The Impact of Inter-firm Heterogeneity on Regression-based Valuation Models: A Comparative Study of GICS Segmentations and the SARD Approach in US Equity Markets

Abstract

Within peer groups chosen for relative valuation, inter-firm heterogeneity is typically addressed by employing measures of central tendency, such as the average multiple. Limited research has been devoted to handling such heterogeneity through regression-based prediction models for valuation multiples – and the impact of heterogeneity on the prediction accuracy of linear models when compared to peer group averages. We explore how model characteristics change when varying peer groups from market to GICS sector groups, and GICS industry groups, and when applying the Sum of Absolute Rank Difference (SARD) approach for peer selection. Furthermore, we seek to narrow the research gap of the SARD approach within the context of linear regression models, and how data subsets sorted by the SARD approach handle the MLR.1-6 assumptions. In doing this, we will explore the degree to which variance in our selected underlying value drivers, growth, profitability, and risk, can significantly explain variances in EV/EBITDA multiples, across different peer segmentations.
Our dataset was comprised of observed EV/EBITDA, EBITDA CAGR, ROIC and WACC figures for a trimmed census of 929 companies from the S&P 1500 Composite Index, which throughout most of the segmentations adhered to the MLR.1-6 assumptions – but with indications of an omitted variable bias. Through running Simple Linear Regressions (SLR), we found there to be varying significant beta coefficients across the different segmentations; mostly positive for growth, close to zero for profitability, and surprisingly – positive for risk. We found the underlying value drivers to have joint significance in around half of the segmentations, with slightly better goodness-of-fit statistics for GICS segmentations than SARD groupings. However, we found that heteroskedasticity in error terms successively decreased, when moving from an aggregate market level, to sector, to industry, to SARD groupings. Relative prediction accuracy followed the same pattern, with SARD groupings having slightly smaller relative prediction errors. SARD groupings also had slightly fewer cases of non-normality and heteroskedasticity, we deemed that the SARD groupings were overall less susceptible to systemic bias and overfitting. Our models cannot fully explain the relationship between selected underlying value drivers and the EBITDA-multiple and are mostly not more accurate than peer group averages. We find that fewer heteroskedastic error terms seem to improve prediction accuracy and that SARD groupings can potentially handle MLR.5-6 assumptions better than GICS segmentations and can be tweaked through changing SARD selection criteria.