Využití velkých jazykových modelů pro predikci cen nemovitostí

This bachelor thesis explores the use of Large Language Models (LLMs) to predict listing prices for apartments on the Prague real estate market. The work directly follows the research of Jindráček (2025), whose hybrid model—combining a ResNet-50 convolutional neural network with XGBoost—achieved an R² = 0.904 and MAE = 1 390 304 CZK. However, an analysis of his methodology identified train-test leakage, causing an artificial overestimation of the results' accuracy. For the purposes of this thesis, a custom dataset of 3,140 apartment listings for sale in Prague was created, scraped from the Reality.iDNES.cz portal in April 2026. For the purposes of this thesis, an original dataset was created consisting of 3,140 real estate listings for apartments for sale in Prague, scraped from the portal Reality.iDNES.cz in April 2026. This dataset included structured tabular data, text descriptions, and photographic documentation. The methodology defines four approaches: (A) direct zero-shot price estimation, (B) estimation supported by the Retrieval-Augmented Generation (RAG) mechanism using the ChromaDB vector database, (C) integration of LLM-extracted semantic features into machine learning models, and (D) estimation based on tabular data and textual descriptions without photographs. Approaches A, B, and D were conducted using both Claude Haiku 4.5 (Anthropic) and Qwen 3.5-35B-A3B (Alibaba) models. Due to high time intensity, approach C was implemented only with the Claude Haiku 4.5 model. A total of 16 model variants were evaluated. The highest coefficient of determination (R2) was achieved by a hybrid model combining stacking with features extracted from property photographs (Approach C); however, the contribution of these features was not statistically significant. A key finding is that Claude Haiku 4.5 with the RAG mechanism achieves R² = 0.735 without any training and does not statistically differ from any tested ML model in its ability to explain price variability; thus, the deployment of RAG demonstrably increases prediction stability. The thesis also confirms that multimodal LLMs represent a promising tool for real estate practice, particularly due to their ability to provide clear justifications for price estimates and interpret the visual standard of a property—qualities that traditional machine learning models lack.