A ML based model to predict the COVID-19 growth using historical data from existing countries
First attempt to predict the growth was by using fbProphet. However, since the scale of data was too small, accurate results were not obtained. A linear increment was predicted, and clearly it was not the case.
The second attempt was by using LSTM networks in Tensorflow to predict data. Such a network needed to be trained first. The WHO dataset https://covid.ourworldindata.org/data/full_data.csv was used. The dataset consisted of time-series growth rates of various countries. We tried using the data from China for training. However, it consisted of outliers and the records are available when the growth was at an advanced stage in China. Therfore the results were not satisfactory.
After noticing similar patterns of growth in early stage in various countries, including India, countries like Austria, Czech Republic were used to train the model. Of them Germany yielded the best results. Loss during training was in ranges of 4.2715e-05and Root Mean Square Error (RMSE) in ranges of 40s. The prediction was spot on.
LSTM based predictions are not suitable for predicting farther time ranges. I have hereby listed for the next 6 days.
Date | Total Cases |
---|---|
21-03-2020 | 277 |
22-03-2020 | 324 |
23-03-2020 | 360 |
24-03-2020 | 383 |
25-03-2020 | 397 |
26-03-2020 | 405 |
If the preventive measures taken were not effective in the last 20 days, the growth will cross 400 within a week