Leveraging Time-Series Foundation Models for Financial Crime Detection

1 Introduction

• Financial crime detection must identify a few illicit transactions among millions of legitimate ones, and illicit behaviour is rare and evolves as criminals adapt.Detection is an edge-classification task for AML, where suspiciousness depends on transaction context, while phishing detection labels accounts.Source: §1 · p. 1
• Account transaction rhythms can distinguish launderers from legitimate customers, but those account-level patterns must inform decisions about individual transactions.The challenge is to enrich each transaction with temporal context drawn from its structural neighbourhood.Source: §1 · pp. 1–2
• Prior approaches either reduce time to a timestamp or use narrow temporal summaries, while pooled TSFM embeddings give every transaction of an account the same vector.Node-level fusion performs well on node-level tasks but trails the strongest structural models on AML edge classification.Source: §1 · p. 2
• TCEE lets each transaction query frozen embeddings of its source and destination accounts using a query formed from that transaction’s own features.The resulting edge representations can be scored by an MLP or supplied to a GNN; two transactions from one account can retrieve different contexts.Source: §1 · p. 2
• 10.8% relative average PR-AUC improvement over each dataset’s strongest evaluated baseline on five of six AMLWorld datasets, with 6.0× faster end-to-end training on average across all six.Adding a GNN backbone yields a further 4.8% PR-AUC on the four datasets trained to convergence; on ETH phishing, TCEE improves PR-AUC by 17.8% over the strongest baseline.Source: §1 · p. 2

• Handengineered AML features encode only patterns specified in advance, whereas GNNs learn features directly from the graph.Source: §2 · p. 2
• Structural GNNs strengthen graph-pattern recognition, but generally treat transaction time as a plain edge attribute.Source: §2 · p. 2
• Temporal graph models use recency weighting, node memories, or snapshots, but learn temporal representations from scratch and summarize histories per account.TCEE instead uses a frozen pretrained model and delivers temporal context to individual transactions.Source: §2 · pp. 2–3
• Chronos-2 encodes each account’s feature series jointly, using in-context attention and producing one 768-dimensional embedding per feature and patch of 16 transactions.Source: §2 · p. 3, Figure 2 · p. 3
• The encoder’s set of embeddings can be queried from graph edges, unlike prior pretrained-encoder approaches described here that fuse representations at the node level.The authors characterize TCEE as the first to query a frozen encoder from a graph edge and apply a TSFM to financial-crime detection.Source: §2 · p. 3

• The method section covers transaction-history encoding, node-level fusion, TCEE, and cost analysis.Source: §3 · p. 3
• All attention in the method is single-head scaled dot-product attention.Source: §3 · p. 3
• The stated attention convention applies across the method’s encoding, fusion, and TCEE components.Source: §3 · p. 3

• Each account’s sequence contains all sent and received transactions ordered by time, with six parallel feature series for AMLWorld.Source: §3.1 · p. 3
• The time-delta series exposes spacing between transactions because Chronos-2 treats its input as equally spaced.Source: §3.1 · p. 3
• A counterparty index numbers partners by first appearance and reuses the number for returning partners, exposing the account’s one-hop neighbourhood.Source: §3.1 · p. 3
• Averaging each account’s Chronos-2 embeddings over patches leaves one embedding per feature.Source: §3.1 · p. 3

• Node-level fusion pools each account’s feature embeddings into one vector, using attention by default.Source: §3.2 · p. 3
• Attention pooling lets each account weight its most informative feature series.Source: §3.2 · p. 3
• Chronos-MLP classifies an edge from its two accounts, while Chronos-GNN uses the account vector as the GNN’s initial node feature.Source: §3.2 · p. 3

3.3 TCEE: Temporal-Context Edge Enrichment

• TCEE lets each transaction attend over frozen embeddings for its two accounts using a query built from the transaction’s own features.The two account-perspective queries retrieve account-specific contexts; the temporal signal enters at the edge.Source: §3.3 · p. 3, Figure 3 · p. 4, §3.3.1 · p. 4

3.3.1 The Edge-Conditioned Query.

• A shared MLP maps sender- and receiver-perspective transaction views into queries that can weight the same account’s feature embeddings differently for each transfer.Each account’s keys and values are fixed, while its query changes with every transaction.Source: §3.3.1 · p. 4

3.3.2 Local Context.

• Pair pooling isolates transactions between the two accounts, while neighbourhood attention reads a 2w-transaction window around the target in each account’s history.The neighbourhood window includes transactions before and from the target time; relative time and a before/after flag encode position.Source: §3.3.2 · p. 4

3.3.3 The Enriched Edge.

• The enriched edge fuses the two queries, retrieved contexts, and optional extensions, then can be scored directly or supplied to a GNN backbone.A residual path preserves the transaction’s own encoded features.Source: §3.3.3 · pp. 4–5

3.3.5 Node Classification.

• For node classification, the GNN pools incident enriched edges into a node state, which a per-node head uses to classify the account.Training uses the loss over accounts rather than edges.Source: §3.3.5 · p. 5

3.4.1 Computational Complexity.

• Edge-minibatch Chronos-MLP and TCEE-MLP training is linear in |E|, and their shared account-embedding projection dominates cost, keeping their epoch costs close.TCEE-MLP adds attention and optional pooling computations; the mean number of parallel transactions per pair and the neighbourhood window enter its cost.Source: §3.4.1 · p. 5
• For sampled-subgraph PNA models, the backbone’s O(L(|E_s|+|V_s|)d^2) cost dominates the per-node embedding-projection saving.The passage notes that the other listed PNA-based baselines pay the same backbone cost.Source: §3.4.1 · p. 5

3.4.2 Storage.

• Averaging account-history patches reduces storage to 9 KB per account when F=6, with the store kept GPU-resident when it fits and otherwise split between resident training data and memory-mapped data.Without averaging, patch count grows with transaction count; Table 1 reports full graph storage cost per split.Source: §3.4.2 · p. 5
• Some runs were not trained to convergence, and Pair-PNA was capped at the 64 most recent transactions per pair.The passage says these choices understate the true result.Source: §3.4.2 · p. 6

4 Experiments

• The experiments ask about TCEE’s AML effectiveness and training cost, the contribution of a frozen TSFM, and transfer to phishing detection.AML evaluation uses edge classification; the ETH phishing task is node classification.Source: §4 · p. 6, §4.1.1 · p. 6

4.1.1 Datasets.

• AMLWorld covers higher- and lower-illicit-ratio regimes at small, medium, and large sizes, while the ETH phishing graph supplies a node-classification task.AMLWorld is an agent-based transaction simulator with complete labels.Source: §4.1.1 · p. 6

4.1.2 Baselines.

• Baselines include GINe and PNAe structural floors and prior MEGA-GNN, Multi-FraudGT, and Multi-GNN models, with the best test-PR-AUC aggregator variant reported per dataset.The ETH comparison also includes DIAM.Source: §4.1.2 · p. 6

4.1.3 Evaluation.

• PR-AUC is the primary metric under heavy class imbalance, with minority-class F1 also reported at a fixed 0.5 threshold.Test metrics are reported at the checkpoint with the best validation score and averaged over five seeds; time-to-best measures cost.Source: §4.1.3 · p. 6

4.2.1 Effectiveness (RQ1).

• 10.8% average relative PR-AUC margin over the strongest dataset baseline on five of six AML datasets; the full TCEE-MLP trains 6.0× faster end-to-end.Adding PNA raises PR-AUC a further 4.8% on the four datasets trained to convergence.Source: §4.2.1 · pp. 6–7, §4.2.2 · p. 7

4.2.2 Efficiency (RQ2).

• 6.0× faster end-to-end than the strongest baseline on average, the full TCEE-MLP reaches its best score with lower cost.TCEE-MLP variants hold every dataset’s Pareto front; adding a GNN backbone improves PR-AUC but trains 7.0× longer on the four converged datasets.Source: §4.2.2 · p. 7, Figure 4 · p. 7

4.2.3 Ablation Studies (RQ3).

• 14.2 and 7.8 PR-AUC points are lost when Chronos context is removed, the largest drop among the ablations.Counterparty information also helps, while transaction-query attention over feature embeddings adds 2.0 and 0.6 points; patch-axis pooling variants stay within two points of plain mean and win nowhere.Source: §4.2.3 · p. 7, Table 4 · p. 7

4.2.4 Phishing Detection (RQ4).

• 17.8% above the strongest baseline in PR-AUC, full TCEE-PNA reaches 72.1 PR-AUC on ETH phishing node classification.The two TCEE extensions add slightly worse minority-class F1 than plain TCEE-PNA; edge-level querying alone shows no measurable benefit over node-level fusion on this task.Source: §4.2.4 · p. 7, Table 5 · p. 7

Deployment.

• On large graphs, full TCEE-MLP finishes in about a day on one A40 GPU, compared with two to six days for structural baselines.The authors say it can support an analyst, not replace one, and whether this training cost suffices for banks’ retraining cadence remains open.Source: §5 · pp. 7–8

Generalisation.

• Whether results transfer to real bank data with incomplete labels remains to be explored.AMLWorld covers fixed typologies, mostly within two hops; TCEE-MLP sees one hop, while PNA sees two.Source: §5 · p. 8

Interpretability.

• Removing the eight highest-attention neighborhood transactions causes over half the PR-AUC drop from removing the whole window, versus about a fifth for eight random transactions.Attention indicates what the model consulted, not why it decided.Source: §5 · p. 8, Figure 5 · p. 8

Limitations and Future Work.

• The findings are limited to one GPU, one frozen encoder, and one domain.The authors suggest testing other TSFM encoders or fine-tuning one; larger GPUs may keep medium and large embeddings resident and make TCEE faster.Source: §5 · p. 8

6 Conclusion

• TCEE builds a separate representation for every transaction from its features, its accounts’ histories, and local context.The authors frame this as bringing context to the individual transaction where the decision is made.Source: §6 · p. 8
• 6.0× faster training accompanies TCEE surpassing the strongest evaluated baseline on five of six anti-money laundering datasets.Adding a GNN for broader structure yields further gains on the smaller datasets.Source: §6 · p. 8
• The same combination also outperforms all evaluated baselines on Ethereum phishing detection.The conclusion presents structural and temporal information as complementary for financial crime detection.Source: §6 · p. 8