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

As a fastidious and diligent AI researcher, I have meticulously analyzed both provided text segments (A and B) pertaining to the paper DSTNet: Dynamic Spectral Trajectory Network for Causal Multi-Horizon Financial Forecasting. The following is a comprehensive, detailed synthesis of the paper's methodology, architecture, results, and key conclusions.

Detailed Research Summary: DSTNet for Causal Multi-Horizon Financial Forecasting

The DSTNet model is a sophisticated deep learning architecture designed specifically for causal multi-horizon financial forecasting. Its core innovation lies in its ability to capture the evolution of spectral features over time while maintaining strict causality, allowing it to generate multiple future forecasts simultaneously in a single forward pass.

I. Core Architectural Philosophy and Causal Constraints

The paper establishes a rigorous causal framework from the outset:

• Causal Front End: The architecture is strictly causal. This is enforced through the use of Morlet-derived filters, which ensure that the model reads no observation after its forecast origin (as detailed in Proposition 1). Corollary 1 further addresses this by fixing a 192-sample exclusion to remove left-boundary transients.

• Multi-Horizon Output: DSTNet is designed to emit forecasts for four distinct horizons (H=1, 3, 5, and 10 days) in a single forward pass.

II. Key Architectural Components

The DSTNet architecture decomposes the multi-horizon forecasting problem into four sequential operations:

1. Denoiser:

• The input OHLCV channels are first processed by a causal undecimated Symlet-4 residual multiresolution denoiser. This step aims to clean the raw financial data before feature extraction.

2. Time-Frequency Representation:

• A strictly causal time-frequency representation is derived using a Morlet-derived complex FIR filter bank. This allows the model to analyze different temporal and spectral characteristics of the input data simultaneously.

3. Dynamic Spectral Trajectory Tensor D(i):

• This is the central innovation of DSTNet. It retains the full evolution of filter-bank magnitudes as a per-sample tensor, denoted as D(i) in R w times S times M.

• Definition: D(i) is defined as the sequence of filter bank outputs across a complete lookback window: D(i) = W[t i - w + 1,:,:], W[t i - w + 2,:,:],, W[t i,:,:].

• This tensor explicitly encodes the temporal evolution of spectral responses over a lookback window (w) across S fixed FIR scales and M technical indicators.

4. Dual-Branch Network:

The core feature extraction is handled by a dual-branch structure:

• Temporal Branch (CNN-BiLSTM): This branch produces a temporal summary, denoted as f(i) 1.

• Spectral Branch (Factorized STST): This branch produces a fused spectral summary, denoted as f(i) 2.

These two summaries are combined by a learned vertical gate (Vertical Attention Fusion - VAF) before being passed to the final decoder.

• Factorized Scale-Temporal Spectral Transformer (STST): The spectral branch employs this mechanism by running two parallel factorized attention streams:

• Stream A models temporal dependencies.

• Stream B models scale dependencies.

• These streams are integrated via a cross-stream Multi-Layer Perceptron (MLP).

5. Horizon-Adaptive Decoder (HAD):

The final stage is the HAD, which manages the multi-horizon output:

• For each horizon h in H, an independent gating layer with activation is learned: g(i) h = (W g,h f(i) fus + b g,h).

• This gate modulates the fused summary (f fus) specifically for that horizon, yielding the final forecast h.

III. Performance Evaluation and Comparative Analysis

The paper rigorously evaluates DSTNet against nine learned baselines and a crucial benchmark: random-walk persistence.

Key Findings:

• Comparison to Persistence: The most decisive comparison is against random walk persistence. The results consistently show that none of the nine learned baselines beat the random-walk persistence benchmark for one, three, or five days on any series. Furthermore, DSTNet itself performs below this persistence benchmark in every cell when measured by MAE and MAPE (e.g., DSTNet is below it in every cell).

• Allocation Diagnostic: The allocation diagnostic does not support an equity-timing advantage for any of the seven tested indices.

• Long-Term Performance Caveat: The paper notes that the random-walk claim does not extend to RMSE at long horizons.

IV. Quantitative Results (Extracted Data)

The provided tables offer concrete quantitative evidence of DSTNet's performance:

Table XVII (RMSE Across Seeds):

This table shows the Root Mean Square Error (RMSE) for DSTNet across different seed runs and various horizons (H=1, 3, 5, 10) for major indices. For instance, on the KOSPI index:

• H=1: 26.76 plus or minus 0.78

• H=10: 82.05 plus or minus 5.92

Table XVIII (Panel-Mean Ablation Summary):

This table quantifies the degradation in Mean Absolute Error (MAE) relative to the full DSTNet model across all eight series and four horizons. The results highlight the importance of its components:

• Without burn-in exclusion, MAE degrades by +0.92%.

• Without the VAF gate, MAE degrades by +2.94%.

• Using a static snapshot (No-DST) results in a significant degradation of +5.79% compared to the full model.

Table XIX (Full Panel Ablation Study):

This table provides detailed MAE comparisons for specific series and horizons, confirming the relative performance of the components:

• For KOSPI, at H=1, DSTNet's MAE is 88.92%, which is significantly lower than the static snapshot baseline (92.42%).

V. Final Conclusion Synthesis

The definitive conclusion of the paper rests on the following synthesis: DSTNet successfully captures the evolution of filter-bank magnitudes as a per-sample Dynamic Spectral Trajectory tensor, encodes this tensor with a factorized Scale-Temporal Spectral Transformer (STST), and achieves four multi-horizon forecasts in one forward pass.

However, despite its sophisticated design and performance metrics shown in the tables, the ultimate deciding factor for the paper is its comparison against simple random walk persistence. The model's inability to outperform persistence across tested horizons suggests that while DSTNet is a highly advanced forecasting tool, it does not currently provide a demonstrable predictive edge over naive persistence strategies on these specific financial series under MAE and MAPE metrics.

1. Bold header: Causal Spectral Trajectory Representation

DSTNet retains the recent evolution of filter-bank magnitudes as a causal Dynamic Spectral Trajectory, which allows the model to exploit changes in spectral structure over time, moving beyond static snapshots that retain only a single temporal slice.

2. Bold header: Factorized Scale-Temporal Spectral Transformer (STST)

The system employs a factorization where it models temporal dependencies and cross-scale interactions within the DST representation separately, reducing attention complexity from O((wS)2d) to O((w2 + S2)d). This allows for more efficient processing of the high-dimensional spectral data.

3. Bold header: Horizon-Adaptive Decoder (HAD)

The HAD applies an independent learned feature-modulation gate before each horizon head, allowing the model to modulate the fused representation separately for each forecast scale, rather than assuming a fixed frequency-to-horizon correspondence.

4. Bold header: Causal End-to-End Pipeline

The system implements a strictly causal end-to-end time–frequency processing pipeline designed to prevent future-information leakage during forecasting, ensuring the spectral transformation depends only on observations available at or before the forecast origin.

Wavelet-based financial forecasters typically use the transform only to denoise, or reduce it to a single spectral snapshot at the forecast origin, and the convolution that produces the coefficients is usually bilateral, so it can read past the forecast origin. DSTNet instead retains the recent evolution of filter-bank magnitudes as a causal Dynamic Spectral Trajectory, built from seven trailing technical indicators over a twenty-day lookback with a one-sided Morlet-derived filter bank and an explicit burn-in for the left-boundary transient. A factorized Scale-Temporal Spectral Transformer attends along the time and filter-bank axes separately, a learned gate fuses the spectral branch with a CNN-BiLSTM, and horizon-specific gates emit one, three, five, and ten day forecasts in a single pass. We evaluate seven equity indices and gold under a common expanding-window protocol and an untouched one-year hold-out, against nine learned baselines and a random-walk persistence benchmark. Under MAE and MAPE, persistence is the strongest of the ten fixed competitors in 29 of the 32 series-horizon cells and DSTNet is the only model below it in every cell, by 0.7 to 0.9 percent at one day and 3.4 to 4.5 percent at ten days. At one day, paired testing favours DSTNet against the weaker learned baselines but is inconclusive against persistence and the strongest learned forecasters. A downstream allocation diagnostic does not support an equity-timing advantage on any of the seven indices.