Developers: Aarushi Verma, Cindy Chiu, Dauren Bizhanov, Sydney Donati-Leach
Home-sharing services like Airbnb can become a viable source of passive income if the host chooses the correct price point and the listing is booked regularly. Previous studies focused on predicting listing prices and few focused on availability of a listing. Availability can be an important factor in predicting revenue as the listing will be profitless without any bookings. This research builds a model that is able to predict the price of a listing as well as anticipated annual availability based on Airbnb data. Investors could do a simple calculation of the annual occupancy rate multiplied by the nightly price to obtain the annual revenue of a listing. In this study, we used XGBoost to train on Hawaii data. We also perform cross-domain testing in Broward County, Florida and Crete, Greece. The price model performed well in both US locations while the performance of the availability model was limited.
We see this project being beneficial to prospective investors by providing easier access to information that could help drive their investment decisions. This model can also fairly easily be adopted by other real estate based companies as well as organizations in the hospitality industry.
The data can be retrieved from the http://insideairbnb.com/get-the-data website. The city we’re training our model on is Hawaii, USA. Since the data is huge and zipped, the user can clone the github repository and run the download_data.py under the setup folder. Running the download_data.py on the local machine will retrieve both training and testing data from the website and create a copy on your local machine. The files will be saved under the same folder as the script with 3 separate folders: Train, Test_Broward and Test_Create. Users can open the files on their local machine.
RMSE and MAE of the Price: XGBoost with default parameters has the smallest RMSE and MAE. Our model was able to predict the price with a mean average error of $186.29
Availability model performance on the validation set (weighted average). All models performed almost at the same level in terms of precision with slight dominance with the optimized Random Forest model. However, the training and prediction speed of the Random Forest model is much slower than XGBoost and LGBM classifiers. Therefore, we decided to use the XGBoost model and be consistent with our model choices. All models perform approximately 50% better than random guessing
This project is built with Python 3, and the visualizations are created using Jupyter Notebooks. In order to run all the files without error, please follow the following steps:
Step 1: Git clone the repo on your local machine
git clone https://github.com/aarushi-vermaa/705_FinalProject.git
Step 2: Install required packages
pip install -r requirements.txt
Step 3: Download the data by running the download_data.py file by navigating to the /00_setup folder
python3 /00_setup/download_data.py
Running this script retrieve both training and testing data from the website and create a copy on your local machine. The files will be saved under the same folder as the script with 3 separate folders: Train, Test_Broward and Test_Create. Users can access the files on their local machine. These source files can also be found in /01_source_data/ folder.
Step 4: Preprocess the train and test data
python3 /00_setup/preprocessing.py
Running the preprocessing.py in /00_setup folder will perform feature transformations on both the training set and testing files.
Step 5: Train the models
For the price prediction model:
python3 /00_setup/train_models_regression.py
For the occupancy prediction:
python3 /00_setup/train_models_classification.py
Those two files contain multiple models that we have implemented for predictions. The train_models_regression.py includes RandomForest Regressor, XGBoost Regressor and LightBGM Regressor. The train_models_classification.py includes RandomForest Classifier, XGBoost Classifier and LightBGM Classifier Running these files will also save the model results as JSON files to /30_results/Models folder. In addition, the evaluation metrics of the validation data, including precision, recall, confusion matrix, RMSE and MAE, are saved in the /30_results/Model Results folder
Step 6: Test Model performance
python3 /10_code/predictions_on_test_data.py
We leverage the model that performed the best on the validation set and predict it on the testing data (Broward County and Crete). The evaluation metrics of the testing data, including precision, recall, confusion matrix, RMSE and MAE, are saved in the /30_results/Model Results folder.
Step 7: Visualize Model performance
python3 /10_code/model_assessment_plots.py
python3 /10_code/confusion matrix.py.py
We can further visualize those testing results and validation results using plots. This includes confusion matrix for occupancy prediction and residual plot for price prediction. These plots can be found in 30_results/Plots folder.
Step 8: Generate Revenue
python3 /10_code/predictions_on_test_data.py
We leveraged the model that performed the best on validation set for both occupancy and price prediction. Then we multiply the results and generate a range of revenue per year for the user. The revenue results are saved in the /20_intermediate_files folder.
The detailed research paper can be found here
The video of our project presentation can be found here