Forecasting COVID-19 cases using dynamic time warping and incremental machine learning methods

The investment of time and resources for developing better strategies is key to dealing with future pandemics. In this work, we recreated the situation of COVID-19 across the year 2020, when the pandemic started spreading worldwide. We conducted experiments to predict the coronavirus cases for the 50 countries with the most cases during 2020. We compared the performance of state-of-the-art machine learning algorithms, such as long-short-term memory networks, against that of online incremental machine learning algorithms. To find the best strategy, we performed experiments to test three different approaches. In the first approach (single-country), we trained each model using data only from the country we were predicting. In the second one (multiple-country), we trained a model using the data from the 50 countries, and we used that model to predict each of the 50 countries. In the third experiment, we first applied clustering to calculate the nine most similar countries to the country that we were predicting. We consider two countries to be similar if the differences between the curve that represents the COVID-19 time series are small. To do so, we used time series similarity measures (TSSM) such as Euclidean Distance (ED) and Dynamic Time Warping (DTW). TSSM return a real value that represents the distance between the points in two time series which can be interpreted as how similar they are. Then, we trained the models with the data from the nine more similar countries to the one that was predicted and the predicted one. We used the model ARIMA as a baseline for our results. Results show that the idea of using TSSM is a very effective approach. By using it with the ED, the obtained RMSE in the single-country and multiple-country approaches was reduced by 74.21% and 74.70%, respectively. And by using the DTW, the RMSE was reduced by 74.89% and 75.36%. The main advantage of our methodology is that it is very simple and fast to apply since it is only based on time series data, as opposed to more complex methodologies that require a deep and thorough study to consider the number of parameters involved in the spread of the virus and their corresponding values. We made our code public to allow other researchers to explore our proposed methodology