Call Volume Prediction for Seattle Fire Department

The project is developed by Yusuf Can Simsek, as an answer to the request from niologic GmbH. Detailed explanation of the design phase can be found in the notebook with last run's outputs (29.09.2022).

Installation

First, clone the repository and install the requirements:

git clone https://github.com/yusiyohpolimi/call_volume_prediction.git

pip install -r requirements.txt

Training

To train the model (XGBRegressor), use train.py script

usage: train.py [-h] [--start START] [--end END] [--download] [--rfe] [--cv]

optional arguments:
  -h, --help     show this help message and exit
  --start START  start date to trim dataset
  --end END      end date to trim dataset
  --download     download dataset from the source
  --rfe          use Recursive Feature Elimination
  --cv           use 5-fold cross-validation for time series, need 5+ years of
                 data

Testing and Future Prediction

Use test.py script to test the trained model with a subset of the dataset. Other script, predict.py, is used to make a future prediction, out of the dataset.

usage: test.py [-h] [--start START] [--end END]

optional arguments:
  -h, --help     show this help message and exit
  --start START  start date to trim dataset
  --end END      end date to trim dataset

usage: predict.py [-h] [--start START] [--end END] [--freq FREQ]

optional arguments:
  -h, --help     show this help message and exit
  --start START  start date to make prediciton
  --end END      end date to make prediction
  --freq FREQ    resolution of the prediction

Notes

• There are recent (29.09.2022) updates for sklearn and xgboost which HIGHLY affected the performance particularly in colab.
• Check the notebook for the details of the project, and the followed steps.
• Downloading weather data takes VERY long time, prefer using from the repository.
• Covid pandemic seems to have a strong effect on call data trend.
• Weather data other than temperature does not yield satisfactory results for this weather dataset. However, other services require subscription.