Predicting private equity enterprise multiples using coarsened exact matching
Abstract
This thesis investigates if it is possible to predict accurate and unbiased Net Asset Values for private equity (PE) portfolio companies using multiple valuation. The study is motivated by PE research that has found that General Partners (GPs) under certain circumstances have incentives to exert opportunistic valuations, made possible by the structure of institutional PE where Limited Partners (LPs) rely solely on the self-reported interim Net Asset Values (NAVs) from GPs. First, we construct a novel time series dataset with quarterly company level data for 141 exited portfolio companies in Argentum’s Nordic buyout portfolio from 2002-2020. Second, we gather equivalent data for publicly traded companies in the Nordics and ultimately consolidate the two datasets. We then match portfolio companies with comparable public peer’s contingent on PE selection criteria, using the matching algorithm Coarsened Exact Matching (CEM). The objective is to test if statistical matching methods in combination with prediction models are able to identify representative Nordic peers and enterprise multiples that can be used to indicate unbiased Fair Market Values for portfolio companies given underlying market conditions. We measure the performance of predictions against each portfolio company’s corresponding exit transaction value. Our findings show that particularly one of our prediction models exhibit consistency and seems to predict NAVs with similar accuracy as the GP when moving further than six months prior to exit. There is a large increase in the GPs prediction accuracy between twelve and six months before exit, which is in line with our expectations given GPs informational advantage near exit. In summary, our results suggest that our best performing specification using CEM may provide a consistent and valid second opinion on the Enterprise Value of portfolio companies. In the final section, we explain model limitations and discuss applicableness. Although the peer median model predicts enterprise values with similar aggregated accuracy as the GP in certain periods, it is still frequently inaccurate on company level, and contingent on relatively strict criteria that prune observations. Further, there are confounding variables that we are unable to capture during matching, which would likely have facilitated better prediction accuracy had they been included.