Abstract
Automated Machine Learning (AutoML) frameworks offer promise for predictive analytics tasks in finance, but limited guidance exists on their domain-specific suitability. This study benchmarks nine open-source AutoML frameworks and the transformer-based model TabPFN across 25 financial predictive analytics datasets, leveraging the AMLB Benchmark. Performance was evaluated across data segments. Results reveal performance variation depending on task characteristics, with no single framework consistently outperforming others. While TabPFN excelled in multiclass tasks, traditional frameworks like Hyperopt-Sklearn and MLJAR offered greater stability in binary settings. Notably, financial industry-optimized solutions like LAMA did not consistently outperform general-purpose frameworks. This study extends and challenges prior AutoML benchmarking results and validates AMLB for financial tasks. Furthermore, it provides empirical comparisons between transformer-based models like TabPFN and traditional AutoML frameworks. This research informs the optimization of AutoML frameworks and outlines directions for future research, including broader predictive tasks and preprocessing strategies.
Recommended Citation
Leyh, Nicolas, "Can AutoML Handle the Constraints of Finance? A
Domain-Specific Benchmark of Automated ML Frameworks and
TabPFN" (2025). ACIS 2025 Proceedings. 28.
https://aisel.aisnet.org/acis2025/28