Comparable Company Valuation through Machine Learning
Paul Geertsemaand - The University of Auckland;
Helen Lu– The University of Auckland
-- Who are the peers in a comparable company valuation? Can machine do a good job at comparable company valuation? More importantly, can machine help us understand the valuation drivers? Our recent paper titled, “Relative Valuation with Machine Learning”, provides some answers.
Comparable company valuation is widely used by practitioners. This approach compares a company’s valuation multiple (or, the market value of assets or equity scaled by an accounting measure) to those of its industry peers to decide whether it is overvalued or undervalued. More sophisticated practitioners consider firm fundamentals such as growth in addition to industry. The choice of comparable companies is often highly subjective because it is hard to find peers that are very similar along all important dimensions (for example, firms with similar growth rate, risk profile and accounting returns). Academics have proposed more rigorous ways to estimate valuation multiples, some of which use regressions that impose linear relations between valuation multiples and firm characteristics such as profitability, growth and leverage ratios. Recent advances in machine learning allow us to take a completely new approach to conducting comparable company valuations. Instead of restricting peers to be from the same industry, we provide decision-tree based machine learning models with 97 variables (including accounting ratios and industry classifications) and let the machine figure out the optimal decision rules for selecting comparable companies.
Decision tree-based machine learning models are an intuitively appealing tool for selecting comparable companies to conduct relative valuation. An analogy to tree-based models is the “I spy with my little eye” game played by children, where sequential questions are asked to narrow down the scope and identify the object the first player chosen within sight. Tree-based models allocate firms to different leaves on a decision-tree. Comparable companies are simply firms allocated to the same leaf. Tree-based models also address the issue of functional form uncertainty and multi-way-interactions among input variables. This is achieved automatically through choosing optimal sequential decision rules that minimise prediction errors. Equally important, the decision rules used in these models can reveal important drivers for cross-sectional valuation. Figure 1 shows a decision-tree from our model.
We use this approach to value US common stocks listed on NYSE, NASDAQ and AMEX between 1980 and 2019. Depending on the multiple used, machine learning models reduce median absolute valuation errors by 5.6 to 31.4 percentage points relative to traditional models. Valuation is more challenging in certain markets and for certain assets.
The median absolute valuation errors from machine learning models were extremely high during the internet bubble in 2000 and during the 2008 financial crisis (see Figure 2). During both periods, a large number of firms were severely miss-valued according to the machine learning models. That is, machine learning models were struggling to relate firm fundamentals to their valuation multiples during these periods. In the cross-section, small firms, firms with extreme market-to-book ratios and firms with low ROEs are harder to value than other firms. Across industries, financial and utility firms are the easiest to value and firms in the healthcare and telecom industries are the hardest to value. Throughout our sample period, machine learning models targeting log market-to-book multiples produce the most accurate valuations among the three multiples considered, but the performance gap has been shrinking across time, perhaps because the book-value of equity has become a less important measure of a firm's return producing capital in recent times.
Remarkably, the input variables machine learning models rely on most are similar to those that would have been used in discounted cash flow (DCF) analysis (by a professional familiar with valuation theory). For example, consistent with the predictions of DCF, machine learning models choose ROE-type variables to predict equity-value multiples and ROA-type of variables to predict enterprise-value multiples. Contrary to common practice in choosing firms from the same industry as comparable firms, we find that industry classification only makes a moderate contribution to the valuation process. The most important types of variables in relative valuation are profitability, growth, efficiency and financial soundness. Since 1990, growth-related variables have become more important in valuation while variables related to profitability have diminished somewhat.
High valuation accuracy in itself does not guarantee that valuations from machine learning models are good proxies for fundamental values. Investors care about fundamental values because deviations from these values suggest relative miss-pricing. We conduct two further empirical tests and show that valuations from our machine learning models behave like fundamental values. First, valuation errors predict returns. Deviations from model valuations are corrected in the next month – model-overvalued stocks on average drop in price and model-undervalued stocks increase in price over the next month. Second, our models can identify severely over-priced IPOs. These over-priced IPOs underperform the market in the long run, generating large abnormal losses of 60 to 80 bps per month based on the Fama-French six-factor model.