WaVeFuse: Regime-Adaptive Equity Index Forecasting via Channel-Wise Wavelet Denoising and Vertical Attention Fusion

Battle Count: 65

Highly relevant to quantitative trading. The paper directly addresses equity index forecasting with directional accuracy exceeding 70% on all four indices, implements realistic backtesting with transaction costs (10 bps), reports Sharpe ratios (mean 3.69), maximum drawdown control (7.5% during COVID-19), and demonstrates regime-adaptive behavior via VAF. The architecture is computationally efficient (152k parameters, 0.68MB, sub-1.3ms GPU inference), making it suitable for production deployment. The walk-forward validation protocol mimics live deployment conditions. The risk-adjusted strategy variant with volatility-conditioned entry filters provides practical trading signal generation. However, results are based on a single 365-day test window and should not be extrapolated to multi-year live deployment without further validation.

Moderate-to-high complexity. The architecture involves multiple specialized components: Sym-4 DWT denoising with MAD soft thresholding, seven technical indicator computations, channel-wise CWT with Morlet wavelet at 32 scales, CNN-BiLSTM encoder, dual-layer Transformer with attention pooling, 2-token VAF gating mechanism, and BiLSTM decoder. Requires expertise in wavelet signal processing, deep learning (CNN, LSTM, Transformer), and financial feature engineering. However, the model is lightweight (152k parameters, 0.68MB), trains in ~100 seconds, and uses standard TensorFlow/Keras components. The WFV protocol adds implementation overhead. No code repository is provided, requiring reimplementation from the detailed mathematical formulations and algorithm pseudocode.

The paper provides detailed hyperparameters (Table 6), algorithm pseudocode (Algorithms 1 and 2), fixed random seed (42), specific hardware (AMD Ryzen 9 9900X, NVIDIA RTX 5060 Ti, 32GB RAM, Python 3.10, TensorFlow 2.10, CUDA 12.9), and complete mathematical formulations. Data sourced from Yahoo Finance (publicly available). However, no GitHub repository link is provided, and the exact preprocessing code is not shared. The WFV protocol is fully specified with window sizes and step parameters.