Graph Neural Networks for Financial Networks & Systemic Risk

Almost every machine-learning recipe in finance begins by committing the same quiet act of vandalism: it takes a world that is densely interconnected — banks lending to banks, suppliers shipping to manufacturers, assets that crash together, wallets that pass coins between them — and flattens it into a table. One row per entity, a tidy column for each feature, and the relationships? Severed. Thrown out with the metadata. Yet the most important questions in finance live precisely in those discarded relationships: If this bank fails, who else falls? Which wallets, individually unremarkable, together form a fraud ring? When a key supplier stumbles, which firms three links downstream are about to miss earnings? A flat table, by construction, cannot see any of it.

This course hands you the toolkit that refuses to throw the edges away. A graph neural network (GNN) learns directly on the network: nodes (institutions, firms, assets, wallets), edges (loans, supply links, correlations, transfers), and a single deceptively powerful idea called message passing — each node repeatedly updates its own representation by aggregating information from its neighbours. From that one mechanism falls everything: scoring an institution by the company it keeps, predicting a hidden counterparty link, classifying a whole portfolio’s fragility, and tracing a shock as it cascades hops away through the lending graph. We build on Deep Learning for Market Data: you already respect tiny effective sample sizes, purged evaluation, and the deflated Sharpe — and you will need every bit of that paranoia, because graphs leak harder than any data structure you have met. The arc:

• Markets are graphs, not tables — the representation shift: nodes and edges, directed vs undirected, weighted vs binary, the node/edge/graph feature stack, and the four canonical finance graphs (interbank lending, supply chains, asset-correlation networks, on-chain wallet/transaction graphs).
• Message passing and the three tasks — aggregate-from-neighbours as the engine of every GNN, why permutation-invariant pooling is non-negotiable, how stacking layers grows the receptive field hop by hop, and the three task types: node-level (classify/score an entity), edge-level (predict a link), graph-level (score a whole network state).
• GCN, GraphSAGE, GAT and temporal GNNs — the architecture lineage: spectral-flavoured GCN (fixed degree-normalised averaging), inductive GraphSAGE (sample-and-aggregate that generalises to unseen nodes), GAT (learned attention over neighbours), and temporal/dynamic GNNs for a network whose topology changes every day.
• Systemic risk and contagion — the killer application: the default cascade through an interbank network, DebtRank and the feedback loops a per-firm tabular model structurally misses, and how a GNN learns a fragility score that embeds the topology.
• Fraud rings and relational alpha — AML ring detection on transaction graphs (the signal is the structure, not any one account), relational/peer-firm alpha propagating a signal across supply-chain and co-movement edges, and on-chain analytics on wallet graphs.
• The honest audit — why graphs leak so badly (a node sees its neighbours, so a careless split lets the test set peek across edges), topology-aware splitting, over-smoothing as you stack too many layers, non-stationary topology, and the brutal data-plumbing cost that sinks most GNN projects before they predict anything.

By the end you will be able to model a financial system as the network it actually is, choose the architecture whose inductive bias fits the problem, and — most valuably — distinguish a genuine relational edge from a leakage artifact that a topology-blind split manufactured for you. A graded final exam runs the whole discipline back at you, one locked question at a time.