For latest predictions, please visit our tracker.

Date Predicted daily cases Actual daily cases Difference
Jun 5 8868 9851 983
Jun 6 9082 9887 805
Jun 7 9339 9971 632
Jun 8 9933 9983 50
Jun 9 10390 9987 -403
Jun 10 10820 9985 -835
Jun 11 11231 9996 -1235
Jun 12 11622 10956 -666
Jun 13 11675 11458 -217
Jun 14 11937 11929 -8
Jun 15 12509 11502 -1007
Jun 16 13060 10667 -2393
Jun 17 13213 10974 -2239
Jun 18 13456 12881 -575
Jun 19 13566 13586 20
Jun 20 14060 14516 456
Jun 21 14364 15413 1049
Jun 22 14431 14821 390
Jun 23 14835 14933 98
Jun 24 15384 15968 584

Predictions

Predictions last updated: 4 Jun

Data source: OWID

The 10 day forecaster that we talked about in the previous post, allowed us to predict only the next 10 days of Covid19 new cases by looking at the past 20 day trend. The look ahead window of 10 days is too small and we started with it as we had limited computing resources. Greater the number of days that we can predict, the more is the time that we get to prepare for the future. Hence we enchanced our neural net to learn from 40 day trends and predict next 20 days. As of now we don not have sufficient data to build any longer predictors. However that comes with a cost: the size of the network increases.

Let’s revisit the concept of neural nets. A neural net is a universal approximator with learnable parameters called weights. That means, a neural network can be used to approximate any curve of the form y = f(x) provided some data points lying on the curve are known. This data is used to identify the relationship between x and y by figuring out the correct values of weights. That process is called training the network.

When we try to predict larger trends by giving larger inputs and expecting larger outputs, the number of weights in the network increases. In our case, when upgrading from a 10 day forecaster to a 20 day forecaster, the weights increase in number from 5170 to 27940. Consequently training time increases too. That means, it takes a lot more time to figure out the right combination of these weights which will satisfy the relationship between x and y.

Since we have access to limited resources at the moment, we couldn’t refine the model as much as we would’ve liked to and hence it does not fit as well as the 10 day forecaster. Take a look at the comparison of the fit in the charts below. The top chart shows the 10 day predictions made everyday by the 10 day forecaster. And the bottom chart does the same for the 20 day forecaster.

Fit comparison

The predictions of the 10 day forecaster made each day, fit very close the the actual cases (blue line), as against the 20 day forecaster, in which case, predictions don’t align very well.

However when we use the models to make predictions for a longer term (say a few months ahead), the 20 day forecaster does much better. The 10 day model predicts the peak too early which is very unlikely. Take a look at the peaks projected by both the models below. Top chart for 10 day model and bottom chart for 20 day.

Peak comparison

There are a few more standard techniques in deep learning to improve these forecasters. However we don’t have enough resources at the moment to perform multiple simultaneous computations. Also, in a few more days, enough data would be available to build a 30 day forecaster as well. Please let us know in case you can arrange for a GPU enhanced computer for training these nets.