DSTNet: Dynamic Spectral Trajectory Network for Causal Multi-Horizon Financial Forecasting

Battle Count: 50

Highly relevant for academic research into causal deep learning architectures for finance. However, the paper explicitly reports a negative result for tradable alpha in equities, suggesting that while the forecasting metrics improve over baselines, the signal may not be economically viable for trading strategies without further refinement or different asset classes (like Gold).

The architecture is complex, involving custom causal FIR filter banks, burn-in exclusions, factorized attention mechanisms, and multi-branch fusion. Reproducing the exact causal guarantees and preprocessing steps requires careful implementation.

The paper provides detailed hyperparameters, algorithm pseudocode, and specifies the use of public data (Yahoo Finance). Code availability is stated as 'upon acceptance', which limits immediate reproducibility but suggests high intent for future availability. Deterministic settings and seed stability analysis are included.