Please rotate your device to landscape mode to view the charts.

Selecting Directors Using Machine Learning

Journal: Review of Financial Studies

Date: 20210701

Author: Erel, Isil; Stern, Lea H.; Tan, Chenhao; Weisbach, Michael S.

Abstract:
Can algorithms assist firms in their decisions on nominating corporate directors? Directors predicted by algorithms to perform poorly indeed do perform poorly compared to a realistic pool of candidates in out-of-sample tests. Predictably bad directors are more likely to be male, accumulate more directorships, and have larger networks than the directors the algorithm would recommend in their place. Companies with weaker governance structures are more likely to nominate them. Our results suggest that machine learning holds promise for understanding the process by which governance structures are chosen and has potential to help real-world firms improve their governance.

Link: Google Scholar

Background and Context

Research Problem

Corporate boards need to predict which directors will perform well, but traditional methods of selecting directors may be influenced by agency conflicts and behavioral biases.

Study Approach

The researchers developed machine learning algorithms to predict director performance using data from 41,015 independent directors appointed to 4,887 U.S. corporate boards between 2000-2014.

Methodology

The study uses machine learning models (XGBoost, Ridge, Lasso, Neural Networks) to predict director performance measured through shareholder votes, dissent levels, and director turnover.

Machine Learning Models Outperform Traditional Methods in Predicting Director Performance

  • Machine learning algorithms (XGBoost) more accurately predict director performance compared to traditional statistical methods (OLS)
  • Directors predicted to perform poorly by XGBoost received -3.1% excess votes, while those predicted to perform well received +1.2% excess votes
  • Traditional OLS models showed no meaningful relationship between predicted and actual performance

Higher Dissent Against Algorithm-Predicted Poor Performers

  • Only 1.3% of directors in the bottom decile of predicted performance faced strong dissent
  • 23% of directors in the top decile of predicted poor performance faced strong shareholder dissent
  • Shows clear relationship between algorithm predictions and actual shareholder disapproval

Predictably Bad Directors vs Alternative Candidates

  • Predictably bad directors are more likely to be male (88% vs 83%)
  • They have larger professional networks (1714 vs 1261 connections)
  • They sit on more boards (3.1 vs 1.6) despite worse performance

Market Reaction to Director Appointments

  • Market reacts negatively (-1.94% returns) to appointments of directors predicted to perform poorly
  • Positive reaction (+0.75% returns) to appointments of directors predicted to perform well
  • Suggests market participants recognize director quality consistent with algorithm predictions

Early Director Turnover Predictions

  • Only 2% of directors in bottom decile of predicted turnover left within 2 years
  • 43% of directors in top decile of predicted turnover left within 2 years
  • Demonstrates algorithm's ability to predict poor director-firm matches

Contribution and Implications

  • Machine learning algorithms can significantly improve director selection processes by providing objective predictions of director performance
  • Current selection practices often favor traditionally networked male directors despite evidence they may not perform as well
  • Implementation of algorithmic decision aids could help boards make more effective and diverse director selections

Data Sources

  • Performance comparison chart based on Table 3 showing predicted vs actual excess votes
  • Dissent prediction chart based on Figure 3 showing relationship between predicted and actual dissent
  • Director characteristics comparison based on Table 6 comparing predictably bad directors to alternatives
  • Market reaction chart based on Table 5 showing CARs around director appointments
  • Turnover prediction chart based on Figure 4 showing relationship between predicted and actual turnover