Baseline accuracy of forecasting COVID-19 cases in Moscow region on a year in retrospect using basic statistical and machine learning methods

Accuracy of COVID-19 evolution models for different forecast horizons

“…Introduction Since the beginning of the coronavirus pandemic, several predictive models have been proposed to determine the number of new cases [1]. Along with generally accepted approaches motivated by modeling of the spread of the virus, such as SIR [2], SEIR [3] and their modifications [4,5], various methods for analyzing one-dimensional data were applied to forecasting, such as exponential smoothing [6], ARIMA [7] and others. A number of works demonstrated the effectiveness of neural network models based on Long Short-Term Memory layers (LSTM) [1,6,8].…”

“…A recent study [9] showed that the effectiveness of the existing methods applied to the COVID-19 dynamics forecasting is comparable to the estimates based on the "tomorrow as today" method, which indicates the complexity of the task and the need for further improvement of predictive methods. In our previous works [5,7] it was shown that: 1) the accuracy of machine learning models strongly depends on the number of training samples, 2) the usage of the models of this type is inefficient on short prediction horizons (up to 28 days). Due to the fact that retrospective data for 2 years of the development of the pandemic are currently available, including details for individual regions and countries, this work is aimed at creating a model that takes into account the newly-accumulated data sets.…”

“…Due to the fact that retrospective data for 2 years of the development of the pandemic are currently available, including details for individual regions and countries, this work is aimed at creating a model that takes into account the newly-accumulated data sets. Here we evaluate our best-performing models from the last year [5,7] on this year's data. In an ongoing work, we compare the "tomorrow as today" (Dummy) baseline model, Exponential Smoothing (ES), and LSTM-based neural networks (NN) on the new data and show that LSTM with the current data amount can outperform the baseline and ES in accuracy.…”

“…The Holt-Winters exponential smoothing method from [13] with damped additive trend, boxcox normalization and no seasonality was used, which proved itself efficient on the data from the first year of the pandemic [5,7]. The output of the prediction consisted of the new daily cases for the FH days ahead.…”