Scholarlink

<p>Accurate financial market prediction requires collaborative learning across institutions that hold sensitive, mutually inaccessible data. Centralized approaches are incompatible with this constraint, while existing federated learning (FL) solutions for finance lack principled defenses against model poisoning, formal privacy guarantees, and regulatory-grade auditability. To address these gaps, we propose <i>FinSecure-FL</i>, a Blockchain-Coordinated Federated Support Vector Regression framework that integrates four complementary components: (1) distributed SVR nodes with adaptive lookback selection and FedProx regularization to handle non-IID market-regime heterogeneity; (2) a four-stage blockchain security pipeline combining norm clipping, Laplace differential privacy, cosine Byzantine Fault Tolerance validation, and SHA-256 hash-chaining; (3) a combined size-and-reputation weighted FedAvg aggregation scheme; (4) and a Proof-of-Authority ledger providing a tamper-evident, timestamped audit trail compliant with MiFID&#xa0;II&#xa0;Article&#xa0;16 and SEC&#xa0;Rule&#xa0;17a-4 record keeping requirements. The framework is evaluated on nine years of NASDAQ Composite daily data partitioned across three non-IID market-regime nodes: Pre-Volatility Baseline, COVID-19 Crisis, and Post-COVID Rate-Hike Regime. Performance is assessed through five regression metrics (MSE, RMSE, MAE, MAPE, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^{2}\)</EquationSource> </InlineEquation>) and Directional Accuracy (DA), complemented by model-complexity analysis. The aggregated global model consistently outperforms both individual local models and two competitive baselines (ARIMA and LSTM) across all metrics and regimes, while the full blockchain pipeline adds negligible computational overhead. These results demonstrate that FinSecure-FL enables the co-design of federated learning and blockchain technology to simultaneously achieve high prediction accuracy, formal privacy guarantees, Byzantine fault tolerance, and regulatory-grade auditability in a multi-institution financial setting. The framework provides actionable forecasts to both local decision-makers (commercial banks, central banks, hedge funds) and global regulators (SEC, ESMA, central bank consortia), establishing a scalable and legally compliant foundation for next-generation financial analytics.</p>

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

FinSecure-FL: a blockchain-secured federated support vector regression framework with adaptive aggregation for privacy-preserving financial market prediction

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Zakia Zouaghia,
  • Zahra Kodia,
  • Lamjed Ben Said

摘要

Accurate financial market prediction requires collaborative learning across institutions that hold sensitive, mutually inaccessible data. Centralized approaches are incompatible with this constraint, while existing federated learning (FL) solutions for finance lack principled defenses against model poisoning, formal privacy guarantees, and regulatory-grade auditability. To address these gaps, we propose FinSecure-FL, a Blockchain-Coordinated Federated Support Vector Regression framework that integrates four complementary components: (1) distributed SVR nodes with adaptive lookback selection and FedProx regularization to handle non-IID market-regime heterogeneity; (2) a four-stage blockchain security pipeline combining norm clipping, Laplace differential privacy, cosine Byzantine Fault Tolerance validation, and SHA-256 hash-chaining; (3) a combined size-and-reputation weighted FedAvg aggregation scheme; (4) and a Proof-of-Authority ledger providing a tamper-evident, timestamped audit trail compliant with MiFID II Article 16 and SEC Rule 17a-4 record keeping requirements. The framework is evaluated on nine years of NASDAQ Composite daily data partitioned across three non-IID market-regime nodes: Pre-Volatility Baseline, COVID-19 Crisis, and Post-COVID Rate-Hike Regime. Performance is assessed through five regression metrics (MSE, RMSE, MAE, MAPE, \(R^{2}\) ) and Directional Accuracy (DA), complemented by model-complexity analysis. The aggregated global model consistently outperforms both individual local models and two competitive baselines (ARIMA and LSTM) across all metrics and regimes, while the full blockchain pipeline adds negligible computational overhead. These results demonstrate that FinSecure-FL enables the co-design of federated learning and blockchain technology to simultaneously achieve high prediction accuracy, formal privacy guarantees, Byzantine fault tolerance, and regulatory-grade auditability in a multi-institution financial setting. The framework provides actionable forecasts to both local decision-makers (commercial banks, central banks, hedge funds) and global regulators (SEC, ESMA, central bank consortia), establishing a scalable and legally compliant foundation for next-generation financial analytics.