Scopus Indexed Publications

Paper Details


Title
Mitigating Market Complexity Through Synergistic Neural Networks: A Hybrid Deep Learning Framework for Multihorizon Insolvency Forecasting

Author
, Md. Tofayel Ahmed Bhuiyan, Mohammad Azam Khan,

Email

Abstract

Corporate insolvency forecasting is a critical component of financial risk management, yet predictive reliability is frequently compromised by severe class imbalance and nonstationary market volatility. Conventional statistical methods and standard recurrent neural networks often exhibit computational inefficiencies when modeling extended temporal dependencies within financial ratios. To overcome these structural limitations, a hybrid deep learning architecture designated as the Mamba-inspired synergistic transformer network (MST-Net) is proposed. This framework integrates a custom selective state space model for efficient temporal dynamic modeling with transformer encoders to capture global contextual awareness. A synergistic gating mechanism dynamically fuses outputs from both architectural branches. To address inherent data irregularities, a rigorous preprocessing pipeline utilizing nonlinear quantile transformations normalizes heavy-tailed financial distributions. Furthermore, a specialized asymmetric loss function decouples gradient influence from easy negative samples to prioritize hard-to-classify distress signals during training. Empirical validation was conducted on a comprehensive longitudinal dataset from major American stock exchanges, spanning 20 years with over 78,000 firm-year observations. The proposed framework was benchmarked against traditional approaches and advanced tree-based ensembles, including XGBoost and Random Forest. For the challenging 3-year prediction window, the synergistic network achieved a macro-F1 score of 0.7595 and an ROC–AUC of 0.9053, outperforming the strongest baseline XGBoost by a significant margin in balanced classification metrics. Additionally, a Matthews correlation coefficient of 0.5288 confirmed a highly balanced binary agreement despite pronounced data sparsity. Beyond quantitative performance, model transparency is established through SHAP and LIME algorithms, validating the applicability of the system for regulatory compliance.


Keywords

Journal or Conference Name
Applied Computational Intelligence and Soft Computing

Publication Year
2026

Indexing
scopus