Detecting earnings management beyond accrual models
Earnings management is a mainstay of the finance and accounting journals, still measured mostly with accrual regressions such as the modified Jones model. Only 4 of 156 recent papers on it carry a machine-learning or text keyword. Do learned models, including language models reading the filings, flag manipulation that accrual models miss?
What prompted this#
Since October 2025, 156 papers carry an earnings-management or accrual keyword (88 in finance journals, 47 in accounting). Only 4 of them also carry a machine-learning, deep-learning, language-model, text or neural-network keyword. Accrual models such as the modified Jones model, on the equation ranking's list, remain the standard measure.
Questions#
- Against known cases (restatements, enforcement actions), do learned models built on raw financial statements and filing text beat accrual residuals?
- Do the measures disagree on which firms manage earnings, and do published results depend on the measure used?
- Can a language model explain why it flags a firm, in terms an auditor could check?
Data that could help#
Restatements and enforcement actions (Audit Analytics, SEC releases), financial statements (Compustat or CSMAR) and filing text (SEC EDGAR).
Next step#
Train a gradient-boosting model on raw accounting items to predict restatements, and compare its ranking of firms with modified-Jones residuals.