A multi-view contrastive learning framework for spatial embeddings in risk modelling

Battle Count: 60

The paper focuses on insurance risk modelling and real estate pricing rather than financial markets or trading. While spatial embeddings could theoretically be applied to geographic risk factors in commodity trading or catastrophe bonds, the paper does not address trading strategies, market microstructure, or portfolio optimization. The relevance is primarily to insurance/actuarial risk management rather than quantitative trading.

The framework involves multiple components: ResNet18 image encoder, hexagonal CNN for OSM data, multi-view fusion network, spherical harmonic positional encoding, SIREN network, and contrastive training. However, the authors demonstrate training on consumer-grade hardware (RTX 4070, ~12 hours), provide complete code on GitHub, and pretrained models on Hugging Face. The hexagonal-to-square transformation and contrastive loss add moderate complexity. Inference is fast (<1ms per coordinate).

GitHub repository with complete code (https://github.com/freekholvoet/MultiviewSpatialEmbeddings), pretrained models published on Hugging Face (Holvoet 2025), public datasets used (French DVF, Belgian flood claims via Assuralia), detailed hyperparameter settings provided, training on consumer-grade hardware (NVIDIA RTX 4070, 12GB VRAM), complete dependency list with version numbers provided.