DSTNet: Dynamic Spectral Trajectory Network for Causal Multi-Horizon Financial Forecasting
DSTNet introduces a causal spectral trajectory representation that retains the recent evolution of wavelet filter-bank magnitudes rather than only denoising or using a single-point snapshot, combined with a scale-temporal factored Transformer and gated fusion. On seven equity indices and gold under an expanding-window protocol, the model beats the random-walk persistence benchmark in all 32 series-horizon cells, but paired testing is inconclusive against the strongest competi
01 ABSTRACT
The work addresses the common limitation of wavelet-based financial forecasters that use the transform only for denoising or reduce it to a single spectral snapshot at the forecast origin, with a bilateral convolution that can read past that origin; it proposes a one-sided Morlet-derived filter bank with explicit burn-in to build a causal Dynamic Spectral Trajectory. Evaluated on seven equity indices and gold with an expanding-window protocol and an untouched one-year hold-out against nine learned baselines and a persistence benchmark, DSTNet is reported to be the only model below persistence under MAE and MAPE in every cell. The authors also report a negative downstream allocation diagnostic and do not claim an equity-timing advantage.
02 KEY FINDINGS
- Constructs a causal Dynamic Spectral Trajectory from seven trailing technical indicators over a twenty-day lookback using a one-sided Morlet-derived filter bank with explicit left-boundary burn-in.
- Uses a factorized Scale-Temporal Spectral Transformer, a learned gate fusing with a CNN-BiLSTM, and horizon-specific gates to emit 1, 3, 5, and 10 day forecasts in a single pass.
- Across 32 series-horizon cells, persistence is the strongest of ten fixed competitors in 29 cells, and DSTNet is the only model below it in every cell.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG · ARXIV Q-FIN