Do Better Volatility Forecasts Lead to Better Portfolios? Evidence from Graph Neural Networks

Battle Count: 84

Highly relevant to quantitative trading. The paper directly addresses the critical question of whether improved volatility forecasts translate to better portfolio performance, with practical implications for position sizing, risk management, and portfolio construction. The finding that MSE-optimal, ranking-optimal, and Sharpe-optimal models differ is directly actionable for practitioners. The comparison of four portfolio constructions (min-var, inverse-vol, vol-targeting, long-short) with realistic transaction costs provides immediately applicable guidance. The HAR baseline's competitiveness is a practical warning against over-engineering.

Moderate to high complexity. Requires implementing GraphSAGE with neighborhood aggregation, constructing three different graph types (rolling correlation, sector, Granger causality with Bonferroni correction), managing macro feature engineering, and running four distinct portfolio optimizations. The correlation graph precomputation and caching, Granger causality testing across all ordered pairs (465×464), and portfolio simulation with transaction costs add engineering overhead. However, the architecture is standardized (shared GraphSAGE backbone) and hyperparameters are well-documented.

Code is publicly available on GitHub (https://github.com/waderylan/sp500-gnn). Data sourced from yfinance (publicly available). Hyperparameters and search spaces are fully documented in Table 1. Chronological train/validation/test splits are clearly specified. However, some implementation details of the GraphSAGE backbone and portfolio optimizer may require code inspection. The 465-stock universe filtering criteria are stated but exact constituent list is not provided.