Agent-based economic models represent households, firms, banks and other actors individually and simulate how their decisions and interactions create wider economic outcomes. As the number of agents, stored information, interactions and repeated simulation runs grows, execution becomes increasingly demanding. The challenge is to determine which model operations can run at the same time without changing the intended model logic and which computing resources are suitable for them.

The thesis combines a literature-based analysis with a practical experiment in Julia. It first examines which workload characteristics affect parallel execution and develops a three-layer framework linking the economic model, computational tasks and execution resources. A simplified wealth-exchange model is then implemented in sequential and multithreaded CPU versions. Tests compare the runtime of the complete simulation with the parallelised transaction phase as the number of threads and agents changes. Correctness checks verify the simulation results.

The study shows that agent-based models should not be treated as one uniform workload. Their processes should first be divided into meaningful tasks and assessed for independence, dependencies, task size, memory access and synchronisation. In the Julia experiment, parallelising non-overlapping transactions improved performance: the transaction phase reached a maximum speedup of about 2.02, while the complete simulation reached about 1.35 with eight threads. Performance did not improve further at sixteen threads, and efficiency declined when the workload grew with the thread count. This demonstrates that adding threads alone does not guarantee proportional speedup because sequential preparation, memory access and coordination remain limiting factors. For the Institute für Wirtschaftsinformatik FHNW, the thesis provides a structured three-layer framework for analysing future agent-based economic models and selecting suitable CPU, GPU, hybrid or task-based execution approaches. It also identifies concrete directions for further development, including improved memory locality, reduced sequential preparation, parallel scenario execution and explicit dependency-aware scheduling.