Can We Stop the Backtest Yet? Anytime-Valid Promotion Tests for Adaptive Financial Research Agents

Abstract

Adaptive research systems can generate, discard, revise, and monitor trading rules faster than their selection history can be audited. We ask when such a system may promote a frozen rule to governed forward evaluation. The proposed protocol assigns each candidate a predictable, nonrefundable share of one lifetime testing budget and accumulates a bounded, cost-adjusted e-process only from post-registration returns. Under the null that a candidate’s clipped net score has nonpositive conditional mean at every scored date, a conditional Ville argument controls the probability of any false promotion despite dependent rules, adaptive births, and indefinite monitoring. In 3,000 shared-sign null replays of a frozen 24-rule public-market library, repeated unadjusted tests cross at least once on 94.5% of paths, while the funded rule records 0/3,000 crossings. A separate exact mean-zero experiment with adaptive identities and births records 1.77% familywise crossings over 3,000 programs. The 2015–2026 market replay promotes no rule after costs. The result is an auditable promotion primitive for agentic and open-finance infrastructure: executable identity, evidence, budget, and delayed execution are recorded in one append-only ledger.