Cognitive Load and Information Processing in Financial Markets: Theory and Evidence from Disclosure Complexity

Abstract

Cognitive-load research in financial markets generally treats disclosure complexity as a property of the document and information acquisition as a direct interaction between an investor and that document. Machine-readable reporting and AI intermediaries make both assumptions incomplete. We develop a reader--task--interface framework in which processing load depends jointly on disclosure content, the representation through which it is accessed, and the transformation technology available to the reader. We evaluate the framework using a historical XBRL-transition diagnostic, a census of 207,684 SEC filings, a paired multi-model interface experiment, and held-out criterion validation. The historical evidence is consistent with interface substitution but has a modest statutory first stage and is interpreted as a measurement diagnostic. Modern filings reveal that human-facing burden and machine accessibility are distinct: large issuers produce longer reports while supplying richer structured coverage. The experiment crosses 432 filing-grounded questions with twelve interfaces and six hosted model implementations, producing 31,104 responses. Grounded accuracy rises from 47.2\% under BM25 HTML to 99.5\% under perfectly localized text; fact-matched oracle XBRL reaches 99.6\%. Most of the observed XBRL--HTML gap in these numerical tasks therefore arises from evidence localization and table reconstruction rather than syntax or taxonomy alone. Cognitive load is consequently not a stable scalar attribute of disclosure: it is local to a reader, task, interface, and processing stage.

31 pages. Substantially revised and expanded version. Reframes cognitive load for AI-mediated disclosure and adds a modern SEC filing census, a paired multi-model interface experiment, and counterfactual evidence tests