AI_CDG_Earnings_Management_Replication_Package
Description
This repository contains the replication package for the manuscript: Substantive AI Disclosure, Corporate Digital Governance, and Earnings Management: Evidence from U.S. Listed Firms The package provides the final firm-year regression dataset, intermediate construction files, validation files, and output tables used to support the empirical analysis. The study examines U.S. publicly listed firms over 2015-2023 and links SEC Form 10-K and DEF 14A textual evidence with financial data derived from SEC XBRL company facts. The package includes: - Final regression dataset - SEC filing match panel - AI contextual disclosure measures - Corporate digital governance coding - Financial variables - Modified Jones accrual-based earnings management measures - Roychowdhury real earnings management measures - Manual validation files for AI and CDG classification - Regression and correlation output tables - README, codebook, data availability statement, and software requirements Raw SEC filings are not redistributed due to file-size constraints. Filing accession numbers, filing dates, filing URLs, and source identifiers are provided to allow retrieval from SEC EDGAR.
Files
Steps to reproduce
Raw data: Raw SEC 10-K and DEF 14A filings are not redistributed due to file-size limits. Retrieve them from SEC EDGAR using the accession numbers, filing dates, and filing URLs listed in data/intermediate/filing_match_panel_full_history.csv (see documentation/CODEBOOK.txt). To regenerate this package from raw filings, run the scripts in scripts/ in this order: 1. make_firm_master.py - builds the firm universe from SEC company tickers. 2. make_10k_index_full_history.py - indexes historical 10-K filings per firm from EDGAR. 3. make_filing_match_panel_full_history.py - matches 10-K and DEF 14A filings into a firm-year panel (reproduces filing_match_panel_full_history.csv). 4. make_firm_profile.py - builds firm-level profile information. 5. Download the raw 10-K/DEF 14A text referenced in the filing match panel into local raw_filings_full/10k and raw_filings_full/def14a folders (set the RAW_DATA_DIR environment variable accordingly). 6. create_contextual_ai_measures.py - classifies AI disclosure as substantive/symbolic/mixed (reproduces ai_contextual_measures_full_sample_public.csv). 7. extract_cdg_evidence_full_sample.py then make_cdg_coding_rule_based_full_sample_v2.py - extract and code corporate digital governance evidence (reproduces cdg_coding_rule_based_full_sample_v2.csv). 8. extract_financial_variables_full_sample_v3.py - pulls financial statement variables from SEC XBRL company facts (reproduces financial_variables_full_sample_v3.csv). 9. Merge the intermediate files above on firm-year, winsorize continuous variables at 1/99 percent, then run estimate_aem_modified_jones_full_sample_v2.py and estimate_rem_roychowdhury_full_sample.py to obtain accrual-based and real earnings-management measures. Combining all variables yields data/final/final_regression_dataset.csv (21,509 firm-year observations, 3,154 firms, 2015-2023). To reproduce the manuscript tables directly from the final dataset: 10. run_baseline_regressions.py - baseline and robustness regressions. 11. make_descriptive_statistics_final.py - descriptive statistics table. Run scripts/verify_package.py at any point to confirm file presence and that key sample sizes and validation accuracy rates match the manuscript.