Beyond linear models: Explainable AI for merger & acquisition outcome prediction

Title (eng)
Beyond linear models: Explainable AI for merger & acquisition outcome prediction
Author
Alexander Ryusandi Pratama
Advisor
Adam Zaremba
Advisor
Ruihai Dong
Description (eng)
This study investigates the effi cacy of the Tabular Prior-informed Bayesian Neural Network (TabPFN) for predicting Merger and Acquisition (M&A) outcomes, positioning it as a transparent alternative to opaque "black-box" models within regulated fi nancial domains. The performance of TabPFN is benchmarked against fi ve established models: Logistic Regression (LR), Random Forest (RF), Gradient Boosting (GB), Support Vector Machine (SVM), and Multilayer Perceptron (MLP) on a curated dataset of 4,400 historical U.S. M&A transactions. The central evaluation assesses predictive discrimination (AUC-ROC) and probability calibration (Expected Calibration Error). Results demonstrate that TabPFN achieves superior discriminative performance (AUC-ROC: 0.9168) and strong calibration (ECE: 0.0676). Furthermore, SHAP-based explainability analysis of the TabPFN model reveals crucial non-linear feature eff ects, such as the interplay between deal attitude and completion time, that are not apparent in linear models. The fi ndings establish TabPFN as a robust and interpretable framework for M&A risk assessment, aligning with the principles of trustworthy AI for high-stakes fi nancial applications. The primary aim was to assess diff erences in predictive discrimination (AUC-ROC), probability calibration (ECE), and explanatory insights. Empirical fi ndings revealed TabPFN achieved the highest discriminative performance (AUC-ROC: 0.9168), outperforming LR (0.7171) and other baselines. TabPFN also demonstrated strong probability calibration (ECE: 0.0676), signifi cantly better than LR (ECE: 0.1372), although RF exhibited the best ECE (0.0299). SHAP analysis of TabPFN identifi ed 'realized_return' as a key predictor for the tested instances and elucidated crucial non-linear feature eff ects not apparent from linear models; the initial feature pruning, guided by Random Forest, highlighted the general importance of factors including 'attitude_Friendly' and 'completion_time' for the prediction task. This study concludes that TabPFN's superior discriminative performance and strong calibration, synergistically combined with XAI techniques, provide a more valuable and actionable tool for M&A strategic decision-making. The enhanced probabilistic reliability and transparent, granular feature-level explanations off er a robust foundation for risk assessment within the TISE framework, aligning with demands for trustworthy AI in high-stakes fi nancial applications
Keywords (eng)
Mergers and Acquisitions (M&A) Outcome PredictionExplainable Artifi cial Intelligence (XAI)Tabular Prior-informed Bayesian Neural Network (TabPFN)Probability CalibrationSHapley Additive exPlanations (SHAP)
Type (eng)
Master theses
Language
[eng]
Persistent identifier
https://door.donau-uni.ac.at/o:5729
Date created
2025
License
All rights reserved