Stock market forecasting faces severe multi-scale non-stationarity and low signal-to-noise ratios, yet existing VMD-Transformer models suffer from open-loop decoupling (static decomposition followed by fixed attention) and heavy parameter dependency. To address these issues, we propose the Closed-Loop Adaptive Dynamic Frequency-Aware Transformer (CADST), a novel framework that integrates frequency-adaptive VMD with ACF-guided windowing and a Spectral-Gating mechanism modulated by closed-loop Kalman filtering.The two-stage adaptive VMD first computes global spectral entropy (SE) and dominant frequency (MF) to linearly correct the penalty factor α, then selects the optimal mode number K via frequency standard deviation (FSD) maximization, while ACF-based adaptive windowing eliminates scale mismatch and future leakage. A custom SpectralGate layer, driven by real-time Kalman-smoothed spectral states (dominant frequency, entropy, variance ratio), dynamically modulates multi-head attention distributions; the Kalman process noise Q is self-adapted according to validation-loss gradient, enabling closed-loop regime-shift perception.Extensive experiments on four heterogeneous global indices (SZSE, NASDAQ, JKSE, UK100) show CADST outperforms strong baselines (VMD-Transformer, CNN-LSTM, Transformer) by 10–25% in MAE/RMSE/MAPE and 5–15% in directional hit rate. Backtesting yields Sharpe ratios >3.0 with maximum drawdowns <11%, demonstrating superior robustness and economic value. Ablation studies confirm that the closed-loop Spectral-Gating + Kalman feedback contributes the majority of performance gains.

Amid global energy transitions, the stable operation of energy markets is crucial for national economic security. To accurately capture China's energy market volatility, this study employs the CSI 300 Energy Index and its core constituent stocks. We reformulate the highly non-stationary index forecasting problem into a constituent-driven log-return modeling framework, constructing a system-level feature space grounded in a micro-asset correlation network.Existing studies using graph neural networks and signal decomposition often suffer from isolated univariate modeling and static decomposition decoupled from downstream forecasting objectives. Consequently, they struggle to capture complex cross-asset spillover effects and multi-scale spatiotemporal resonance.To overcome these limitations, we propose a constituent-driven closed-loop spatio-temporal-frequency graph fusion forecasting framework (HGFCT). HGFCT integrates task-driven adaptive decomposition (TRG-VMD) with spatial graph topological evolution, introducing endogenous regularization via a closed-loop dynamic temperature annealing mechanism.Experimental results demonstrate HGFCT achieves superior log-return forecasting performance (R^2 = 0.8996, DA = 88.56%). Furthermore, reconstructed index closing prices closely match actual trajectories (R^2= 0.9886, MAPE = 0.6298%). In summary, outperforming both traditional econometric models and advanced graph neural networks, HGFCT provides a reliable paradigm for risk monitoring and intelligent quantitative decision-making in complex energy markets.

With the rapid development of intelligent transportation systems, the increasing number of connected vehicles and onboard sensing devices has generated substantial data traffic and placed stringent demands on communication resources in the Internet of Vehicles. In three-dimensional urban environments, physical blockage, high mobility, and co-channel interference can cause temporary link disruptions, preventing the timely dissemination of safety-critical vehicular information. This study investigates joint resource allocation for Cellular Vehicle-to-Everything (C-V2X) networks with heterogeneous vehicular platoons. We propose an Age of Information (AoI)- and Fairness-Aware Multi-Agent Resource Allocation (AF-MARL) 1 under imperfect and delayed channel state information. Its novelty lies in combining an AoI-dependent power cost with a cross-platoon AoI-dispersion penalty within a multi-task multi-agent reinforcement learning (MARL) architecture, thereby reducing unnecessary high-power transmission by favorable platoons while explicitly protecting disadvantaged platoons without adding Actor or Critic networks. Simulation results show that AF-MARL reduces interference and persistent AoI outages while maintaining reliable Cooperative Awareness Message (CAM) delivery under different platoon spacings and sizes.