Forecasting mergers and acquisitions failure based on partial-sigmoid neural network and feature selection.

Area of Science:

• Business Analytics
• Computational Finance
• Machine Learning

Background:

• Merger and Acquisition (M&A) forecasting faces challenges due to imbalanced datasets, with failure cases significantly outnumbered by successful ones (82% vs. 18%).
• The high dimensionality of bidder and target features complicates accurate M&A forecasting.
• Existing forecasting models often overlook crucial financial indicators relevant to M&A success.

Purpose of the Study:

• To propose a novel forecasting approach for M&A data using a partial-sigmoid neural network (PSNN).
• To evaluate the efficacy of three distinct feature selection methods: chi-square (chi2) test, information gain, and gradient boosting decision tree (GBDT).
• To enhance the accuracy and reliability of M&A failure prediction by addressing data imbalance and feature selection issues.

Main Methods:

• Development of a partial-sigmoid neural network (PSNN) incorporating a partial-sigmoid activation function in the output layer.
• Comparative analysis of chi-square (chi2) test, information gain, and gradient boosting decision tree (GBDT) for feature selection.
• Empirical evaluation of the proposed PSNN and feature selection techniques on M&A datasets.

Main Results:

• The PSNN model demonstrated significant improvements in forecasting metrics, including precision (up to 0.37), recall (0.49), G-Mean (0.41), and F1-measure (0.23).
• Feature selection methods collectively improved forecasting accuracy by 1.83% to 13.16%.
• The chi-square (chi2) test emerged as the most effective feature selection method among those evaluated.

Conclusions:

• The proposed PSNN and feature selection approach effectively mitigates the negative impacts of data imbalance and high dimensionality in M&A forecasting.
• Key overlooked features for predicting M&A failure include prior year's assets, market value, and capital expenditure.
• The chi-square (chi2) test is recommended for feature selection in M&A forecasting due to its superior performance.