Portfolio Optimization Proxies under Label Scarcity and Regime Shifts — Bayesian and Deterministic Students under Semi-Supervised Sandwich Training

Battle Count: 71

Directly addresses portfolio construction, risk management, and trading under realistic constraints. The framework is designed for weekly rebalanced long-only portfolios with turnover caps, position limits, and transaction costs. Key findings (implicit turnover regularization, HIGHVOL paradox, constraint sensitivity) are directly actionable for quantitative portfolio managers. The retail deployment focus and ETF-based universe make it practically relevant. The Bayesian uncertainty quantification addresses a critical gap in production trading systems.

Requires implementing: (1) CVaR optimization as teacher, (2) variational Bayesian neural networks with reparameterization trick, (3) three-stage sandwich training schedule, (4) factor-based synthetic data generation with VAR(1) and t-copula, (5) 576-dimensional feature engineering with rolling ridge regression and PCA, (6) rolling fine-tuning with reset protocol, (7) multi-tier evaluation (GRID, C2A, D2A) with constraint operators. The architecture itself (feedforward + softmax) is simple, but the full pipeline with synthetic generation, Bayesian inference, and evaluation protocol is substantial.

The paper provides extensive appendices detailing the full data pipeline, synthetic generation mechanics, feature construction, constraint operators, and evaluation protocol. However, no GitHub repository or code link is provided. The base sandwich training codebase is attributed to Pareek et al. (2025) but not directly linked. Data is sourced from yfinance (publicly available). Reproduction would require implementing the full pipeline from scratch based on the detailed specifications.