An explainable hybrid adaptive neuro-fuzzy inference system and deep learning framework for stochastic claims reserving
Accurately predicting insurance claims is crucial for insurance companies as it directly impacts cash flow, pricing strategies and overall profitability. This paper addresses the critical need for distributional forecasting, particularly for regulatory requirements that necessitate understanding the...
Enregistré dans:
| Auteurs principaux: | , , |
|---|---|
| Format: | Article (Journal) |
| Langue: | anglais |
| Publié: |
15 December 2026
|
| In: |
Expert systems with applications
Year: 2026, Volume: 331, Pages: 1-24 |
| ISSN: | 1873-6793 |
| DOI: | 10.1016/j.eswa.2026.132994 |
| Accès en ligne: | Verlag, kostenfrei, Volltext: https://doi.org/10.1016/j.eswa.2026.132994 Verlag, kostenfrei, Volltext: https://www.sciencedirect.com/science/article/pii/S0957417426019056 
|
| Notes sur l'auteur: | Arne Johannssen, Ali Yeganeh, Nataliya Chukhrova |
| Résumé: | Accurately predicting insurance claims is crucial for insurance companies as it directly impacts cash flow, pricing strategies and overall profitability. This paper addresses the critical need for distributional forecasting, particularly for regulatory requirements that necessitate understanding the variability of the claims. To this end, a hybrid model is proposed that integrates the chain ladder (CL) method, machine learning (ML) and fuzzy logic for enhanced claims reserve prediction through three major modules. In the first module, a multi-layer perceptron (MLP) is employed to capture patterns of the losses’ location, and then, CL estimations are imported to a second module using the adaptive neuro-fuzzy inference system (ANFIS). The outputs of the MLP and ANFIS are integrated into a recurrent neural network (RNN), as a kind of ensemble learning approach, in the third module. Specifically, a long short-term memory (LSTM) model is employed to obtain the mean and standard deviation of the loss distribution for each cell in the claims development triangle. Moreover, a module-wise explainability strategy is adopted that combines intrinsically interpretable components with post-hoc analyses for neural networks. Through comprehensive simulations, parameter optimization is addressed, the importance of its key components is validated, and the competitive performance of the method in accurately predicting the outstanding loss liabilities is demonstrated. |
|---|---|
| Description: | Gesehen am 04.08.2026 Online verfügbar: 4. Juni 2026, Artikelversion: 26. Juni 2026 |
| Description matérielle: | Online Resource |
| ISSN: | 1873-6793 |
| DOI: | 10.1016/j.eswa.2026.132994 |