1st Place Winner of Spring 2021 Session 1
by Creeper: Ethan Hanover, Nicholas Lin, Jason Yu
Abstract: This experiment seeks to create a model that predicts the future change in the number of new COVID cases based on a variety of factors including current COVID cases, mobility data, vaccine distribution, and state-sanctioned restriction policies, all of which are recorded for each state in the USA on a daily basis. We initially used classical machine learning (linear regression, random forest) to predict percent change in case count 7 days into the future. The random forest model predicts the percent change in case count 7 days into the future with an RMSE of 0.15 (i.e. on average, the predicted percent change in new cases is about 15 percent off of the true change). We also used recurrent neural networks (RNN) to model the data since RNNs are better suited to predict time-series data such as COVID cases over time. Specifically, we experimented with LSTM models to forecast future COVID cases in Virginia based on training data from other states. Running the model on hypothetical timelines with altered variables provides qualitative data describing the impacts of these restrictions or preventative measures, which can ultimately help inform Virginian policymakers about the best steps to take to effectively mitigate the spread of diseases in the future.
Comments