Why do CEOs get fired?

A machine learning study predicting forced CEO turnover using US firm and executive data (1992–2022)

This project investigates the determinants of forced CEO turnover using a dataset of US public firms from 1992 to 2022, combining executive compensation data (Execucomp) with firm financial data (WRDS).

To address data imbalance, we employed matched case-control sampling (1:1 and 1:2 ratios, repeated 3 times each). We compared three modeling approaches: Boosting + Logistic Regression, Lasso + Logistic Regression, and a baseline logistic model without variable selection.

Key findings:

  • Best model: Boosting + Logistic (AUC = 0.8072), outperforming Lasso + Logistic (0.7684) and the baseline (0.7669)
  • Most important executive features: gender, bonus, option grants, age, and tenure
  • Most important firm features: profitability metrics (income before extraordinary items, pre-tax income), followed by capital expenditure and receivables
  • Post-financial crisis shift: firms place greater weight on financial soundness (debt-to-assets, quick ratio) over short-term profitability
  • CEOs with low firm profitability but high compensation face higher firing risk

The model was applied to Samsung (0.6654, high-risk), Meritz Securities (0.1219, low-risk), and Hyundai (0.3501, medium-risk) as real-world demonstrations.

Team project for Data Mining Labs and Methods, Seoul National University (June 2025). Co-authored with Sungwon Ryu and Jihee Lee.