Machine learning in trading: theory, models, practice and algo-trading - page 891

New comment
 
elibrarius:
31 category... no it's more of a 31 step discretization. One of Vladimir's articles uses this and the result is just as good.

Only if the original data itself is good....

 
Mihail Marchukajtes:

First, record the moment when you make a decision. Let it be an event. Then exactly at this moment when the event occurred save the values of indicators.

To be honest I do not quite understand your table. What is in it?

So what is the point of saving the exact value, but not a pattern - for example the well-known RSI, as I have, if it opened above level 70, then 1, below level 30, then -1, but not both, then 0.

In the table there is a set of rules with high score, which were formed based on the results of the tree. I use a program Deductor Studio, which is pretty good for beginners, which I think I am, but the free version is difficult to get the results of analysis, so I filtered the rules and chose the best ones manually.

Today I automated the process and noticed the phenomenon that a rule might not be complete, if after filtering a sample a rule gives only one predictor value (i.e. if, for example, the predictor has 2 positions, and after filtering (applying a rule part) I got one position), resulting in such a weird set, an example on two rules:

with certain rules - green color, any value - yellow color, not defined - red color.

The uncertainty comes in when after filtering by rules we don't have alternative variable values, but in other conditions they can appear and the rule will be executed not because it is proven correct, but because it hasn't encountered the situation before and couldn't evaluate it and will consider it similar although it could be quite the opposite.

 
Aleksey Vyazmikin:

So what is the point of keeping the exact value, and not a certain pattern - for example, the well-known RSI, as I have, if you open above the level of 70, then 1, below 30, then -1, but not both, then 0.

In the table there is a set of rules with high score, which were formed based on the results of the tree. I use a program Deductor Studio, which is pretty good for beginners, which I think I am, but the free version is difficult to get the results of analysis, so I filtered the rules and chose the best ones manually.

Today I automated the process and noticed the phenomenon that a rule might not be complete, if after filtering a sample a rule gives only one predictor value (i.e. if, for example, the predictor has 2 positions, and after filtering (applying a rule part) I got one position), resulting in such a weird set, an example on two rules:

with certain rules - green color, any value - yellow color, not defined - red color.

Uncertainty just stinks when after filtering by rules we have no alternative values of variables, but in other conditions they may appear and the rule will be executed not because it is verified correct, but because it has not encountered this situation before and could not evaluate it and will consider it similar, although it may be quite the opposite.

To answer your main question. When the PSI is over 70, you only have one value of 1 (one), but it is very important for NS to know how much higher than 70 the PSI shows. If it's 70.05, that's one thing, and if it's 73, that's another thing, let alone 90, 80 or any number between 70 and 100. It is in these little things that the model tries to cling to the IMPORTANT...!!! So that's it. And by assigning strictly units you deprive the network of important information for her, IMHO, of course!!!

 

You know, everyone here is not stupid people, I would even say very smart and logical in their reasoning. Programming requires logic. But here's the thing.... When you start to link TS you rely solely on your logical conclusions and all conclusions you draw from your experience and the phrase "Well, it's logical", what you need to do so-and-so. But the tests show the invalidity of your conclusions and you do not understand why it does not work. After all, you have done everything logically correct. But doing something logically does not mean it is right. Thus many become a hostage to their logic, which is naturally correct, but why it does not work...... And commit the same error without even knowing that you do it over and over again based on your logical conclusions.

As an example I have already given: Logically, the bigger the model, the smarter it is...

In practice: Medium models work better than large or small ones. This is as an example...

You thought that assigning 1 if the RSI is above 70 is right, and for a network it's not. And you're like, what about the grid I made your job easier, how can you not understand? Why are you so dumb? But unfortunately she will not be able to answer you that it is not enough. That with your unit you have stupidly killed all the important information....It is also as an example...

And in general the field of MO is alogical. Here logical conclusions do not work well. Here allogical conclusions go off without a hitch!!!!

 
Mihail Marchukajtes:

To answer your basic question. When the PSI is greater than 70, you only have one value of 1 (one), but it is very important for the NS to know how much higher than 70 the PSI shows. If it's 70.05, that's one thing, and if it's 73, that's another thing, let alone 90, 80 or any number between 70 and 100. It is in these little things that the model tries to cling to the IMPORTANT...!!! So that's it. And by assigning strictly units you deprive the network of important information for it, IMHO, of course!!!

You may be right, but I proceed from these considerations, if the system without NS works by such rules successfully, what prevents to understand and find these rules of NS? Yesterday I posted screenshots for comparison. So far my task is that the result of machine learning would be better than optimization and hard enough logic in TS.

 
Aleksey Vyazmikin:

You may be right, but I proceed from these considerations, if the system without NS works according to such rules successfully, then what prevents to understand and find these rules of NS? Yesterday I posted screenshots for comparison. So far, my goal is that the result of machine learning would be better than the optimization and hard enough logic prescribed in the TS.

Throw me the link where you posted the screenshots... I'll have a look at.... And about the rules, you're giving them to her wrong. You think that you feed her right, but in fact she starves for lack of information. You can't explain to her exactly what you want, so she gives you the result as she understands it. Do you want me to help you with your TS. Only within the guidelines ....

First tell me what basic TS you use?

 
Mihail Marchukajtes:

Throw me the link where you posted the screenshots... let me see.... And about the rules, you're giving them to her wrong. You think you're feeding her right, but she's actually starving for lack of information. You can't explain to her exactly what you want, so she gives you the result as she understands it. Do you want me to help you with your TS. Only within the guidelines ....

First tell me what basic TS you use.

Here's the linkhttps://www.mql5.com/ru/forum/86386/page889#comment_7378517

About the hunger, I'm just now thinking that I'm overfeeding the data, since in fact they don't all go to the rules...

Interesting idea that I can't explain what I want, i.e. it's about the wrong targets?

And TC, so the basic simple one is turtles on minutes. And the rest is done by filter cascades based on my observations of the market.

 
Aleksey Vyazmikin:

Here's the linkhttps://www.mql5.com/ru/forum/86386/page889#comment_7378517

About the hunger, I'm just now thinking that I'm overfeeding the data, since in fact they don't all go to the rules...

Interesting idea that I can't explain what I want, i.e., it's about the wrong targets?

And TC, so the basic simple one is turtles on minutes. And the rest is done by filter cascades built from my observations of the market.

How many signals per day does the turtle on minutes give????

The basic TS is needed only for the moment of making a decision. There is a signal on the basic TS. All we start to analyze the market. If you do not know the basic TS, do you have any relation to it?

 
Mihail Marchukajtes:

Turtle on minutes how many signals per day gives????

The basic TS is needed only for the moment of making a decision. A signal is formed by the basic TS. Everything begins to analyze the market. Screenshots of the signals from the basic TS.

My Expert Advisor is improving all the time so I have no statistics - I have 8 signals on 8th May and 5 losing signals, but I closed the day with profit. I have noticed for myself that the evening before the holidays is better not to trade - I will check on the history - in the evening there was 1 stupid entry, which could have been avoided, but which bit 20% of the daily income.

Well, my trading system is not structured in a standard way, i.e. decision to enter the market is generated constantly, depending on the price position in Doncian channel, and then the filters start working and as a result they provide decision to open position or not. In general, ideally, I need to find an external factor that can enable/disable them at different moments of the market. I have just started working with NS earlier this week and I've created environment for my work so far. I haven't investigated the selection of predictors so far, but I've already found interesting patterns that I use in my EA.

 

That's pretty much all the basic conditions for the input

      double HighD=Donchianf(0,0);
      double LowD=Donchianf(1,0);
      double PriceProc=0.0;

      if(HighD-LowD>0) PriceProc=((Open(0)-LowD)/(HighD-LowD))*100.0;
      else
        {
         SellNow=false;
         BuyNow=false;
        }

      priceBuy   =Open(0);
      priceSell  =Open(0);


      if(CountMarketOrder_OS==0 && CountMarketOrder_OB==0 && HighD>0 && LowD>0)
        {
         if(PriceProc<=ProcOpenBuy && PriceProc>50.0)BuyNow=true;
         if(PriceProc>=ProcOpenSell && PriceProc<50.0)SellNow=true;
        }
1...884885886887888889890891892893894895896897898...3478
New comment
Reason: