. Rounded from the working paper, which is not yet peer reviewed; final figures may differ.
The review sets the bar our own systems must clear.
It shapes the design of our multi-agent forecasting work (P2) and is the evidence base for the open evaluation standards we plan to publish (P4).
Kalantarinejad, R. and Amel-Zadeh, A. Working paper.