RV-FLS: A Resource for Retrospective Verification of Forward-Looking Statements
Abstract
Forward-looking statements (FLSs) influence decisions before their outcomes are known, yet they are rarely revisited systematically after relevant evidence becomes available. Retrospective verification of FLSs differs from conventional fact verification because supporting or contradicting evidence may emerge gradually, remain inaccessible, or be insufficient for a responsible judgment. We introduce RV-FLS, a resource for retrospective verification of scenario-level FLSs derived from financial analyst reports. RV-FLS contains 651 human-reviewed instances. Each instance links an original statement and publication date to post-publication evidence collected within a five-year verification window, an evidence-grounded rationale, an agent-generated judgment, a human review decision, and a final label. We construct the resource through an agent-first workflow in which a retrieval-augmented language model assembles evidence packages and proposes preliminary judgments for human adjudication. As standalone label predictors, a closed-book model and a retrieval-augmented agent achieve accuracies of 26.5% and 42.5%, respectively. Under agent-first construction, however, reviewers accept 74.5% of the generated judgments, while average verification time decreases from 12.89 to 3.34 minutes per scenario. RV-FLS supports research on evidence retrieval, evidence-conditioned classification, and end-to-end retrospective verification, while providing an auditable foundation for making future-oriented claims more transparent and accountable.