Dear All & JCL,

I have just tested DeepLearn and would like to do more. Hope to get some supports from the community.

1. I want to add more signals for the function advise().

In the Zorro Online Manual, said: "For using more than 20 signals, collect them in a global array and send them to the machine learning algorithm with NEURAL_LEARN or NEURAL_TRAIN."

But, I don't know how to "send them"?

I read the include\r.h and found that: neural.train called with the Data with signals + 1 more field, so it means that advise() function did do something with the signals ---> I could not understand this.

2. In the Better Strategies #5, hinted 6 possibilities in the conclusion, (1) is one of them, so anybody have anything to do with those? Some I never heard of, so help me please, any clue

Thanks in advance.

Se7en

I have not used more than 20 signals so far, so I have no ready script for this. The general recipe is this. First examine closely the "neural" function framework in r.h until you have completely understood what it does. Then write your own "neural" function with different NEURAL_TRAIN and NEURAL_PREDICT parts. For NEURAL_TRAIN, collect the signals in an array, and write it to the CSV file instead of the signals from the Data parameter. Then do the same for NEURAL_PREDICT: send the current N signals over to R instead of the 20 signals from the Data parameter. For more than 20 signals you also must use the "Objective" parameter of the advise function, rather than the outcome of the next trade.

Thank you JCL.

First of all, I want to confirm this first.

In the "Better Strategies 5", you suggested 6 possibilities to do.

[1] "Use inputs from more candles and process them with far bigger networks with thousands of neurons."

[4] "Use inputs from many candles – such as, 100 – and pre-process adjacent candles with one-dimensional convolutional network layers."

That's why I thought of adding more signals for the function advise(). Is it right?

Thanks.

Se7en.

Yes for (4). In the case (1) I would not use more than 20 inputs. I had the idea to pass the prices from candles with increasing time periods, f.i. from 1, 2, 3, 5, 10, 15, 30, 60, and 120 minutes ago. But that was just an idea, I have not tested it yet.

Dear JCL,

I did what you said to understand the NEURAL function.

1. I could not create the same Data for NEURAL_TRAIN.

In your DeepLearn, you used 8 signals, then Data contained 9 fields included: 8 signals & plus one more (a very special signal).

How could I create the plus one?

I thought of ignoring it. So, I created the signals[Core].csv with the last one empty.

2. Data for NEURAL_PREDICT was EMPTY (NULL).

And the result was:

Train ran OK.

Test ran with nothing happens.

PLEASE HELP ME OUT

OR

give me anything to improve the DeepLearn, such as: LEARN MORE!!!

Thank in advance.

Se7en.

No, it's only 8 signals. The 'one more' is not a signal, but the training target. And I don't think that the Data pointer is NULL in NEURAL_PREDICT, unless the number of signals is also 0. Otherwise you could not run a test, as it would immediately crash.

Oh!!!

I kept NEURAL_PREDICT script the same, changed NEURAL_TRAIN with my 8 signals Data. It was NOT work!!!

So, How could I get the training target?

Which target do you want to train?

Thanks for your asking.

I understood how to use Objective in the function advise().

Trying to include the Objective into my own Data and maybe I did something wrong with the next trade result.

Thanks agains.

Do you think it´s worthless to put more than 20 signals? Because if we are using series like vol[1], vol[2], etc with four signals (ROCP,vol,range...) it seems very poor. I think that more than 20 signals can be interesting because allows us to use exogenous data like COT or sentiment indicators and series.

You can now use as many signals as you want. The systems that we did for clients had mostly about 10 signals. With more signals, it becomes hard to produce a successful system. But there might be cases where you use about 100 signals, f.i. when preprocessing them with a covolution network.

It´s usefull to put the signals in series? For example: vol[1],vol[2],vol[3],ZScore[1],..........

Thanks.

If you want to evaluate the history of the parameters, then series make sense.