What KAN mortality say: smooth and interpretable mortality modeling using Kolmogorov−Arnold networks

Abstract In machine learning-based mortality models, interpretation methods are well established, and they can reveal structures resembling the age or time effects in traditional mortality models. However, in the reverse direction, using such traditional components to guide the initialization of a neural network remains highly challenging due to information loss during model interpretation. This study addresses this gap by exploring how components from pre-fitted traditional mortality models can be used to initialize neural networks, enabling structural information to be incorporated into a deep learning framework. We introduce Kolmogorov–Arnold Networks (KAN) and first construct two shallow models, KAN[2,1] and ARIMAKAN, to examine their applicability to mortality modeling. We then extend the Combined Actuarial Neural Network (CANN) into a KAN-based Actuarial Neural Network (KANN), in which classical model components calibrated via generalized nonlinear models or generalized additive models are naturally used for initialization. Three KANN variants, namely KANN[2,1], KANNLC, and KANNAPC, are proposed. In these models, neural networks assist in improving the accuracy of traditional models and help refine the original parameter estimates. All KANN-based models can also produce smooth mortality curves as well as smooth age, period, and cohort effects through simple regularization. Experiments on 34 populations demonstrate that KAN-based approaches achieve stable performance while balancing interpretability, smoothness, and predictive accuracy.