Researchers release FINESSE, a simulator and benchmark for multimodal financial event sequences

Machine learning research aimed at financial services has been held back by a shortage of representative, open source datasets, the authors argue. Existing public resources tend to focus narrowly on a single modality or a single task and fail to capture the structured, multimodal, and constantly changing nature of real financial problems.

To close that gap, the authors introduce FINESSE, short for Financial Event Sequence Simulation Environment: an agent-based simulation framework that produces synthetic, structured datasets made up of several interdependent event streams. Each stream represents a distinct financial behavior (e.g., transactions, payments, account status changes, or policy interventions), and each has its own action space, schema, and variable types. The streams are linked through the simulated agents' underlying, evolving state, which couples them together and lets the simulator produce temporally rich interactions across different kinds of events rather than isolated, single-type data.

Alongside the simulator, the authors release FINESSE-Bench, a benchmark dataset generated with it. FINESSE-Bench supports four tasks: balance forecasting, transaction fraud detection, missed payment prediction, and next event prediction. The authors report baseline results using established method families: time series forecasting, event sequence modeling, temporal graphs, and temporal point processes. They release the full FINESSE framework, simulator and dataset together, with the stated goal of accelerating research on structured, multimodal event sequence modeling in financial services.

Key facts

FINESSE (Financial Event Sequence Simulation Environment) is an agent-based simulator that produces synthetic, structured, multimodal datasets built from multiple interdependent financial event streams.
Each event stream models one type of financial behavior (e.g., transactions, payments, account status changes, or policy interventions), each with its own action space, schema, and variable types.
The event streams share the simulated agents' latent, evolving state, which couples the streams together and produces temporally rich interactions across event types.
FINESSE-Bench, the benchmark dataset generated by the simulator, supports four tasks: balance forecasting, transaction fraud detection, missed payment prediction, and next event prediction.
Baseline results are reported using time series forecasting, event sequence modeling, temporal graphs, and temporal point processes, and the authors release the full framework, simulator and dataset included.

Why it matters

Machine learning research in financial services is limited by a shortage of representative, open source datasets, according to the authors: existing resources tend to focus on a single modality or task and do not capture how structured, multimodal, and constantly changing real financial systems actually behave. FINESSE addresses that gap by simulating several linked event streams (e.g., transactions, payments, account status changes, and policy interventions) that share the same underlying agents rather than existing as separate, unrelated tables. Because the streams are coupled through each agent's evolving state, the simulator can reproduce temporally rich interactions across event types, something single-modality datasets cannot show.

Who it affects

This work is aimed at researchers who build models for banks, fintech firms, and payment processors, and who currently have little choice but to train and evaluate on scarce, often proprietary financial records. FINESSE-Bench gives them a shared, synthetic testbed instead: its four tasks (balance forecasting, transaction fraud detection, missed payment prediction, and next event prediction) let different groups compare methods on common ground rather than on incompatible private datasets.

How to use it

The authors say they are releasing the full FINESSE framework, the simulator together with the FINESSE-Bench dataset. That makes two things available: a generator for new synthetic event-stream data, and a benchmark with four ready-made tasks (balance forecasting, transaction fraud detection, missed payment prediction, and next event prediction), together with baseline results reported using time series forecasting, event sequence modeling, temporal graphs, and temporal point processes. The abstract does not say where or how the framework and dataset are released: no repository link, license, or venue is named.

How solid is it

The authors anchor the new benchmark with baseline results computed using established method families (time series forecasting, event sequence modeling, temporal graphs, and temporal point processes) rather than a single new best-performing model. However, the abstract gives no baseline performance numbers, such as accuracy, F1, or error rate, for any of the four tasks; it states only that these method families were used to compute them. It also does not state the scale of FINESSE-Bench: the number of agents, event streams, or records the dataset contains is not given. Because the dataset is simulated rather than drawn from real financial records, how closely its dynamics track real-world behavior is a separate question the abstract does not address.

Risks and caveats

The dataset's scale (how many agents, event streams, or records FINESSE-Bench contains) is not stated, and neither are baseline performance numbers for its four tasks. As with any arXiv posting, it is also not stated whether the work has been peer reviewed.