Tabular foundational models represent a relatively recent learning approach that enables the use of previously acquired knowledge when solving new problems. In my thesis, I examined their applicability and performance on classification problems in the financial domain. In the experimental part, I used the foundational models TabPFN and TabICL and compared them with the classical model XGBoost on several datasets describing problems such as financial fraud detection, credit risk assessment, credit rating, and loan evaluation. I compared the models based on classification performance, result stability, and computational complexity, and additionally analyzed their predictions using the SHAP method.
The results showed that no single model is best for solving financial problems. However, I found that TabICL and TabPFN are very strong competitive alternatives to classical methods for this type of task, with their performance remaining dependent on the characteristics of each specific problem and dataset.
|
