Application of a Stateful LSTM in the Forecasting of Covid-19 Cases in South Africa
Graeme Lubbe , Tinashe Chinyati
Abstract: For this project, the use of a stateful LSTM in forecasting Covid-19 is evaluated. A stateful LSTM is useful in predicting autocorrelated time series data, as it does not shuffle sample during training, thereby preserving the time series dependency in the data. The chosen model uses a lag on the daily cases to predict future cases, which is determined by evaluating the autocorrelation at various lags. This was found to be 170 days, with an ACF of 0.54, which implies the possibility of forecasting 170 days. Cross-validation and stability is tested through back testing, by evaluating multiple subsets, split in training and testing set. This shows good performance on data earlier in the pandemic, but the RMSE increases from a range of 3300-4400 on the first six slices, to 8100-17100 for the final slices, showcasing some instability and poorer predictability latter in the pandemic.