Relative valuation is among the most widely applied methods in valuation practice. However, traditional approaches based on industry averages or peer group medians exhibit significant conceptual weaknesses: the selection of comparable companies is often subjective, industry classifications remain coarse, and fundamental differences between firms are insufficiently captured. The econometric approaches developed in the literature, typically based on regression analyses, systematically incorporate fundamental factors and improve predictive accuracy, but reach methodological limits due to their underlying assumption of linearity.
A new generation of machine learning techniques – in particular tree-based approaches such as gradient boosting – has demonstrated substantially higher predictive accuracy in academic research. The study by Geertsema and Lu (2023) provides the first systematic evidence of this superiority on a broad empirical basis. In their published form, however, these models are only of limited applicability in valuation practice: they rely on a large number of input variables that are typically unavailable in real-world valuation settings, and their implementation requires considerable technical expertise.
This presentation addresses this gap and offers preliminary insights into how these promising machine learning models can be adapted such that they become genuinely applicable under the conditions of valuation practice while preserving predictive accuracy superior to that of conventional approaches.
This presentation will be held in English.
