Nyström Attention Matches Full Attention for Cross-Sectional Stock Prediction

Battle Count: 67

Highly relevant to quantitative trading. The paper directly addresses cross-sectional stock ranking (a core quant task), provides practical guidance on efficient attention for scaling stock Transformers, reveals that MASTER's inter-stock attention implements an implicit 6-factor risk model (volatility, momentum, sector rotation, defensive/cyclical, liquidity, reversal), and demonstrates that attention anti-correlates with return similarity (complementarity-seeking). The Nyström result enables O(mN) cross-stock attention at equivalent predictive quality, critical for large-universe strategies. The finding that cross-stock modules do not outperform per-stock LSTM at N≈3,500 is practically important for strategy design. The IC vs. Rank IC dissociation has direct implications for return-focused vs. ranking-focused strategies.

Moderate complexity. The Nyström approximation itself is straightforward (m=32 landmarks, 6 Newton-Schulz iterations for pseudo-inverse, O(mN) computation). However, the full analytical framework (spectral decomposition, TOST equivalence testing, multi-seed validation, controlled ablations across 9+ configurations) requires substantial experimental infrastructure. The MASTER architecture has 5 stages with 775K parameters. Implementing all Step 3 variants (GraphMask, GraphBias, GCN, TopK, DeviationAttn, LearnedLowRank, StaticAttn, UniformAttn, MeanPool, Performer) as subclasses adds engineering effort. The large-scale experiments require adapted architectures and Qlib infrastructure.

The paper uses the MASTER open-source CSI300 dataset and Qlib for full A-share data. All Step 3 variants are implemented as subclasses without modifying the original codebase. Multi-seed validation (5-10 seeds) with TOST testing is provided for key claims. However, some mechanistic decomposition results rely on single-seed (seed 0) analysis. The large-scale experiments use adapted architectures (17 features, d_model=64, no market gating) rather than the full original architecture, limiting direct comparability. No explicit GitHub repository URL is provided in the extract.