RESEARCH · GLOBAL
arXiv: Graph neural networks improve volatility forecasts but portfolio gains are mixed
Empirical study on 465 S&P 500 equities comparing GraphSAGE (graph NN) to LSTM/AR baselines for realized volatility. Finds lowest forecast error does not yield highest portfolio Sharpe ratio—suggesting model selection for portfolio construction differs from forecast accuracy.
WHY IT MATTERS
Caution against AI model selection bias in portfolio ops. Asset managers must validate end-to-end (forecast → portfolio → returns) rather than optimizing forecast metrics alone. Relevant for risk quant teams building ML pipelines.
Source: arXiv · 2026-05-21