Centro de Documentacion de Fundación MAPFRE - What KAN mortality say: smooth and interpretable mortality modeling using Kolmogorov-Arnold networks

MAP20260002064

What KAN mortality say: smooth and interpretable mortality modeling using Kolmogorov-Arnold networks / Lianzeng Zhang and Yuan Zhuang

Sumario: This study explores how traditional mortality model components can be directly used to initialize neural networks, addressing the common loss of structural information that occurs during model interpretation. We introduce KolmogorovArnold Networks (KAN) and develop both shallow models (KAN[2,1] and ARIMAKAN) and a KAN-based Actuarial Neural Network (KANN), which extends the existing CANN framework. The proposed KANN variantsKANN[2,1], KANNLC, and KANNAPCincorporate elements from classical mortality models and enhance them through deep learning. These models produce smooth mortality curves and smooth age, period, and cohort effects through simple regularization. Experiments across 34 populations show that KAN-based approaches offer stable performance, successfully balancing interpretability, smoothness, and predictive accuracy

En: Astin bulletin. - Belgium : ASTIN and AFIR Sections of the International Actuarial Association = ISSN 0515-0361. - 19/01/2026 Volume 56 Issue 1 - January 2026 , p. 32 - 59

1. Mortalidad . 2. Modelos matemáticos . 3. Cálculo actuarial . 4. Machine learning . 5. Longevidad . 6. Matemática del seguro . 7. Modelos estadísticos . I. Zhuang, Yuan . II. International Actuarial Association . III. Title.