In the last days, we had a lot of technical details.

It was the idea to have sinking blocks.

And there's one more sinking block

which we have touched in the beginning.

And this is this discussion.

We have developed so nice models.

For example, this model here.

But what do I learn as a human being out of such a structure?

Hopefully, this is a good explanation between the inputs

and the outputs.

But how do I can clap this?

And this is the story with the sensitivity analysis,

which we have mentioned before.

But now I have to explain you how to do it really

in practice.

And so here we are.

So whitening the black box is the counter argument

against an often used argument against neural networks

that people say, yeah, this is only black box modeling, which

is a negative argument in this wording.

So you know nothing about what's really going on.

The machine is doing something.

But yeah, what is it?

And so the point here is the neural network

is a quantitative model, which in itself is a good news.

But you also have to be able to squeeze out

qualitative interpretations out of this.

Otherwise, you will lose your customers.

And there's a second point which is going the same direction.

And this is if you speak about a customer project,

then often you are sitting there in the situation

that customers tell you, yeah, we want to model the,

for example, the Siemens stock index.

But this thing depends on the complete, not only

on the German economy, it depends on worldwide economy.

So you might be forced in the direction of using hundreds

of input variables.

This is an overkill.

You might, in principle, be able to use

the hundreds of input variables to explain

whatever you want to explain.

But then you would need so many data examples

to find the interrelationship between all the variables there.

So that is hopeless.

We do not have so, and normally we do not

have so long time series there.

And so therefore, a very practical question

is how to select the reasonable input factors if I

want to do the model building.

And this, again, is a question on input-output analysis,