Machine Learning Cuts the Error in Estimating Risk for Private Companies by 42 Percent

Valuing a company that has no traded stock is a routine but awkward job. The valuation needs a discount rate, and the discount rate needs a measure of market risk called beta. Beta captures how much a stock tends to rise or fall when the overall market rises or falls. A private firm has no stock price, so it has no beta of its own.

The usual workaround is comparable company analysis. Analysts find a set of public peers, take their betas, adjust for debt, and apply the result to the private firm. Earlier research has questioned how accurate that method is. A paper in Applied Economics by four faculty members of the McCoy College of Business at Texas State University tests whether machine learning can do better.

The authors are Emmanuel Alanis, Associate Professor of Finance; Vance Lesseig, Associate Professor of Finance; Janet Payne, Professor of Finance; and Margot Quijano, Associate Professor of Finance. All four are in the Department of Finance and Economics.

The study

The researchers set up a horse race. On one side is the standard comparable company analysis. On the other are machine learning algorithms, statistical models that learn patterns from large amounts of data without being told the rules in advance.

The test is out-of-sample, meaning each forecast is made using only information available before the period being predicted. That guards against the models simply memorizing the answers. The tests run from 1990 to 2021, more than three decades that include several booms and busts.

Accuracy is measured by mean absolute error, the average gap between the predicted beta and the actual beta, ignoring whether the miss was high or low.

What the researchers found

Machine learning won decisively. The models reduced mean absolute error by more than 42 percent relative to comparable company analysis.

The improvement was not evenly spread. It was most pronounced for smaller and younger firms, and for firms whose mix of debt and equity differed from their peer group. Those are exactly the cases where a peer-based estimate is on shakiest ground, because the peers are not really comparable. The machine learning approach appears to handle those differences rather than averaging over them.

What it means for practitioners

The abstract points to private firm valuation as the natural application. Private equity investors, corporate development teams, business appraisers, and courts handling valuation disputes all depend on beta estimates for firms without traded shares. A 42 percent cut in error is large enough to change deal prices and litigation outcomes.

The results also suggest where to focus. For a large, mature private company with a conventional balance sheet, the peer method may be good enough. For a young company, a small one, or one with unusual leverage, the gains from a machine learning estimate are likely to be worth the added effort.

A caution is in order. The study measures how well the models predict beta, not how a specific valuation would have changed. Practitioners will want to see the full paper for the algorithms used, the data inputs, and how the models were trained before adopting the approach.

This summary is based on the paper’s abstract. The full article reports the data, methods, and detailed results.

What it means for managers

• Do not settle for peer-group averages. Machine learning forecasts of beta reduced the mean absolute error by more than 42 percent compared with comparable company analysis.
• The gain is largest where it matters most. Smaller, younger firms whose capital structure differs from their peers are the ones where peer-based estimates go most wrong.
• Private company valuation is the obvious application. A better beta means a better discount rate, and a better discount rate means a more defensible value.

Alanis, E., Lesseig, V., Payne, J. D., & Quijano, M. (2025). Can machine learning methods predict beta? Applied Economics, 57(21), 2742-2756. 10.1080/00036846.2024.2331039