A multimodal model for commercial real estate valuation based on the integration of geospatial data and language-model-derived features

Abstract

The aim of this paper is to substantiate methodological approaches to commercial real estate valuation through the integration of heterogeneous data, and to demonstrate their application within the Moscow market. A methodology for constructing a multimodal automated valuation model (AVM) is presented, built on the integration of structured, geospatial, and semantic data. The study was conducted in the street retail segment of the Moscow commercial real estate market. Three sequential feature configurations were developed: a baseline model, a model extended with geospatial characteristics, and a model further augmented with interpretable features extracted from property text descriptions via the GigaChat language model. Modeling was performed with CatBoost and LightGBM algorithms. Hyperparameters were optimized through cross-validation on the training sample, and final model quality was assessed on a held-out test set. Adding geospatial features reduces the mean absolute percentage error by 17–20% for both algorithms-a significant reduction. The inclusion of LLM-derived semantic features yields a further accuracy gain of 1.6–3.3%. SHAP analysis confirms the dominant role of spatial factors, while qualitative property characteristics also contribute meaningfully to model output. The multimodal approach delivers higher accuracy and interpretability than conventional single-source models, demonstrating its applicability to mass appraisal tasks and its practical value for the banking sector, real estate developers, and valuation firms. The theoretical contribution of this work lies in identifying a viable direction for the development of valuation methodology in the era of digital data. The approach does not substitute for expert judgment. It reinforces it, providing practitioners with a scalable, reproducible, and objective analytical instrument that formally incorporates both spatial and semantic context. The results have practical relevance for real estate lending institutions, valuation companies, and entities owning or managing commercial property portfolios. They are also directly applicable to regional state budgetary institutions of the Russian Federation responsible for cadastral valuation, and to Rosreestr (the Federal Service for State Registration, Cadaster and Cartography) in its efforts to advance mass property appraisal methodology at the national level.