Short-term COVID-19 forecast based on Back Propagation Neural Network (11-15 May 2020)

Date of publication 10.05.2020

The results of short-term predictive modeling of the number of patients with COVID-19 in Ukraine and Kyiv on 11.05.20 - 15.05.20 (Fig. 1, 2, 3, 4; Table 1) are obtained using a multilayer Back Propagation Neural Network based on the mechanism "sliding window", the length of which is 12 data points. The average absolute error as a percentage of the forecast does not exceed MAPE = 2.2%.. 

As can be seen for Ukraine for the ten days of 06.05.2020-15.05.2020 the nature of the pandemic can probably become clearer linear with the level of "plateau" of 500-550 new infected per day (Fig. 2). According to the results of the forecast modeling of the FORESIGHT COVID-19 project team, this nature of the process may continue until the third decade of May 2020, during which a pandemic peak is likely to occur. After that, the pandemic is likely to decline (the number of daily infected will be consistently lower than the number of daily recovered people).

For Kyiv city, during the ten-day period (06.05.20-15.05.20) the process of pandemic development is still more non-stationary with the probable relatively high volatility of this process next week (Fig. 4).