Using neural networks in trading - page 16

[solar]

The error must be monitored. otherwise it is not clear what we are teaching ))))

[Дмитрий]

This is already section 1.1.2008-1.1.2009. Also training mode. Drawdown 665.87, net profit 703.14 Forward test mercilessly drains:)))

[Дмитрий]

grell:

This is already section 1.1.2008-1.1.2009. Also training mode. Drawdown 665.87, net profit 703.14 Forward test mercilessly drains:)))

Inattention and haste. Forgot to add the "off-market" condition.

[Роман]

solar:

The error must be tracked. otherwise it is not clear what we are teaching ))))

Nice. :-)

I read somewhere that there is an option:

You should stop training when the error grows... (i.e. it should be constantly monitored (error value) and as soon as it grows, training should be stopped...)

[[Deleted]]

grell:

Inattention and haste. I forgot to add the "off-market" condition.

I think it's all the same weak. It is unlikely that the NS from the sheet will perform better than during the training period. And there's already nothing to eat on it...

[Дмитрий]

Figar0:
It seems weak to me all the same. It is unlikely that the NS from the sheet is going to perform better than in the training period. And there's not much to eat on it already...

We need an idea.

[[Deleted]]

grell:

I need an idea.

What idea? And frankly I don't understand why you have a multi-year study. The frame appears to be shallow, trying to find "eternal values" on the verge of noise?

[Дмитрий]

Figar0:
What's the idea? And frankly I don't understand why you have a study in many, many years. The frame appears to be shallow, trying to find "eternal values" at the edge of the noise?

No, trying to show the hidden layer as many examples as possible. So to say not to "learn", but to bring them up to speed. And then, with small samples, to finish learning. If you go straight to the small sample, he'll just learn it. That's in a nutshell.

[Дмитрий]

I'll post a screenshot of the retrained network soon.

[[Deleted]]

grell:

No, I'm trying to show the hidden layer as many examples as possible. So to say, not to "learn", but to bring them up to speed. And then learn the rest with small samples. If you go straight to the small sample, he'll just learn it. That's in a nutshell.

You sound as if you have this hidden layer is something quite isolated and has a life of its own) But I think I understand what you mean ... I'm practicing the opposite method. I find a lot of different solutions, for example on one month's data, then I check it for 2 months, discard the solutions that don't fit the new data, add a lot of solutions, etc.