Learning from failures: A hybrid approach to warranty claims forecasting

Wei XIE , Zelong LIN , Qiyang LI , Yun GU , Haitao LIAO

Eng. Manag ››

PDF (2354KB)
Eng. Manag ›› DOI: 10.1007/s42524-026-6173-4
Research Article
Learning from failures: A hybrid approach to warranty claims forecasting
Author information +
History +
PDF (2354KB)

Abstract

Accurate short-term forecasts of warranty claims are essential for reserve accrual, spare-parts provisioning, service-capacity planning, and quality improvement. However, forecasting is difficult under product heterogeneity, right censoring, and sparse observations because monthly claim counts aggregate historical sales cohorts whose failure risks evolve with service age. Parametric lifetime models are sensitive to distributional misspecification, whereas purely data-driven models can be unstable when data are limited. We develop tree-gradient warranty claims forecasting (TG-WCF), a hybrid framework that uses claim-generation structure to guide the correction of data-driven forecasts. Its distinctive feature is to combine disagreement between a Weibull-based sales–failure prior and a light gradient boosting machine (LGB) baseline with learned empirical residuals in a bounded logarithmic adjustment. Monthly feedback and sample-specific gating then adapt the correction to temporal drift and differences across observations, making the integration of structural knowledge and statistical learning responsive to changing claim patterns. We evaluate TG-WCF using real-world warranty records for 21 product models in the North American electronics market. TG-WCF outperforms benchmark methods across most evaluation dimensions. Compared to the second-best performing method for each metric, TG-WCF reduces weighted absolute percentage error by 19.56%, peak-weighted absolute percentage error by 37.87% and dynamic time warping distance by 26.72%, while maintaining near-zero aggregate bias. These findings support a domain-informed design logic for operational forecasting under data scarcity: predictions should be anchored in claim-generation structure, systematic misspecification should be learned from data, and adaptation to temporal change should be enabled. For managers, the resulting monthly forecasts support coordinated decision-making regarding warranty reserves, spare parts inventories and service capacity.

Keywords

warranty claims forecasting / claim-generation mechanism / gradient boosting algorithm / dynamic calibration / physics-informed machine learning

Cite this article

Download citation ▾
Wei XIE, Zelong LIN, Qiyang LI, Yun GU, Haitao LIAO. Learning from failures: A hybrid approach to warranty claims forecasting. Eng. Manag DOI:10.1007/s42524-026-6173-4

登录浏览全文

4963

注册一个新账户 忘记密码

References

Rights & permissions

Higher Education Press 2026

PDF (2354KB)

0

Accesses

0

Citation

Detail

Sections
Recommended

/

〈 〉