On the Performance of LLMs for Real Estate Appraisal
_{Margot Geerts, Manon Reusens, Bart Baesens, Seppe vanden Broucke, Jochen De Weerdt [2025]}

This study explores the potential of Large Language Models (LLMs) in real estate appraisal, leveraging prompt optimization with in-context learning (ICL). This project evaluates various prompting strategies, and investigates LLMs' ability to provide a viable alternative to traditional machine learning (ML) models, in terms of prediction accuracy, uncertainty estimation and interpretability.

Repository structure

This repository is organised as follows:

.
├── config/
│    ├── barcelona_test_indices.txt: the test indices for the Barcelona dataset
│    ├── beijing_test_indices.txt: the test indices for the Beijing dataset
│    ├── KC_test_indices.txt: the test indices for the King County dataset
│    ├── context.json: the market reports for the three datasets in different time periods
│    ├── features.json: the features used in the three datasets
│    ├── prompt_template.json: the prompt templates
│    └── task_definition.json: the task definition prompts
│
├── data/
│    ├── raw/
│    │    
│    └── processed/
│            ├── barcelona.csv: the Barcelona dataset
│            ├── beijing.csv: the Beijing dataset
│            └── KC.csv: the King County dataset
└── scripts/
        ├── baselines_hedonic.py: script to run the hedonic lightgbm baseline
        ├── baselines_knn.py: script to run the k-nearest neighbours baseline
        ├── baselines.py: script to run the lightgbm and k-nearest neighbours baselines
        ├── gather_features_ollama.py: script to gather feature rankings from Ollama
        ├── gather_features_openai.py: script to gather feature rankings from OpenAI
        ├── gather_features_vllm.py: script to gather feature rankings using VLLM
        ├── gather_intervals_ollama.py: script to gather prediction intervals from Ollama
        ├── gather_intervals_openai.py: script to gather prediction intervals from OpenAI
        ├── gather_intervals_vllm.py: script to gather prediction intervals using VLLM
        ├── gather_prices_ollama.py: script to gather price predictions from Ollama
        ├── gather_prices_openai.py: script to gather price predictions from OpenAI
        ├── gather_prices_vllm.py: script to gather price predictions using VLLM
        └── utils.py: utility functions for creating prompts and processing responses

Installing

We have provided a requirements.txt file:

pip install -r requirements.txt

Please use the above in a newly created virtual environment to avoid clashing dependencies.

Reproducing the results

To reproduce the baseline results, taking the barcelona dataset as an example, run the following commands:

python scripts/baselines.py --dataset barcelona
python scripts/baselines_hedonic.py --dataset barcelona
python scripts/baselines_knn.py --dataset barcelona

Arguments:

• --dataset: the dataset to use, options are barcelona, beijing, and KC.

To reproduce the results with Ollama, first install Ollama according to the instructions in the Ollama website, then run the following command:

python scripts/gather_prices_ollama.py --dataset barcelona --model llama3.2 --output_file results/barcelona_ollama.csv 
python scripts/gather_intervals_ollama.py --dataset barcelona --model llama3.2 --output_file results/barcelona_ollama.csv
python scripts/gather_features_ollama.py --dataset barcelona --model llama3.2 --output_file results/barcelona_ollama.csv

Arguments:

• --dataset: the dataset to use, options are barcelona, beijing, and KC.
• --model: the Ollama model to use, options are llama3.2 and deepseek-r1:7b.
• --output_file: the output file to save the results to.
• --examples: the number of examples to use for the prompt strategy, options are 0, 3, 10.
• --example_selection: the prompt strategy to use, options are geo, hedonic, mixed.
• --context: whether to use context or not, options are False, True.

e.g. using the 10 ex. mixed prompt strategy:

python scripts/gather_prices_ollama.py --dataset barcelona --model llama3.2 --output_file results/barcelona_ollama_prices_prompt_strategy.csv --examples 10 --example_selection mixed

To reproduce the results with OpenAI, make sure you have an API key and run the following command with the same arguments as above:

python scripts/gather_prices_openai.py --dataset barcelona --model gpt-4o-mini --output_file results/barcelona_openai_prices.csv --key <API_KEY>

Citing

Please cite our paper and/or code as follows:

@InProceedings{Geerts2025LLMsRealEstate,
  title={On the Performance of LLMs for Real Estate Appraisal},
  author={Margot Geerts and Manon Reusens and Bart Baesens and Seppe {vanden Broucke} and Jochen {De Weerdt}},
  booktitle={Machine Learning and Knowledge Discovery in Databases. Applied Data Science Track},
  year={2026},
  publisher={Springer Nature Switzerland},
  pages={195--211},
  doi={10.1007/978-3-032-06118-8_12}
}