Time series forecasting results from my Data Science 2 class project (at The University of Vermont)
All code in this repository is original code. Actual research was performed in project_main.py, but this can be hard
to follow without visuals. To provide some intuitive insight into my data science process, check out the attached notebook
demo-forecasting-covid-prisons.ipynb which will walk through my process of cleaning, transforming, and modeling the
data, and my main findings.
Pictured below is a sample visual for one of my fitted ARIMAX models, visualizing the out-of-time cross validation prediction accuracy of the fitted model on unseen data.
In this project, I attempt to model new weekly cases of COVID-19 amongst the state prison population in California and Washington using an ARIMAX modeling framework. Under this framework, I attempted to model the Residents time series (incarcerated people in the state prison system) using two potential explanatory variables: The time series of Staff cases, and the time series of Statewide cases.
In the attached notebook, demo-forecasting-covid-prisons.ipynb, I walk through how I performed exploratory data analysis of the three time series variables, flagged and cleaned any flawed data, transformed the data, and modeled the Residents time series using one or more of the explantory variables. I show Out-Of-Time cross validation predictions to compare the predictive accuracy (MAPE) of models against one another, to aid in model selection for the best model, and show a few figures of the fitting process and predictive modeling steps.
From this project, I found that the Staff variable could best explain the Residents time series, better than a baseline model using just Residents by itself, and better than the State time series.
Data sources are the UCLA Covid Behind Bars Data Project found here and the New York Times found here.