Relevance
9/10
Directly addresses cross-sectional stock ranking for portfolio construction, a core quantitative trading task. Provides state-of-the-art RankIC and Sharpe ratios on major indices. The factor-based interpretation aligns with institutional portfolio construction workflows. The TopK-DropN portfolio protocol with transaction costs reflects realistic trading constraints. Expert specialization analysis provides insights into regime-dependent strategy behavior. The discrete codebook approach offers a novel inductive bias for noisy financial data.
Implementation Complexity
7/10
Two-stage training with frozen codebook between stages adds complexity. Multiple components (GRU encoder, cross-asset Transformer, VQ codebook, contrastive learning, FiLM-conditioned decoder, temporal Transformer with RoPE, MoE with stochastic gating, loading projection). Market-specific hyperparameters (number of experts, attention heads). However, code is publicly available and architecture is well-documented in the appendix. Training on single GPU (RTX 3090) is feasible.
Reproducibility
4/5
Code is publicly available on GitHub (https://github.com/finxlab/PRISM-VQ). Detailed architecture specifications, hyperparameters, training procedures, and data preprocessing pipelines are provided in the main text and Technical Appendix. Experiments use 5 random seeds. Data sourced from Qlib and JKP Global Factor Library. However, some implementation details (exact codebook initialization, EMA re-initialization procedure) could benefit from more explicit pseudocode.