GNNs for cross-asset risk
Practitioners showcased graph neural networks to encode relationships across assets for adaptive position sizing and risk contagion mapping—GNNs are being used more for risk allocation than raw alpha prediction. The implementation path they recommended includes PyTorch Geometric and visualization of spillover dynamics. (youtube.com)
An International Journal of Forecasting paper (2025) introduced GNN-based realized-volatility models that incorporate multi-hop spillovers and published accompanying Python code on GitHub (project GNNHAR) for 5‑minute to daily aggregation and out‑of‑sample evaluation. (github.com) An arXiv working paper from September 2024 by Lukas Gonon et al. formalized systemic‑risk computation using permutation‑equivariant GNNs and reported numerical experiments where GNN/PENN architectures outperformed benchmark allocation methods on default‑cascade and bailout‑allocation tasks. (arxiv.org) A Springer article published 24 January 2026 (TVP‑VAR‑GNN) mapped time‑varying cross‑asset connectedness across stocks, crypto, commodities and ETFs for 2018–2024 and identified Bitcoin and Ethereum as net shock transmitters while gold, silver and crude oil behaved as relative safe‑haven receivers in stress periods. (link.springer.com) PyTorch Geometric (PyG) is the de‑facto practitioner stack for these projects, with an active GitHub repository showing frequent commits and example scripts for temporal/heterogeneous GNNs and a PyG 2.0 paper describing native support for temporal and heterogeneous graph primitives aimed at scaling real‑world financial graphs. (github.com) Public codebases and tutorials used in recent work demonstrate an end‑to‑end pipeline: rolling correlation or spillover matrices → dynamic graph construction → PyG message‑passing models → node‑level volatility forecasts and spillover heatmaps (see baharkudarzi’s gnn‑cross‑asset repo and the GNNHAR repo for compute_vol and Summary_Results scripts). (github.com) A recent MDPI study comparing fixed versus dynamically re‑estimated spillover graphs found that strict expanding‑window out‑of‑sample tests showed no consistent advantage from constantly updating the graph over using a well‑specified fixed spillover network, highlighting a practical robustness trade‑off for implementation. (mdpi.com)
