This is our submission for Data Crunch 2025, organized by CSE UOM.
The main notebook used for the final submission can be found in Notebooks/DataCrunch_The_Final_Notebook.ipynb. Other notebooks can be found in Notebooks/experiments.
After much debating, experimenting and sleepless nights, Facebook Prophet yielded the most performance - more than ARIMA, XGBoost and even various LSTM derivatives.
Neural Prophet might have been somewhat better still, but setting it up was a big hassle, with lots of dependency issues (mismatching numpy versions etc.), that we ultimately ended up with Prophet as our final solution, since it seemed to outperform pretty much every other solution we came up with.
It is recommended to run this notebook in a Google Colab with Google Drive mounted and the CSV data files stored in a folder named DataCrunch in your Google Drive.
For local development, the following dependencies are required to be installed:
pip install numpy pandas scipy statsmodels matplotlib seaborn prophet pmdarima xgboost tqdm ipykernel jupyterMake sure to remove / avoid running Colab-related code (e.g. mounting Google Drive) before running the notebook locally.
Made with ❤️ by Caramel Labs