Research on multi-agent systems has extensively modeled real-world phenomena across various domains including epidemiology, urban planning, and financial transactions. These systems often struggle to produce agent behaviors that comprehensively capture the dynamics of the real-world ecosystem and the unique behaviors of each agent type. Furthermore, the limited explainability of these models due to non-iterative calibration poses significant challenges. This paper introduces an iterative model calibration algorithm that dynamically adjusts the multitude of parameters in a multi-agent simulation platform. Initially treating the simulation model as a black box, our method refines simulation parameters through cycles of adjustments based on clusters of observed behaviors comprising the behavior of both agents and actors. This approach allows for the identification and correction of inaccuracies, introduces new parameters, and discards erroneous ones within the agent-based model as demonstrated in a Mobile Money Transaction Simulator (MoMTSim). The calibration algorithm enhances the realism and applicability of the simulation model by ensuring that the generated synthetic datasets closely mirror real transaction data. The effectiveness of this calibration method was determined by validating the generated data.We computed the delta between the real and synthetic data using the sum of squared errors approach. This shows that MoMTSim can produce synthetic data that resemble real mobile money transaction data, thereby confirming the efficacy of the calibration algorithm for use in simulating complex financial ecosystems.
This data article introduces a labeled synthetic mobile money transaction dataset created using MoMTSim, a multi-agent-based simulation platform designed and validated specifically for mobile money transactions. MoMTSim toolkit simulates mobile money interactions, ensuring that the generated synthetic dataset closely mimics the statistical properties of real transaction data. This dataset encapsulates a wide range of transaction features, such as timestamps (step), transaction amounts, the initial and new account balances of both the initiator and recipient, participant IDs, and the types of transactions conducted. The included transaction types span deposits, withdrawals, transfers, payments, and debits. Each record in the dataset also carries a label that identifies whether the transaction is legitimate or fraudulent. The synthesis of this dataset using MoMTSim is described in this article and its structure and summary statistics are also presented. The dataset is particularly suitable for training and testing machine learning algorithms to detect financial fraud. Additionally, it holds the potential for benchmarking fraud detection algorithms and systems and validating synthetic data generation methodologies.