Model-based evaluation of synthetic financial time series data: A comparative study with multi-metric validation

Abstract

This research presents a behaviorally informed framework for synthesizing financial time-series data, specifically designed to emulate the complex dynamics of foreign exchange markets. Deviating from conventional generative adversarial networks (GANs) or purely statistical distribution-matching, the proposed methodology adopts a game-theoretic architecture. This framework integrates trader-interaction dynamics, stochastic strategies, and information asymmetry, treating the market as a strategic participant to reproduce authentic volatility patterns and structural dependencies. To ensure numerical stability across extensive simulations, the study introduces a uniform upward scaling procedure and controlled initialization, preventing pathological price behaviors without compromising the underlying statistical properties. The framework’s analytical fidelity was rigorously evaluated against a suite of econometric and machine learning models, including ARIMA, XGBoost, LSTM, N-BEATS, and DLinear. Experimental results involving 12,960 hourly observations demonstrate that the synthetic data maintains strong alignment with empirical benchmarks. DLinear emerged as the superior model, exhibiting exceptional stability with an R2 frequently exceeding 0.98 and a Mean Absolute Scaled Error (MASE) near unity. While XGBoost and N-BEATS yielded competitive results, ARIMA and LSTM showed anticipated performance degradation due to temporal noise. Comprehensive residual diagnostics, including Ljung-Box tests and stationarity assessments, confirm that the generated series are behaviorally consistent and analytically reliable. This framework thus provides a robust foundation for comparative modeling and experimental financial research. • Privacy-Preserving Synthetic Data for Forex Analysis This study addresses data privacy challenges by demonstrating that behaviorally informed synthetic forex data can serve as a viable alternative to real trading data, supporting regulatory compliance and broader research accessibility. • Game-Theoretic Generation of Financial Time Series A game-theoretic framework simulates trader-market interactions to generate synthetic forex data that retain behavioral and structural realism beyond conventional statistical or generative approaches. • Stylized Fact Replication and Structural Fidelity Assessment The synthetic series reproduce key stylized facts, non-stationarity, volatility clustering, and heavy-tailed return distributions, and are further validated through Wasserstein, Hausdorff, and Cramér’s V similarity measures. • Model-Based Validation with ARIMA, LSTM, XGBoost, N-BEATS and DLINEAR Predictive transferability between real and synthetic data is evaluated using five forecasting models. DLinear and XGBoost achieve the highest accuracy, demonstrating robustness across both datasets. LSTM effectively captures sequential dependencies, while ARIMA serves as a statistical baseline. N-BEATS provides an additional deep learning perspective. The inclusion of DLinear notably enhances the experimental depth and strengthens conclusions on model performance and synthetic data fidelity. • Limitations and Future Directions for Trading Applications Although high predictive consistency is achieved, predictive accuracy alone does not ensure trading profitability. Future work will incorporate volatility-aware models, profit-based metrics, and cross-currency generalization to enhance real-world applicability.