Machine learning for robots - page 7

[Igor Makanu]

Yuriy Asaulenko:

But in general, I don't see how TC can squeeze anything meaningful out of 200 trades.

Ask Wizard, he used to be very handy with nerocheldaytrader, he could make a neural network indicator out of any indicator or deal history, which would then put buy/sell arrows on the chart.

[Yuriy Asaulenko]

Dmitriy Skub:

Why not? Just present)

It is possible to choose with a large number. The main thing is to have a higher expectation.

I didn't even know about it. I've never been there, never been interested. Now I've taken a look - it's a gold mine.) You can teach me anything you want.)

[Dmitriy Skub]

Yuriy Asaulenko:
Teach me.

Very true.)

[Yuriy Asaulenko]

Igor Makanu:
Ask Wizard, he used to manage neural network indicator smartly, from any indicator or history of deals he could make a neural network indicator, which then set buy/sell arrows on the chart.

That's not a trick. I am surprised only by a small number of deals for the TS training. I don't understand it, for the life of me.

If there is an NS being trained there, with 10 or so neurons. If it's a tree, it's a very small tree. There's nothing to do with big ones.

[Igor Makanu]

Yuriy Asaulenko:
Not a trick at all. I'm only surprised by the small number of trades to learn from TC. I don't understand it, for the life of me.

If it's learning NS, it's 10 or so neurons. If it's a tree, it's a very small tree. There is nothing to do with large ones.

Well, this is the eternal problem in data preparation, whatever NS you create, you still have to prepare and filter data.

I've written many times that if you feed a NS with rubbish at the input, you'll get rubbish at the output, like in the joke: mechanic's instruments! - the answer is: 15!... what 15!... what about the instruments? .....

there was a very smart neural networker on this forum@Andrey Dik, somewhere in his posts he told the truth about NS: if you have a good mathematical model, you don't really need NS... so here's the problem, if you give NS a set of data in the form of trade open prices: 1.35 Bai, 1.40 Sell, 1.32 Bai... then we get the result of NS response on the forward one like in the anecdote ... what instruments????

[Yuriy Asaulenko]

Dmitriy Skub:

Very true.)

In general, TC has a very good idea. And it all seems clear, but, from his own bell tower, move in no way possible. Imho, it will not be enough)).

[Yuriy Asaulenko]

Igor Makanu:

There was a very, very smart neural networker on this forum@Andrey Dik, somewhere in his posts he said...

Our opinions are diametrically opposed on this one. I've stopped reading his posts altogether. Not a stupid reader...

[Ivan Negreshniy]

mytarmailS:

Here you go, I think I got it done...

There aren't many trades because I set them up manually, but you can try

it is for 5 min chart

I generated your EA, although the template has too few signals and that is probably why it is not very nice, but I'm not afraid of copyright claims, so I'm safely posting the source code together with the testing template:)

[mytarmailS]

Ivan Negreshniy:

I generated your Expert Advisor, although the template has too few signals and that is probably why the EA is far from being perfect, but I'm not afraid of copyright claims and I'm also putting it in the source code together with the testing template:))

Yes, it's expected))

1) few deals

2) Possibly, the pattern itself is not standard and is not seen by the grid.

I explain, the rebound (where the arrow) is only the price of the pattern, but not the time of its appearance, that is not the pattern itself was formed earlier and the arrow is placed when the price reaches the price pattern

it is possible that the network cannot search for such structures due to its implementation

[Ivan Negreshniy]

mytarmailS:

Yes, expected))

1) Few trades

2) Maybe because the pattern is not standard, it is not seen by the grid.

Let me explain, the rebound (where the arrow) is only the price of the pattern, but not the time of its appearance, it is not the pattern itself was formed earlier and the arrow is put when the price reaches the price of the pattern

it is possible that the network cannot search for such structures due to its implementation

There should be more signals, that's for sure, but it's important that the input data is adequate for the task at hand.

I purely intuitively, from what you said about the maximum size of the pattern and seeing the pattern indicator, entered 6 stochastic values and 150 price bars into the training sequence.

Perhaps you, knowing the signs that define your level pattern, can suggest other features and / or formulas to calculate them, at the moment the EA uses these:

#define  CALC_IND_1(n) iStochastic(NULL,0,5,3,3,0,0,MODE_MAIN,n+1)
#define  CALC_IND_2(n) iStochastic(NULL,0,5,3,3,0,0,MODE_SIGNAL,n+1)
#define  CALC_IND_3(n) iStochastic(NULL,0,5,3,3,0,0,MODE_MAIN,n+2)
#define  CALC_IND_4(n) iStochastic(NULL,0,5,3,3,0,0,MODE_SIGNAL,n+2)
#define  CALC_IND_5(n) iStochastic(NULL,0,5,3,3,0,0,MODE_MAIN,n+3)
#define  CALC_IND_6(n) iStochastic(NULL,0,5,3,3,0,0,MODE_SIGNAL,n+3)
#define  CALC_X0(n)    iOpen(symbol,period,n+1)
#define  CALC_X1(n)    iHigh(symbol,period,n+1);
#define  CALC_X2(n)    iLow(symbol,period,n+1);
#define  CALC_X3(n)    iClose(symbol,period,n+1);
#define  CALC_BAR(x0,x1,x2,x3) ((x3-x2)>(x1-x3)?((2*x3-x2-x1)>(x1-x2)?1:0.5):((x3-x2)<(x1-x3)?((x1+x2-2*x3)>(x3-x2)?-1:-0.5):0))