Scholarlink

Accurate, resilient, and privacy-compliant financial forecasting remains a major challenge in high-stakes environments characterized by data sparsity, volatility, and regulatory constraints. This paper presents a novel multi-objective deep learning framework that integrates Generative Adversarial Networks (GANs) with Long Short-Term Memory (LSTM) networks, optimized concurrently for prediction accuracy, output stability, and differential privacy. Unlike traditional single-objective models, the proposed GAN-LSTM hybrid utilizes a custom loss function that balances accuracy, Mean Squared Error (MSE), volatility-sensitive stability, and privacy penalties via differential privacy (DP). The model is trained on a composite dataset comprising stock prices (Yahoo Finance), macroeconomic indicators (FRED), sentiment scores (FinBERT), and GAN-generated synthetic data to ensure structural and contextual robustness. Experimental results across global indices such as FTSE and Nikkei demonstrate superior performance, achieving up to 18% lower RMSE and enhanced resilience under market regime shifts. Privacy-preserving efficacy is validated through adversarial attack simulations and ε-scores, while interpretability is ensured using SHAP-based feature attributions. An extensive ablation study confirms the necessity of each loss component. This unified, multi-criteria approach positions the GAN-LSTM framework as a deployment-ready solution for ethically aligned and regulation-compliant financial forecasting systems. The framework's performance is further validated through a comprehensive benchmarking study against recent state-of-the-art architectures, where it achieved a superior composite score by balancing accuracy, stability, and privacy.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Privacy-Preserving, Stability-Aware GAN–LSTM Framework for Multi-objective Financial Time Series Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Gudla Anjaneyulu,
  • Pappula Praveen

摘要

Accurate, resilient, and privacy-compliant financial forecasting remains a major challenge in high-stakes environments characterized by data sparsity, volatility, and regulatory constraints. This paper presents a novel multi-objective deep learning framework that integrates Generative Adversarial Networks (GANs) with Long Short-Term Memory (LSTM) networks, optimized concurrently for prediction accuracy, output stability, and differential privacy. Unlike traditional single-objective models, the proposed GAN-LSTM hybrid utilizes a custom loss function that balances accuracy, Mean Squared Error (MSE), volatility-sensitive stability, and privacy penalties via differential privacy (DP). The model is trained on a composite dataset comprising stock prices (Yahoo Finance), macroeconomic indicators (FRED), sentiment scores (FinBERT), and GAN-generated synthetic data to ensure structural and contextual robustness. Experimental results across global indices such as FTSE and Nikkei demonstrate superior performance, achieving up to 18% lower RMSE and enhanced resilience under market regime shifts. Privacy-preserving efficacy is validated through adversarial attack simulations and ε-scores, while interpretability is ensured using SHAP-based feature attributions. An extensive ablation study confirms the necessity of each loss component. This unified, multi-criteria approach positions the GAN-LSTM framework as a deployment-ready solution for ethically aligned and regulation-compliant financial forecasting systems. The framework's performance is further validated through a comprehensive benchmarking study against recent state-of-the-art architectures, where it achieved a superior composite score by balancing accuracy, stability, and privacy.