FINESSE: An Agent-Based Simulator and Benchmark Dataset for Multimodal Financial Event Sequences

1 Introduction

• Synthetic datasets can support research on financial-services modeling, while distribution-matching techniques raise privacy concerns and limit applicability.The paper presents agent-based models as an alternative for generating synthetic datasets independent of sensitive data.Source: §1 · p. 1
• Existing financial datasets are often monomodal, designed for narrow use cases, or limited to static account-level attributes, missing evolving, interdependent event sequences.FINESSE is introduced to capture dynamic relationships across transactions, payments, account status changes, and policy interventions.Source: §1 · p. 1
• FINESSE-ABM generates long-term, interdependent event sequences with latent state dynamics, while FINESSE-Bench supports four representative financial modeling tasks.The tasks are spending behavior forecasting, missed payment prediction, transaction fraud detection, and next event prediction.Source: §1 · p. 2
• The paper reports initial baselines for all four tasks using sequence models, temporal graphs, time-series forecasting, and temporal point processes.Source: §1 · p. 2

4 Event Sequences

• FINESSE’s event streams cover transactions, payments, account state, and policy changes.Source: §4 · p. 2

4 Methodologies

• Figure 1 is identified as a high-level summary of FINESSE-Bench.Source: Figure 1 · p. 2
• FINESSE aims to facilitate research on underexplored structured, multimodal event-sequence modeling challenges in financial services.Source: §4 · p. 2
• The benchmark tasks are for research and do not reflect granular real-world tasks associated with any organization.Although FINESSE generates synthetic datasets with latent state dynamics, it does not aim to support formal causal inference.Source: §4 · p. 2

2.1 Agent Based Models

• Prior agent-based finance work includes market dynamics, interbank liquidity risk, and credit-card incentive modeling, but credit-card applications have focused on account-level rather than granular event data.FINESSE instead targets long-term, interdependent event sequences spanning transactions, payments, account states, and policy adjustments.Source: §2.1 · pp. 2–3

2.2.1 Transaction Fraud Detection

• Open-source fraud datasets often focus on single-modal transaction data or fundamental features, including device-level and synthetic transaction features.These datasets include examples aggregated into Amazon’s fraud detection benchmark [20].Source: §2.2.1 · p. 3

2.2.2 Loan Default Prediction

• Loan-payment logs are underrepresented; cited alternatives provide account-level credit-bureau and demographic features or aggregated statement-level features.FINESSE offers event-level logs and latent state dynamics related to credit-card policy adjustments and risky merchants.Source: §2.2.2 · p. 3

2.2.3 Multimodal Benchmarks

• Existing multimodal datasets vary in their temporal detail and focus: Berka lacks granular temporal features and policy dynamics, while the Data Fusion Challenge centers on stream matching.MBD supports campaign analytics and modality matching; FINESSE targets structured event-sequence tasks, and the paper encourages cross-sequence fusion techniques.Source: §2.2.3 · p. 3, Table 1 · p. 3

3 FINESSE-ABM: Simulation Framework for Multimodal Event Sequences

• Agents transition through latent states over discrete time steps, influencing their actions and interactions across financial event streams.The simulation outputs transaction, payment, account, policy, and hidden-state event logs.Source: §3 · pp. 3–4

3.1 Simulation Setup: Merchants, Customer Behavior, and Bank Policies

• Merchant-affinity and payment-strategy policies generate diverse customer spending and payment patterns, while bank interventions adjust credit limits, interest rates, and minimum-payment factors.Customer agents receive one policy of each type; policy parameters include frequency and amount factors.Source: §3.1 · pp. 4–5, Table 2 · p. 4

3.2 Latent State Dynamics

• Customer policies can transition over time, and interventions may increase transition likelihood depending on account state; risky merchant interactions also shape evolving fraud risk.Repeated transactions at merchants with higher identity-theft risk increase a customer's hidden fraud risk, which contributes to later fraud probability.Source: §3.2 · p. 5

4 FINESSE-Bench Dataset Overview

• Each FINESSE event type has its own schema with distinct attributes.The table covers schemas across the dataset's event types.Source: Table 3 · p. 6
• Event types are recorded at different frequencies.Source: Table 3 · p. 6
• The schema overview connects event-type attributes with their differing recording frequencies.Source: Table 3 · p. 6

5 Task Evaluation

• Evaluations used a single A10 GPU, with task training times of 12–24 hours.Source: §5 · p. 6

5.1 Detecting Fraudulent Transactions

• Fraud detection is framed as binary classification of transaction-level fraud associated with latent identity-theft risk and risky merchants.Source: §5.1 · p. 6

5.1.1 Edge Classification on Temporal Graph

• The fraud task treats customer–merchant transactions as timestamped edges in a continuous-time dynamic bipartite graph, classified using TGN [30] and DyRep.The timestamp-based split assigns the first 70% of transactions to training, followed by validation and test sets; evaluation reports ROC-AUC and PR-AUC.Source: §5.1.1 · p. 6

5.2 Missed Payment Prediction

• Missed-payment prediction uses a customer's historical sequence to predict the number of missed payments over a future horizon, with a missed payment defined by failure to pay the minimum by the due date.Source: §5.2 · pp. 6–7

5.2.1 Baselines

• Missed-payment baselines include MLP, GRU, Transformer [33], MAMBA [12], and CoLES, trained with default model settings.Source: §5.2.1 · p. 7

5.2.2 Metrics

• Models predict missed-payment counts over 3-, 6-, 9-, or 12-month horizons using multiclass ROC-AUC for evaluation.Source: §5.2.2 · p. 7

5.2.3 Results

• Average multiclass ROC-AUC declines across almost all baselines as the missed-payment horizon increases; Transformers lead at 3 and 6 months, while GRU leads at 9 and is slightly ahead at 12 months.At 12 months, baseline performance is nearly identical.Source: §5.2.3 · p. 7, Figure 3 · p. 7

5.3.1 Experiment Setup

• The baselines forecast 60 days of future account balances from 60-day sequences of past balances.Source: §5.3.1 · p. 7

5.3.2 Metrics & Results

• MambaSimple ranks lowest, as the 60-day sequence length does not benefit from its O(L) scaling advantages.Source: §5.3.2 · p. 8
• PatchTST leads three of five balance-forecasting metrics, while the Transformer performs best on MAPE and MPSE.Source: §5.3.2 · p. 7, Figure 4 · p. 8

5.4 Next Event Prediction

• Next-event prediction is framed as a discrete temporal point process, with evaluations for next payment and transaction-category prediction using four named model families.The methods are RMTPP, NHP, THP, and IFTPP; Table 5 is titled as reporting next-event type and time prediction baseline results.Source: §5.4 · p. 8, Table 5 · p. 8

• FINESSE generates synthetic multimodal event sequences with latent state dynamics, while FINESSE-Bench evaluates four representative financial tasks.The tasks include transaction fraud detection, missed payment prediction, balance forecasting, and next event prediction.Source: §6 · p. 8
• The flexible framework and benchmark are intended to enable research on structured, multimodal event-sequence modeling.The reported baselines serve as an initial reference point, and the authors encourage extensions to both the simulator and benchmark.Source: §6 · p. 8