# Intelligent Early Warning Model for Financial Risks of New Energy Vehicle Enterprises Based on Deep Learning

MEAL & SPFS: Intelligent Early Warning Model for Financial Risks of New Energy Vehicle Enterprises A unified deep learning framework for policy-aware and interpretable financial risk prediction in new energy vehicle (NEV) enterprises using the Modular Energy Accountability Lattice (MEAL) and the Semantic Policy Fusion Strategy (SPFS). OverviewThis repository implements MEAL and SPFS, two complementary modules for structured financial representation, semantic policy integration, and intelligent early warning of corporate financial distress in NEV enterprises. The framework is designed to model the nonlinear, time-varying, and policy-sensitive financial dynamics of the NEV sector by integrating structured financial indicators, policy-related variables, and ESG/environmental signals into a unified prediction architecture. The project combines symbolic accounting representations with deep learning–based temporal modeling to provide adaptive, interpretable, and regulation-aware financial risk assessment. It is particularly suited for high-volatility industries where deferred subsidies, carbon credits, and policy-driven liabilities materially affect firm-level financial stability. Key Features MEAL (Modular Energy Accountability Lattice): Constructs symbolic and time-indexed accounting representations for NEV firms across assets, liabilities, equity, policy claims, and carbon-related components. Introduces Temporal Asset Tensorization, Policy-Aware Propagation Kernel, and Stochastic Equity Dynamics for modeling financial flows under uncertainty and regulatory change. Supports structured representation of battery-related capital, deferred subsidy assets, carbon credit valuation, and regulatory liabilities. SPFS (Semantic Policy Fusion Strategy): Dynamically maps policy tensors into accounting modules through semantic matching logic. Applies time-decay modulation to capture the diminishing yet persistent effect of historical policy interventions. Enables traceable policy fusion through deterministic semantic hashes and rolling audit chains for interpretability and auditability. Performance: Evaluated on four datasets with strong performance in supervised financial distress classification, including Thomson Reuters Eikon, Financial Distress, CSMAR, and Credit Scoring. On the CSMAR benchmark, the proposed MEAL + SPFS framework achieves 91.12% Accuracy, 87.77% Recall, 89.02% F1 Score, and 92.18% AUC in the main supervised setting; on another benchmark table it reports 90.66% Accuracy and 91.75% AUC on CSMAR depending on the comparison protocol. The model consistently outperforms traditional baselines such as Altman Z-score and logistic regression, as well as deep baselines including AutoEncoder, DAGMM, XGBoost, TabTransformer, LSTM + Dense, and FT-Transformer. Dataset PreparationThis repository is designed to interoperate with multiple financial and cross-domain datasets for experimentation and benchmarking. Below are the main datasets corresponding to the paper’s evaluation section. Dataset Description Link Thomson Reuters Eikon Dataset Comprehensive financial data source with market, statement, and macro-level indicators. General corporate financial distress benchmark Financial Distress Dataset Firm-level financial indicators labeled as distressed or non-distressed. General distress prediction benchmark CSMAR Dataset Chinese listed-company financial database; includes the NEV subset used in this study with policy-sensitive and carbon-related variables. Core NEV evaluation dataset Credit Scoring Dataset Individual-level credit risk dataset. Cross-domain generalization / robustness evaluation The paper explicitly notes that only the CSMAR dataset contains a true NEV subset with authentic policy-related and carbon-related features, while the other datasets are used as general benchmarks or for cross-domain robustness testing. Usage Note:Some datasets used in the study are publicly accessible, while others are subject to third-party licensing or access restrictions. The paper states that raw third-party datasets may not be redistributed directly and should be obtained from their official providers. The archived repository is intended to provide code, preprocessing scripts, feature-construction pipelines, configuration files, and temporal split definitions to support reproducibility. Quick Start Requirements Python ≥ 3.10 PyTorch 2.1 recommended NVIDIA GPU recommended for full experiments Optional experiment tracking via Weights & Biases (WandB) Architecture Overview MEAL ModuleProcesses structured financial statements, accounting variables, policy-adjusted assets, and ESG/environmental indicators through symbolic accounting transformations, temporal tensorization, policy-aware propagation, and stochastic equity modeling. SPFS ModuleUses policy signals for: semantic matching between policy tensors and accounting modules time-decay weighting of historical policy effects traceable fusion through semantic hashes and audit chains End-to-End PipelineDocumented inputs, including financial variables, policy-related variables/text, and ESG/environmental indicators, are harmonized, symbolically mapped, and temporally aligned before entering the MEAL and SPFS modules. Their outputs are fused into a joint representation and passed into a temporal deep learning pipeline for distress prediction, risk exposure evaluation, policy sensitivity analysis, and audit-oriented traceability. Model Accuracy Recall F1 Score AUC Altman Z-score 76.85% 68.90% 70.34% 78.11% Logistic Regression (ratios) 79.91% 72.10% 74.88% 80.90% AutoEncoder 86.44% 82.50% 84.01% 87.24% DAGMM 87.60% 83.81% 85.27% 88.64% MEAL + SPFS 91.12% 87.77% 89.02% 92.18% These results show that MEAL + SPFS substantially improves classification performance, policy sensitivity, and interpretability in both standard and policy-stress evaluation settings. Applications Financial Risk Monitoring: Early warning of distress in NEV enterprises under volatile market and policy conditions. Regulatory Technology: Dynamic monitoring of subsidy-linked risk, carbon-credit exposure, and policy compliance. Sustainable Finance: Integration of ESG and carbon-related signals into financial risk analytics. Investment Decision Support: Audit-friendly, interpretable risk analysis for portfolio screening and policy-sensitive valuation. Future Work Automated policy parsing for real-time adaptation to evolving regulatory environments. Interactive visual dashboards and human-AI collaboration interfaces for financial analysts and regulators. Semi-supervised learning and transfer learning to reduce dependence on fully structured symbolic inputs. Scalable deployment across supply chains, value-chain monitoring, and federated regulatory surveillance. LicenseThis project may be released under the MIT License if the repository author chooses to follow a permissive open-source distribution model. AcknowledgementsThanks to: CSMAR dataset providers for financial and NEV-related benchmark data Thomson Reuters Eikon dataset providers Financial Distress dataset contributors Credit Scoring dataset contributors The original authors of the MEAL + SPFS framework for the conceptual and methodological foundation of this repository.