C19 Neural Network

I mentioned in my first C19 blog post that I had fiddled with machine learning, specifically neural networks (NN) but quickly moved on to models that have a known mathematical form (logistic, lognormal) with unknown parameters that are estimated by fitting the data. That is what I have been giving each day in the C19 update.

Here is an example of NN fitting of the world cumulative death data. For geeks, this uses the scikit learn mlpregressor. Below that is today's update logistic model fit.

Briefly, the problem with neural nets in this case is it can only estimate the future from past data points and nothing in the past would indicate a topping out of the curve. The logistic fit predicts a flattening because it is an equation based on population dynamics and other physical systems. I will be adding the NN fit to each day's update at the bottom for reference. As the cumulative data points roll over and flatten out we can expect the NN prediction to do the same. But that is not much of a prediction.

Neural network model

Logistic function model