This is the graded finale for Graph Neural Networks for Financial Networks, and it runs across the whole arc the course built. You started by refusing the small act of vandalism that opens most ML pipelines — flattening an interconnected world into a feature table and silently deleting the relationships where the expensive risks live. You learned to see markets as graphs instead: nodes carry the old columns, edges carry who-owes-whom, and the choice of directed-versus-undirected and weighted-versus-binary is never cosmetic because direction is risk direction. From there you built the engine — message passing, where each node aggregates from its neighbours with a permutation-invariant pool and then updates itself, growing its receptive field one hop per layer — and pointed it at the three jobs: node-level, edge-level, and graph-level. You walked the architecture lineage from fixed degree-normalised GCN through inductive GraphSAGE and attention-weighted GAT to temporal GNNs for a topology that will not hold still. Then the killer application: systemic risk, the default cascade round by round, DebtRank as a walk-once measure of impact, and fire-sale feedback that makes the system nonlinear. You aimed the toolbox at money — AML ring detection, relational peer-firm alpha, on-chain wallet analytics — and finally took the cold shower: graphs leak harder than any table, over-smoothing punishes depth, topology is non-stationary, and the data-plumbing tax decides the outcome while everyone argues about the architecture. No hints are shown, each answer locks the moment you submit, and your score stays hidden until the very end.
Course Recap
Big picture
Graph Neural Networks for Financial Networks — the whole arc
- GNNs for Financial Networks
- 1 · Markets are graphs
- Flattening to a table deletes the adjacency
- Graph = nodes plus edges; directed and weighted choices matter
- Four finance graphs: interbank, supply chain, correlation, on-chain
- 2 · Message passing and three tasks
- Aggregate from neighbours, then update self
- Permutation-invariant pool; K layers reach K hops
- Node-level, edge-level, graph-level jobs
- 3 · GCN, GraphSAGE, GAT, temporal
- GCN: fixed degree-normalised average, transductive
- GraphSAGE: sample, learn aggregator, inductive
- GAT: learned attention; temporal GNN for moving topology
- 4 · Systemic risk and contagion
- Cascade: fail when cumulative loss exceeds the buffer
- DebtRank: walk-once impact, not probability
- Fire-sale feedback multiplier one over one minus feedback
- 5 · Fraud rings and relational alpha
- AML: fan-in, chains, cycles, dense clusters
- Relational alpha: propagate a surprise along economic edges
- On-chain: public pseudonymous wallet graph
- 6 · The honest audit
- Leakage radius equals the layer count
- Topology-aware split; over-smoothing punishes depth
- Non-stationary topology; the data-plumbing tax
- 1 · Markets are graphs
One run, one shot
This is a graded, irreversible exam. There are 24 questions, shown one at a time. The instant you submit a question it locks — there is no Back button, no retry, and no Restart. A wrong answer simply fails that question and the exam moves on; you cannot revisit it. Your running score is hidden until the final screen. The pass mark is 70%. Some questions accept more than one correct option — read every option before you commit, because once you submit you own the answer.
A risk team loads a CSV with one row per bank — assets, leverage, capital ratio — and trains a model. Which fact about a bank's risk is structurally impossible for this representation to hold, no matter how many columns it adds?
Select an answer to continue.
Where this leaves you on the quant ladder
Pass or fail, you now carry the habit that separates a graph practitioner from someone running a tutorial on the wrong data: you refuse to flatten an interconnected market into a table, you can name the right finance graph and its edge semantics, you know that message passing is just aggregate-then-update growing one hop per layer, and you can pick the architecture — GCN, GraphSAGE, GAT, or a temporal GNN — whose inductive bias fits a growing, non-stationary network. You can trace a default cascade and a DebtRank walk by hand, size a fire-sale multiplier, and aim the whole toolbox at AML rings and relational alpha without mistaking suspicion for proof or a stale edge for signal. Above all you can audit a graph result — demanding a topology-aware temporal split, a shallow network that dodges over-smoothing, a regime-checked topology, and a metric deflated for trials — before you trust a single number it prints. That judgment, more than any architecture, is what lets a financial GNN earn its keep.