Adaptive Weighting and Deep Neural Networks for Automated Multi-Indicator Financial Statement Analysis and Risk Prediction

Abstract

This study proposes an innovative financial statement analysis model combining deep neural networks with an adaptive weighting algorithm. The model includes five hidden layers with neuron counts of 128, 256, 128, 64, and 32, and applies an adaptive weighting mechanism that dynamically adjusts feature importance using the coefficient of variation and Pearson correlation. The dataset consists of 8,500 financial records from companies across eight industries, spanning from 2005 to 2023, and includes over 20 key indicators from the balance sheet, income statement, and cash flow statement. The model was evaluated against traditional approaches, including support vector machines (SVM), random forest, and Transformer-based models. Results demonstrate that the proposed model achieves 90% accuracy in financial risk prediction, outperforming SVM by 12%, with an F1 score of 87% and RMSE reduced to 0.06. This highlights the model’s effectiveness and robustness in handling complex financial data. In financial risk prediction (a classification task), the model achieved an average accuracy of 88%, recall of 85%, and F1 score of 86.5% across 50 experimental runs. For profitability analysis (a regression task), the model reduced RMSE to 0.045. These results outperform traditional baselines such as logistic regression and SVM, and approach the performance of emerging Transformer-based models, demonstrating both predictive effectiveness and generalizability across industries.