What to feed to the input of the neural network? Your ideas... - page 47

[Yuriy Asaulenko]

Maxim Dmitrievsky #:

history window... normalise the entire training history or renormalise at intervals

No. Normalise each window of NS or MO inputs.

[Maxim Dmitrievsky]

Yuriy Asaulenko #:

No. Normalise each window of NS or MO inputs.

Got it. Well, here is food for experiments for the author of the topic.

[Andrey Dik]

mytarmailS #:
Well, crawl back to your hole...

and you shouldn't have taught me and corrected me when I gave good advice to Alexei, if you're a complete ignoramus yourself.

I give you a link too, maybe because your eyesight is bad, you may not have noticed it:

https://www.mql5.com/ru/forum/87536

[Yuriy Asaulenko]

Maxim Dmitrievsky #:

Got it. Well, here is food for experiments for the author of the topic.

))

[mytarmailS]

Andrey Dik #:

I'll give you a link too, maybe

You give me a link to the definition of what you were asked.

Or apologise and fume.

[Andrey Dik]

Stanislav Korotky #:

IMHO, the condition of uniqueness and stationarity of the FF maximum is impossible to fulfil because the market itself is by definition a non-stationary process, subject to many unpredictable external influences (no detrending or taking derivatives will save us). The only thing we can use for successful optimisation (and subsequent forecasting) is the relative inertia of the market, but of course, provided that we consider the most liquid instruments with large volumes of trades and participants. Then we can find a sufficiently wide FF wave, which, although moving with time, still gives a FF value close to the extremum at the step between optimisations.

There is a (high) probability that no wide wave is found in the FF. I would not call such a FF inappropriate to the process and throw it out immediately, but try to add another layer of meta-optimisation/forecasting - on the sequence of FF wave surfaces on history (i.e. generalise/formalise the step-by-step transformation of waves and be able to synthesise the waveform for the next step). Ideally, this would be logically built into Walk-Forward Optimisation, but I haven't got round to it yet.

Of course, we always keep in mind that we are dealing with non-stationary data, it is not even worth mentioning. But there is also the fact that the price series itself is discrete. Therefore, when I said about a stationary island of FF, I meant something like a shimmering island, the FF values on this island will change slightly in the window.
The other sets of parameters in the window will look like sprawling waves.
It follows that robust parameters should be sought among those that "shake" the least in terms of FF values. This can act, as you wrote, as a Meta-FF that minimises sub-FF fluctuations

And about process mismatch - if there are no more or less visible stable sets forming "land areas", then there is no reason to believe that the system has stable sets. So either the FF does not match the process, or the process has no stable sets at all, which is what I meant. You can, of course, keep adding/removing metrics to the FF, but that would just be a different FF.

[Ivan Butko]

Yuriy Asaulenko #:

First hit NS training directly on the quotes. Evaluation of training on an independent sample by epochs. Those who have dealt with TensorFlow will understand.

This is already quite enough for some profit. It may not be enough, but the first copy I found.

Look quickly, I'll delete it.)))

You don't understand the results of Python

It sticks the predicate line to the fact line and the predicate line hangs behind it like a tail, like a lagging moving average with a period of 5.

Just check it not at the macro level (the picture is beautiful at the macro level), but at the micro level - NS does not guess the next price. 50/50 with the direction.

It can be MLP with 100500 layers and 100500 neurons, or CNN+LSTM+MLP+dropouts+regularisations+sacred writing+guessing on coffee grounds.

Same with RL. Nothing works in Python, unfortunately.


UPD

You need some creative/creative trading approach or some complex system.

The only grail that works is trading without spread. It's easy to make one. But, there are no brokers like that.

[Ivan Butko]

Yuriy Asaulenko #:

Once again, slowly - estimation by epoch on an independent sample of non-participants in the study).

We've been through everything. And on the independent sample, and on the neighbouring sample, and on someone else's sample, and on the fifth sample.


Let's make it clear: show a working stat. A working EA. If you have one, you're right. And if you don't - you are on the same stage of development as all forum participants - in the context of searching for a consistently profitable system. Many people have academic knowledge here. They are not quoted in the context of the art of trading.

[Ivan Butko]

Yuriy Asaulenko #:

I'll be right back. I'm just gonna run out and get a discharge order.

Well, that makes sense.

Filter the upstarts.

[mytarmailS]

Yuriy Asaulenko #:

Actually, the topic is about ideas, not about measuring...))))

"What to feed to the neural network input? Your ideas..."

You've got something wrong again.))

You can do a backtest.

And you have to.

What's the point of looking at the MAE