Outlier or Signal? An Auditable Framework for SEC Filing-Derived Financial Data Cleaning

Abstract

Researchers often clean financial-statement panels before the research question is fully visible. Extreme SEC XBRL values can be construction errors, one-time accounting events, operating signals, or unresolved cases, but common preprocessing rules often collapse them into a single outlier category. This paper develops an auditable alternative. The pipeline downloads official SEC submissions and companyfacts data, constructs a provenance-preserving firm-year panel, identifies anomaly candidates without deleting them, and links each candidate to filing evidence used for review. In a 20-company retail and consumer discretionary sample covering 156 unbalanced firm-years within a 2016-2024 target window, the pipeline identifies 73 anomaly candidates across 50 firm-years. Human review classifies 38 candidates as business signals, 24 as ambiguous, 10 as accounting events, and 1 as a construction issue. The central case, Costco 2019, shows why the audit trail matters: a gross-profit reconciliation failure is not an operating anomaly but a line-item definition mismatch. In a within-sample inference demonstration, mechanical winsorization shifts the operating-margin persistence estimate and inflates its standard error, while audited cleaning preserves the evidence-supported estimate. The contribution is methodological: data lineage is tied to outlier-treatment decisions, so a researcher can see why a value was retained, corrected, set aside, or flagged.