Machine learning in trading: theory, models, practice and algo-trading - page 60

 
Alexey Burnakov:
What is the duration of the training and validation data? Does it look like a couple of days? This does not say anything, to be honest.
Well actually it's 100 buy signals and 100 sell signals on 5 minutes is three weeks.... so.....
 
Mihail Marchukajtes:
Well actually it's 100 buy signals and 100 sell signals on 5 minutes is three weeks.... so.....
It's still not enough for me. Count the probability that it is not pure luck for 3 weeks.
 
Mihail Marchukajtes:

And it is possible to bring any input data to the output data and the system will work for some time, so he who is looking will always find :-)

So, really, a little trick with data and we have a level of generalization grows to acceptable numbers in 90%.....

Thanks for the tip! Because we're trying to predict the future price trend here. But it turns out that it is already trivially known from the TA signals and we only need to classify the signals by their reliability.
 
Alexey Burnakov:
It's still not enough for me. Count the probability that this is not pure luck in 3 weeks.
No. It is important to have the same patterns during three weeks and the main thing that the market reaction to them was exactly the same as in the future for a week at least.....
 
Mihail Marchukajtes:

Well, the current example that works in the picture I have now made from this file

Thank you!

It's better to feel once than hear a hundred times. It's always much easier to understand with real examples, because something may not be told in words.

 
Yury Reshetov:
Thanks for the tip! We are trying to predict the future price trend here. But it appears that it is already trivially known from the TA signals and we only need to classify the signals by their reliability.

There are two levels of experts on the NS. The first are developers, which is you Yuri, the second are users, which is me. It is one thing to write a network, another thing to know how to use it. Networks are divided into two kinds, some proznoziruyut. Others classify. A predictive network answers our question... "In the future the forecasted value (in N bars, the forecast threshold) will be so-and-so or so-and-so". From this we draw a conclusion and make a decision. The classification network tells us. "The current situation is true or false" and we again infer from this information. Since it is IMPOSSIBLE to know the future, I decided long ago that it is necessary to determine exactly where we are and then draw a conclusion and appropriate steps.

Garry Kasparov, the world chess champion, was once asked how many moves ahead in the game he thought he was planning his next move. Everyone thought that Garry was going to say a huge figure and reveal the secret of the winner. However, what the chess player said proved that not everyone understood even the essence of the game: "The main thing in chess is not how many moves you think ahead, but how well you analyze the current situation.

The whole essence of Yury's method, for which I would like to express my gratitude, and for my part I'm ready to help in constructing and using his brainchild :-)

 
Mihail Marchukajtes:
After the red line out of the sampler or out of sample for beginners. In my opinion, it is quite viable. The truth has already made mistakes for today, but it's not terrible..... There's no such thing as a mistake....

You show a beautiful picture. I don't believe in miracles.

I can find a picture where after your buy signal the market will continue to move down a couple of figures. The same for selling. How do you determine that the market will behave normally within a week after the learning curve?

 
Alexey Burnakov:

You show a beautiful picture. I don't believe in miracles.

I can find a picture where after your buy signal the market will continue to move down a couple of figures. The same for selling. How do you determine that the market will go down in a week after the learning curve?

Nothing. You have to watch the number of errors from the total number, and if the errors go up, you should overtrain the network. For example, if you get 4 mistakes out of 20 signals, it is ok, if there are more mistakes, you need to retrain the network. There is another question ... what model to choose so that it can be trusted for 10 signals. Well, Yuri has already kind of described it. Choose the model that shows the maximum level of generalization of both binary and trinary models and start working on it. And to qualitatively increase the working time of a strategy, we should increase the trainable interval, and to increase it we should increase the number of inputs. I.e. 10 inputs can saw 100 signals to zero. 15 inputs can break 225 records, and this is 6 weeks of signals, respectively, the network running time out of sample will be longer, not a week, but two... With the proper level of error. It is not possible to work without errors. I would like to, but it is not possible. The main thing is to reduce the impact of this error on the deposit and I'm all set :-)
 
Alexey Burnakov:

You show a beautiful picture. I don't believe in miracles.

I can find a picture where after your buy signal the market will continue to move down a couple of figures. The same is true for selling. How do you determine that the market will behave in the same way during the week after the learning curve?

If you want to buy, the market will go against you, you will get a loss and wait for another signal that will kick your loss and earn from the top. Another thing is that TS starts to make a mistake after a mistake, that`s right :-( But I hope it won`t be like that :-)
 
Mihail Marchukajtes:
I don't know why I started to make such kind of errors, but I'm sure I won't.

You have no idea how your model works over long periods: there is no big forward test and no walking forward test. Your model may be a set of learned noises that you mistook for signals in a few well-matched pictures.

I suggest you reconsider your approach and do extensive testing first. You'll see troughs and mountains and maybe, but unlikely, you'll beat the zero-mat expectation a little. And you can do all this before you start losing real money.