Private equity-backed public-to-private transaction activity

: a machine learning approach to quarterly deal volume prediction in the US

Abstract

This thesis addresses a critical gap in private equity research by developing quantitative models to forecast quarterly PE-backed public-to-private (P2P) transaction volumes in the United States. While extensive literature exists on firm-level takeover prediction, systematic approaches to forecasting aggregate PE deal activity remain largely unexplored. This study employs machine learning techniques to predict quarterly P2P deal counts using a comprehensive dataset spanning 1986-2024, incorporating 152 macroeconomic and financial variables. The methodology implements three distinct modeling paradigms4Lasso regression, Random Forest, and XGBoost4within a rigorous time-series cross-validation framework. Variables were systematically processed through principal component analysis to address multicollinearity while preserving economic interpretability. Feature engineering incorporated temporal dynamics through Fourier transforms, momentum indicators, and multi-scale rolling statistics. Results demonstrate modest but consistent predictive improvements over naive baselines, with mean absolute error reductions of 3-9.8% and R² values ranging from 0.145-0.254. Lasso regression emerged as the most balanced approach, maintaining interpretability while achieving competitive accuracy. The analysis reveals strong path dependence in PE activity, with lagged deal counts and momentum indicators consistently ranking as top predictors. Multi-scale temporal patterns, including seasonal and decade-long cycles, contribute significantly to forecast accuracy.

Date of Award	23 Jun 2025
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Dan Tran (Supervisor)