Revealing Geography-Driven Signals in Zone-Level Claim Frequency Models: An Empirical Study using Environmental and Visual Predictors

Battle Count: 77

The paper is primarily focused on actuarial science and insurance risk modeling rather than quantitative trading. However, there are tangential connections: (1) the use of alternative data sources (geospatial, imagery) parallels the use of alternative data in quantitative finance; (2) the Poisson modeling framework and regularization techniques are transferable to count-based financial models; (3) the multimodal feature engineering approach could inspire similar techniques in trading signal generation; (4) the robustness analysis methodology is relevant to any predictive modeling context. The geographic risk assessment methodology could potentially inform insurance-linked securities pricing or catastrophe bond modeling.

The implementation requires multiple components: (1) GIS data processing with ArcGIS Pro for extracting environmental features from OSM and CORINE at multiple radii; (2) Image tile extraction from NGI tile service; (3) Training of ResNet18 CNN adapted for grayscale input; (4) Vision Transformer embedding extraction (Nomic-v1.5); (5) Multiple model families (GLM, ElasticNet GLM, XGBoost, MLP, CNN); (6) Complex multi-fold cross-validation scheme with 6 outer folds and 5 inner folds; (7) Feature engineering at multiple spatial scales. The statistical modeling itself is standard, but the data pipeline construction and multimodal feature engineering add significant complexity.

The beMTPL97 dataset is publicly available in the CASdatasets R package. OpenStreetMap and CORINE Land Cover 2000 data are publicly accessible. NGI Belgium orthoimagery is available for academic use via Cartesius.be tile service. The methodology is well-documented with specific hyperparameters (ElasticNet η=0.2, α=1.0 for fold 3). However, the paper does not provide a GitHub repository link, and the exact preprocessing pipeline in ArcGIS Pro is described but not fully reproducible without the software. The multi-fold evaluation scheme follows Henckaerts et al. (2021) methodology.