Carl Hennigan

Anybody who wants to join on that, it'd be great to hear from you because it's a huge issue now.

Yeah, we're starting it because we put so much weight on a single model and yet daily all the data is changing, isn't it?

How on earth do you end up joining models together like that?

Well, you don't.

It should be a live platform.

It shouldn't be a static production and saying, here's our model.

Once you've got the input to the model, you can produce outputs on a daily basis with updated data.

And you can keep producing that.

And what you want to see is how stable it looks.

And if it's not stable, for instance, the estimates are changing significantly on a daily basis.

So one day you've got 50,000 deaths, the next day you've got 20,000.

You know you've got a problem.

And so you can't publish it at that point, but you can produce the information.

That's a classic example of where we need a radically different approach.

I think it's really interesting about prediction rules.

We have to understand what they're really useful for.

What are they useful for?

Well, what they're useful for is communicating information across different systems.

So, for instance, let's take the news score, which is a score you use to try and say, what's the chances of you're in primary care?

And I'm speaking to an ambulance driver and I'm trying to say, I'm worried about this patient.