An Investment Analysis Research Enhanced Enterprise Development Algorithm for Bankruptcy Prediction

Source abstract

Corporate bankruptcy prediction requires reliable optimization methods capable of identifying effective machine-learning configurations while maintaining robust global search performance. This study proposes an investment-analysis-inspired Enhanced Enterprise Development Optimizer (EED) for global optimization and KNN-based bankruptcy prediction. Rather than introducing additional generic perturbation operators, EED incorporates historical performance appraisal, covariance- and correlation-based search allocation, and drawdown-aware adaptive recovery into the original Enterprise Development Optimizer. The optimization performance of EED is evaluated on the CEC2017 and CEC2022 benchmark suites under five dimensional settings and compared with eleven competing algorithms. For CEC2017, F2 is excluded because of documented implementation inconsistencies and unstable numerical behavior, and the remaining 29 functions, namely F1 and F3–F30, are used in the evaluation. Each algorithm is independently executed 30 times, and the results are assessed using the mean and standard deviation, convergence and distribution analyses, the Friedman rank test, and the Wilcoxon signed-rank test. EED achieves the best or tied-best mean performance on 28, 27, and 23 of the 29 retained CEC2017 functions at 30, 50, and 100 dimensions, respectively, and on 9 and 8 of the 12 CEC2022 functions at 10 and 20 dimensions. Its overall Friedman mean rank is approximately 1.41, while the Wilcoxon comparison with the original ED yields 107 wins, 3 ties, and 1 loss, indicating that the proposed mechanisms provide consistent improvements across different benchmark landscapes rather than gains on only a small number of functions. To examine practical applicability, EED is further used to optimize the neighborhood size and distance measure of KNN on the Wieslaw and JPNBDS bankruptcy datasets. The resulting EED-KNN models show competitive and relatively stable performance in ACC, MCC, F1-score, and precision. The benchmark experiments therefore establish the general search capability and robustness of EED, while the bankruptcy-prediction application demonstrates that these optimization advantages can be transferred to a discrete machine-learning parameter-selection task.