Visual Information and Geographic Reliability in Residential Valuation: A Multimodal, Multi-Source Benchmark from Shanghai — analysis code and aggregate results

Abstract

Companion analysis code and aggregate results for “Visual Information and Geographic Reliability in Residential Valuation: A Multimodal, Multi-Source Benchmark from Shanghai” by Cong Xu, Xinwei Gao and Jinglan Hu. Version 1.0.0. This deposit contains 22 analysis scripts, aggregate model/fold/protocol metrics, calibration and spatial residual diagnostics, measurement and external linkage summaries, extended predictor and temporal evaluation results, data dictionaries, provenance, and a Python standard-library replay script. The primary benchmark evaluates 17,716 Shanghai residential community records across 16 districts and 14,908 image footprints. The public replay checks aggregate consistency and package integrity; it does not retrain models or reconstruct individual predictions. Raw community and transaction records, exact locations, third-party RGB/GIS inputs, image patches and embeddings, row-level predictions and split assignments, and model weights are excluded because redistribution rights were not established. A distance-risk bin with n=1 is suppressed. See DATA_ACCESS.md for the source and access boundaries. This is a code and aggregate-results deposit, not a fully open raw-data replication package. Mixed-date predictors and unresolved source dates limit prospective and causal interpretations. Licensing: author-generated aggregate data and documentation are CC BY 4.0; analysis/replay code is MIT. These licences do not extend to excluded third-party inputs or dependencies. AI-assisted development with ChatGPT/Codex and Claude is documented in the package. Funding: National Social Science Fund of China, General Project, grant 24BJY144. Research contact: xuc14@qdhhc.edu.cn.