Sealed Before the Open: A Verifiable Standard for Measuring AI Forecasts of Financial Markets
Year: 2026
Type: preprint
Abstract: Related: Verification Theater: Measured Failure Modes of AI Verifier Panels, accepted (poster) at the NeurIPS 2026 workshop Who Verifies the Agents? Version of record: SSRN 7653679, DOI 10.2139/ssrn.7653679. Cite that version. The same text is also at polyhelper.ai/research/sealed-before-the-open/ and https://github.com/vadimchernets/papers/tree/main/sealed-before-the-open. The registered predictions of Section 8 are on OSF: 10.17605/OSF.IO/9Q86N. The immutable release of the open league files on which every number of the paper rests: 10.5281/zenodo.23552700. The live league and its open data (CC BY 4.0): trends.polyhelper.ai/markets, trends.polyhelper.ai/api/v1. This deposit is an archival copy made so that the full text stays retrievable; it makes no separate claim. When an AI model says gold will rise, anyone can score the call at the close, but few can check when it was made. A backtest mixes memory with skill, and an operator's log asks for trust. I propose a standard for measuring AI market forecasts, written as a ladder of checkable requirements, and I run its reference implementation, AI Council · Markets: an open league where fourteen models, seven from US and seven from Chinese laboratories, face one question on one price snapshot. Before each New York open the ten on its Council, five from each side, give probabilities for twelve indicators at four horizons. Each answer is sealed into a hash chain before the price it is scored from, the chain is anchored in Bitcoin and signed by independent time-stamping authorities, and one run per session counts by a rule declared in advance. On 21 October 2026, the first counted morning, every counted answer was signed before the 09:30 ET open. Agreement among the models proved to be a count of independent voices, not a headcount: on two development mornings the seven and eight models sealed before the open were as alike as about 1.6 independent forecasters, nine counted ones were worth 2.23, and the closest pairs crossed the US-China line. Check a seal with three commands, and score my thirty predictions, registered on OSF, before I do.
Author: Chernets, Vadym
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7223459106