Combining Multi-Aspect Policyholder Review Analytics with Edge-Hosted CNN–LSTM Time-Series Forecasting for Dynamic Property Insurance Risk Assessment

• Billy Elly

  Lautech

  Author

Property insurance carriers face mounting pressure from escalating catastrophe losses, rising claims frequencies, and tightening underwriting margins, yet conventional risk assessment methods remain anchored to static historical claims data and coarse geospatial proxies that fail to capture dynamic property-level risk evolution. This study addresses the research gap at the intersection of policyholder behavioral analytics and real-time property risk forecasting by proposing a hybrid framework that integrates multi-aspect sentiment analysis of policyholder reviews with an edge-hosted CNN–LSTM neural network for time-series risk prediction. The methodology combines aspect-based sentiment classification using DeBERTa-v3 encoders to extract granular policyholder satisfaction signals from customer feedback, with a collaborative CNN–LSTM architecture deployed across edge-cloud infrastructure to capture spatiotemporal risk patterns in near real-time. Empirical evaluation using a synthetic property insurance dataset demonstrates that the proposed framework achieves 89.4% predictive accuracy for dynamic risk classification, outperforming standalone CNN (85.2%) and LSTM (83.7%) models, while reducing inference latency by approximately 65% through edge-based deployment. The framework provides insurers with a replicable, computationally efficient tool for continuous risk reassessment, enabling proactive intervention before claims materialize. Key implications include enhanced underwriting precision, improved policyholder retention through sentiment-driven service adjustments, and a scalable architecture suitable for integration into existing carrier workflows.

• pdf

Copyright (c) 2026 Billy Elly (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.