A Novel approach to portfolio construction

Battle Count: 140

Highly relevant to quantitative trading and portfolio management. BPASGM provides a principled, data-driven approach to asset selection that directly addresses the estimation error problem in high-dimensional portfolio construction. The method is computationally efficient, does not require covariance matrix inversion, handles both linear and non-linear dependencies, and produces sparse portfolios with improved risk-adjusted returns. The empirical application to US equities, global indices, and FX rates demonstrates practical applicability. However, the method is primarily a preprocessing/screening tool rather than a trading signal generator, and the aggressive cardinality reduction (358 to 3 assets) may limit its direct use in diversified trading strategies. The out-of-sample validation period is short, and the method's performance under stress/regime changes is not fully explored.

Moderate to high complexity. The method requires: (1) construction of minimal AIC/BIC trees via HDGM (Edwards et al., 2010), (2) iterative application of BPA to each variable, (3) construction of signed adjacency matrices using covariance signs, (4) computation of Hadamard products for direct/indirect/simple link detection, (5) a three-step pruning procedure with performance-based criteria, (6) re-estimation of portfolio moments on the selected subset, and (7) standard mean-variance optimization. The information-theoretic components (mutual information, KL divergence, entropy coefficient) require careful numerical implementation. The DCC-GARCH validation adds further complexity. However, the modular structure (screening + optimization) makes it adaptable, and the BPA component has been previously published.

The paper provides detailed mathematical formulations, algorithmic steps, and a complete data-generating process for simulations. However, no code repository is mentioned. The empirical application uses Refinitiv Workspace data (commercial, not freely available). The BPA component references prior work (Riso et al., 2023) for implementation details. The three-step pruning procedure is well-described but requires careful implementation of signed adjacency matrices, Hadamard products, and link-removal criteria. Monte Carlo simulation parameters are fully specified in Appendix D.