peter-amerkhanian / San-Francisco-Service-Analysis

Comparing various machine learning methods -- random forest, RNN, LSTM -- with ARIMA for forecasting street sweeping demand in San Francisco

San-Francisco-Service-Analysis

Project Title: Forecasting and Inference for Public Sector service provisioning

Team Members: Peter Amerkhanian, Jared Schober

Instructions

• Run long_run_ARIMA.py for output of all long-run ARIMA forecasting results (~3 hours). This script will use public_data/processed_data.csv
• Run Non_ARIMA_Models.py for output of all next-hour and long-run models besides ARIMA (~4 hours). This script will use public_data/processed_data.csv
• Run Geospatial_Models.py for output of OLS and Random Forest Next-Hour models applied at the census tract level. This script will download the 311 data directly and requires an internet connection.

Presentations and final paper are in Written Materials.

Datasets:

• SF 311 calls for service (2008-2023)
  • n=2.28 million
  • Geo-spatial distribution:

WORKFLOW

git pull
# start working
git add .
git commit -m "your message here"
git push

Project Description:

We seek to forecast the hourly need for street cleaning services in San Francisco, so that the city can more accurately prepare for service delivery. We will use the data as a time series and will aim to produce forecasts that are accurate at one week.