How Machine Learning Could Improve the Selection of Board Directors

How Machine Learning Could Improve the Selection of Board Directors

Think about a boardroom, with all the people sitting around a table. If there is an opening, how do you pick the next director?

It’s a problem that goes back to the days of Adam Smith, author of The Wealth of Nations, which was published in 1776. He wrote that management could control boards and take actions that are not in shareholders’ interests. During the Great Depression, Adolf Berle and Gardiner Means were concerned that control tended to fall in the hands of those who selected a proxy committee appointed by existing management, so they virtually dictated their own successors. Today, if you pick up an issue of Fortune or Business Week, there is often an article about how boards are not doing their job well.

Boards tend to pick people they know, since they’re going to have to work with that person. They don’t usually want a stranger, and as a result they choose people with the same backgrounds. Typically, that’s middle-aged white men who may not always be the best person for the job. Institutional investors are not happy about this situation. They’re constantly trying to interfere and encourage managers to pick more diligent boards.

Consider an algorithm that weighs possible choices of directors to identify better options along with the expected performance of those options. An algorithm removes people’s implicit biases. Its choices shouldn’t be forced on boards, but it could serve as a tool for the recruitment process.

The questions are: Could this kind of process potentially work better than what’s been in place for 250 years? How do you compare the two approaches? Machine learning holds the solution. My colleagues Lea Stern, Isil Erel, Chenhao Tan, and I wrote a paper on this very topic.

The Power of Machine Learning

Machine learning algorithms are designed to maximize predictive accuracy out of samples. To do that, they sift through many variables and combine them in nonlinear ways to identify complex patterns and structure in the data.

Machine learning is pretty powerful. I can post a picture of some random friend and me on Facebook. It can be a blurry picture that I took with an old camera. Facebook can still tell from that picture who my friend is out of the hundreds of millions of people in its data sets. The algorithm knows because it can use the data that it has about the shape of the person’s nose, the individual’s hair, face, and everything else, then predict with astonishing accuracy exactly which person that picture represents.

Hiring can be thought of as a prediction problem: you try to predict how well a person will do in a new job. Prediction problems are what machine learning does extremely well. That’s why we can ask the computer to take all of the information that we know about potential directors as well as the performance of historical directors, then come up with a prediction of how well a potential director will do.

With an algorithm, we can look at the difference between the performance of algorithm-selected directors and management’s selected directors. Then we can potentially influence the quality of the hiring decision. The second thing an algorithm can do is look at what features of directors people tend to overrate.

This discussion is predicated on how we define director performance. This is difficult because most director actions are “collective.” The board acts as a group to hire a CEO, fire a CEO, accept a merger bid, etc. To construct an algorithm to select directors, we need a measure of not just how effective the board is but how effective an individual director would be. To think about this, we go back to the board’s mandate, which is to represent shareholders’ interests.

Shareholder Value Versus Company Value

There is a famous paper by Nobel Prize winner Oliver Hart and Luigi Zingales, a well-known economist. They argue that maximizing shareholder welfare is not necessarily the same as maximizing the value of the firm. For example, shareholders may care about not polluting the environment. The traditional view is that shareholders should care only about maximizing profits, which goes back to Milton Friedman. He would argue that companies should care only about maximizing profits, where Hart and Zingales argued they should maximize what shareholders actually care about. Therefore, if we want to think about what directors do, we should focus on how they will represent shareholders’ interests.

How do you measure that? Annually or every three years, there is a vote for directors. These are a bit like Soviet-style elections in that the average director gets 95% of the votes. But what most people fail to realize is that there is substantial variation around that. Some directors get 80%, while others get close to 100%. Previous research shows that if directors do less than average in these elections, they tend to make worse decisions. If they do better than average, they tend to make better decisions. So even though somebody who gets 80% of the vote does get reelected, this is a signal from shareholders that this person is less likely to represent their interests.

There were 25,000 new directors on U.S. public company boards between 2000 and 2014 with voting outcomes up to 2016. We trained our algorithms on directors who were appointed between 2000 and 2011, then used this data to predict the performance of directors who were appointed between 2012 and 2014. That allowed us to look at the fraction of bad outcomes by various directors and board characteristics.

Which directors do shareholders like? We found that male directors have 10% of the bad outcomes, while female directors have only 7.9%. Another data point that stands out is if directors are busy — on more than three or four boards — shareholders are more likely to dislike them.

Making Better Decisions

The next question is: Can an algorithm really improve on board decisions? There are two factors in that: selective labels, so we can observe only outcomes for hired directors, and unobservables, features the algorithm does not see, such as whether people are actually good at the business they are in. How do we deal with this?

Looking at the data with the aid of machine learning allowed us to find features of predictably unpopular directors — potential directors who were added but shouldn’t have been. It turns out that men tend to do worse as directors than women if they have a larger network size, a finance background, and several previous and current board seats. Diverse boards are good.

One issue that some people had with our experiment is should we really care about a director’s popularity, or should we care about profitability? It might not matter — either way, profits increased. The important thing was a potential director’s predicted performance: the market reaction to the appointment of directors in the bottom rungs of predicted performance was negative 1.9%, while choosing a director that the algorithm liked resulted in the stock going up almost 0.75%.

The results confirm the observation that dates to Adam Smith, that the board selection process leads to directors who are nearest at hand and not necessarily the ones who are the best choices to serve shareholders’ interests.

Institutional investors should ask their firms to look at the data from an algorithm such as this. Maybe they don’t feel like they can tell the firms to pick this person or that person, but they can ask them to pick somebody that the algorithm thinks is in the top 10% of potential directors. Certainly, they can ask not to appoint those in the bottom 10% to 20%.


About Michael Weisbach

Dr. Michael Weisbach is the Ralph Kurtz Chair of Finance at The Ohio State University, where he has focused his research on issues concerning corporate financial policy. Dr. Weisbach’s work applying quantitative methods to capital structure choices won the prestigious Jensen Prize, and his research on the motivations for stock repurchase programs led him to advise the Securities and Exchange Commission concerning these programs’ regulations. Dr. Weisbach is an editor of the Review of Financial Studies. He is also an associate editor with Financial Management, Journal of Financial Economics, Journal of Corporate Finance, and Journal of Multinational Financial Management.


This article is adapted from the August 6, 2020, GLG webcast “Using Machine Learning in Board of Director Selection.” If you would like access to this webcast or would like to speak with Dr. Michael Weisbach, or any of our more than 700,000 experts, contact us.