Local and Global Balance in Financial Correlation Networks: an Application to Investment Decisions

Battle Count: 73

The paper presents a novel network-theoretic criterion for stock selection during financial crises, which is directly relevant to quantitative trading strategies focused on crisis alpha and tail-risk hedging. The approach provides both asset selection and timing signals (when to concentrate vs. diversify). However, the practical implementation requires careful threshold calibration per dataset, the strategy is specifically designed for crisis periods (high global balance), and the paper lacks comparison with established quantitative strategies. The Sharpe ratio improvements are demonstrated but the economic magnitude and statistical significance testing are limited.

The core computation involves: (1) constructing correlation matrices from rolling windows of returns, (2) computing eigenvalues of C and |C|, (3) computing matrix exponentials, (4) calculating global and local balance indices, (5) applying threshold-based selection rules. The eigenvalue decomposition and matrix exponential are standard linear algebra operations available in numerical libraries (NumPy, MATLAB). The rolling window framework and threshold calibration add moderate complexity. The 10,000-iteration random sampling for distribution comparison adds computational overhead. Overall, the mathematical framework is well-defined but requires careful numerical implementation, especially for large N (e.g., NIKKEI with N=199).

The methodology is clearly described with explicit formulas for global and local balance. Data sources (DAX, ESX, FTSE, NIKKEI daily log returns from 2005-2020) are standard financial indices. However, no code repository is provided, specific threshold values are dataset-dependent, and the rolling window parameters (ΔT ≥ N, Δt = 10 days) are stated but implementation details for matrix exponential computation and eigenvalue decomposition are not fully specified. The 10,000 random sampling procedure for comparison is described but seed values are not provided.