Machine learning in trading: theory, models, practice and algo-trading - page 1449

 
Mihail Marchukajtes:
So, what was the result on the white site???? Something I did not see. It's all very strange that it's so easy.

white site - from March 2017


 
elibrarius:

white site - from March 2017


And you believe that??? Surprisingly sure there's a catch somewhere, because if it's a reflection of reality, then why sell it for such ridiculous money. Makes sense? It makes sense and in this case I totally agree with cutting the white patch from mt with subsequent refreshing of quotes in mt, then we will see what this beast is worth. But the best thing is to trade in real account and in a couple of weeks we will be able to say anything. But it cannot be that simple
 
Mihail Marchukajtes:
And you believe it??? Surprisingly, I am sure there is a catch, because if this is a reflection of reality, why sell it for such ridiculous money? Logical? It makes sense and in this case I totally agree with cutting the white patch from mt with subsequent refreshing of quotes in mt, then we will see what this beast is worth. But the best thing is to trade in real account and in a couple of weeks we will be able to say anything. But it cannot be that simple

I don't. My practical experience has never been that good.

Let's see what Ivan Butko's signal will show

 
Mihail Marchukajtes:
Well, what is the result on the white area???? What I did not see. All this is very strange, it is so simple. I do not know if it is supported and not updated, but I could be wrong.
elibrarius:

I don't. My practical experience has never been that good.

Let's see what Ivan Butko's signal will show.

Of course it's a fake, especially on such slow timeframes.

I will not share revelations from Captain Hindsight here, but working with such "cats in a bag" is very dangerous when it comes to further trading of such models. All should be open, the data, the MO, the backtester, if there are closed modules it is a risk, there are many options how to organize fakes in the MO and in the indicators and backtester, this is if we talk about deliberate sabotage, but there are also errors or erroneous models. This is all good for selling, but toxic for personal use.

 

I know that many people played guessing the color of the bar, what was the result, in terms of Accuracy - on the fly I got something like 0.52.

Profit graph on the test sample is the best on the preliminary data


 
Aleksey Vyazmikin:

I know that many people played guessing the color of the bar, what was the result, in terms of Accuracy - on the fly I got something like 0.52.

If I look at the profit graph of the test sample I'll see the best one on preliminary data


The bar color is well defined in the articles by Vladimir Perervenko with the accuracy of about 0.8. And it is easy to repeat. I repeated it, including bare quotes, and it was 2-3% worse than with digital filters.
It is strange that your results are so bad. Although much depends on the instrument and the TF.
 
elibrarius:
The color of the bar in the articles of Vladimir Perervenko is very well defined with an accuracy of about 0.8. And it is easy to repeat. I repeated it, including bare quotes, and it was 2-3% worse than with digital filters.
It is strange that your results are so bad. Although, a lot depends on the symbol and the TF.

Can you point me to a specific article?

If it is true and the accuracy is so high, then why not build a strategy on this data? I have a mathematical expectation of just over 3 pips, but that's at 52%, and if they were 80%, it would clearly be a grail.

 
Aleksey Vyazmikin:

Can you point me to a specific article?

If this is true, and the accuracy is so high, then why not build a strategy on this data? I have a mathematical expectation of just over 3 points, but that's at 52%, and if they were 80%, it would clearly be a grail.

Here's the last one https://www.mql5.com/ru/articles/4722

            Sensitivity : 0.8494          
            Specificity : 0.8230          
         Pos Pred Value : 0.7921          
         Neg Pred Value : 0.8731          
             Prevalence : 0.4427          
         Detection Rate : 0.3760          
   Detection Prevalence : 0.4747          
      Balanced Accuracy : 0.8362          
                                          
       'Positive' Class : -1    
I believe that well guessed night candlestick directions give a small increase in balance, and the daytime incorrect guesses have a large size eats 2-5 night candlesticks.
Глубокие нейросети (Часть VIII). Повышение качества классификации bagging-ансамблей
Глубокие нейросети (Часть VIII). Повышение качества классификации bagging-ансамблей
  • www.mql5.com
В предыдущих двух статьях (1, 2) мы создавали ансамбль нейросетевых классификаторов ELM. Тогда мы говорили о том, как можно улучшить качество классификации. Среди многих возможных направлений я выбрал два: снизить влияние шумовых примеров и выбрать оптимальный порог, по которому непрерывные предсказания нейросетей ансамбля переводятся в метки...
 
elibrarius:

Well, here's the last one https://www.mql5.com/ru/articles/4722

I believe that well guessed night candle directions give a small increase in balance, while daytime misguesses having a large size eat up 2-5 night candles.

So he refers there to already prepared data, the way of preparation of which he described here, and there:

"

Now let's form two sets of data - data1 and data2. In the first set as predictors will be digital filters and their first differences, and as the target - the sign of change of the first difference of the ZigZag. In the second set the predictors will be first differences of High/Low/Close quotes and first differences of CO/HO/LO/HL quotes, and the target is the first difference of the ZigZag. The script is below and can be found in thePrepare.R file.

"

It doesn't sound like we're talking about bars...

And, I don't think forex is very convenient for MO, better go to Moex exchange.

 
Aleksey Vyazmikin:

I know that many people played guessing the color of the bar, what was the result, in terms of Accuracy - on the fly I got something like 0.52.

Profit graph of the test sample is the best on preliminary data


52% is not garbage, it's a real uterus and besides if you look at the hours it's not such a bad result, it's tradable with annual SR ~1

80% this is humor or just a market gimmick