Title: LIESEL: A software framework for prototyping Bayesian models and exploring estimation methods
Authors: Hannes Riebl - Goettingen University (Germany)
Paul Wiemann - TU Dortmund University (Germany) [presenting]
Thomas Kneib - University of Goettingen (Germany)
Abstract: LIESEL is a software framework for Python to facilitate statistical research on Bayesian models focusing on modularity, extensibility, and reliability. In the area of statistical software, LIESEL is located between specialized implementations of specific model classes and general-purpose software packages for Bayesian inference. The framework can be used for quick model prototyping or serves as a basis for the implementation of complex models or novel statistical inference algorithms. High-performance, albeit being implemented in Python, is achieved by the extensive use of the modern libraries JAX and TensorFlow Probability providing access to just-in-time compilation, automatic differentiation, and vectorization. Furthermore, LIESEL runs on high-performance computing devices like GPUs or TPUs. The framework provides an easily extensible library for MCMC estimation, including the HMC and NUTS sampling algorithms. Because of the strong modularity, LIESEL's estimation and modeling modules can be used independently of each other. LIESEL comes with tools to set up structured additive distributional regression models. Distributional regression enables researchers to explore complex relationships between explanatory and response variables beyond the mean. We present LIESEL's architecture and demonstrate its applicability in several case studies featuring distributional regression, including copula regression.