Sebastian Benthall - New Developments in Open Source Computational Economics | SciPy 2024

Computational economics is undergoing major changes with the adoption of deep learning, GPU computing, and more sophisticated algorithms for solving complex economic models

Traditional representative agent modeling is being replaced by heterogeneous agent models that can better capture economic diversity and market dynamics

Key computational challenges include:

• Curse of dimensionality in state spaces
• Need to discretize continuous values
• Expensive grid-based computations
• Complex numerical integrations

New approaches using neural networks and sampling techniques are making previously intractable high-dimensional problems solvable:

• Sampling from ergodic distributions rather than full grid computation
• Using loss functions and gradient-based optimization
• Leveraging GPU acceleration

Open source libraries like HARK and DOLO are enabling:

• More modular and reusable model components
• Human and machine readable model specifications
• Cross-platform compatibility between Python and Julia
• Improved reproducibility

Economic models are becoming more realistic by incorporating:

• Aging and mortality
• Heterogeneous agents with different preferences
• Market mechanisms and aggregation
• Banking sectors and financial markets
• Time-inconsistent preferences

Structural estimation techniques using tax records and other data help validate and calibrate these increasingly complex models

There is growing recognition that pure rationality assumptions need to be relaxed in favor of more bounded rational models that better match observed behavior

The field is moving toward more standardized, reproducible approaches through domain-specific languages and improved software development practices

Computational advances are enabling more complex and realistic models while reducing implementation time and computational costs