Time Series Forecasting with SARIMA, Holt-Winters’, and Prophet
An introduction into the three models with a project from Flatiron School
For my latest project in Flatiron’s Data Science bootcamp, I was tasked with running a time series prediction. I chose to go with data on interstate traffic volume on I-94 westbound from St.Paul to Minneapolis, found on the UCI machine learning repository (here). This particular stretch of road is known for backups and closures, drastically effecting commute times. There were over 40,000 observations of hourly volume for a six year period between 2012 and 2018, with some major gaps leading up to 2016.
In order to standardize this data, I chose to work with 2016–2018, and fill in any gaps which I did with an average of the hour before and after, then create a daily summation of these observations. This resulted in an average of 80,000 total vehicles being recorded on a daily basis. After cleaning, I was left with 802 days. This was split into 738 days for my training data, and 82 for my test. Exploring the data, this gave me a weekly seasonality to run my model.
