RESEARCH · GLOBAL
arXiv: Better volatility forecasts don't always yield better portfolios; GNNs underperform on Sharpe
Comparative study (S&P 500, 2015–2025) shows Graph Neural Networks (GNNs) improve realized volatility forecasts vs. traditional models (LSTM, HAR), yet three different models rank first on different metrics: MSE, cross-sectional ranking accuracy, and portfolio Sharpe ratio. Suggests forecasting accuracy ≠ portfolio performance.
WHY IT MATTERS
Warns BFSI quant teams against optimizing single-metric ML models. Multi-objective evaluation is essential: a model with best loss may not generate best risk-adjusted returns. Relevant for portfolio managers using AI for allocation.
Source: arXiv q-fin · 2026-05-21