Machine learning in trading: theory, models, practice and algo-trading - page 443

New comment
 
Albinochka6:

I tried to read a book on machine learning, not even to master it, but just to start reading - I could not, it was very complicated and I understand that the topic is really good to study, probably not enough life.

And try to start with articles on neural networks here on the site, just type neural networks into the search. They are difficult to understand without practical examples. I learned it this way - theory, examples, theory, examples, theory, examples. For example, you can just read how a perseptron is built, and then by analogy it will be easier.
 
Mihail Marchukajtes:

Today is not a good day.... If TC is really wrong, she's doing it in a big way. It's a kind of payment for extra profitability. But I would not want to start from this very place :-(.

Either that is the answer, that the TS has stopped working, do another one. BUT for a week I'll still see..... Either it's a temporary hiccup that will more than pay off later. It is important what happens next, right? :-)


Well your ns is retraining means adjusting to the story. It takes a lot of testing to understand whether it will earn or not :)

Mine was 30+% the week before last and everything was perfect. Last week it was -7%, volatility was absurd and there was no directional movement on nonfarm. On Monday I will re-train again :) I train in 2 months on 15-minute symbols.

I have found the basic one - 3-month market cycles. That's why I teach us in more than 2 months. The market changes approximately every 3 months and it ceases to work, it is clearly visible in the tester.

It is also useful to visualize the state of all predictors during trading. For example, I found one trend - in flat conditions signals of NS often change from buy to sell and vice versa, while it works almost perfectly at trend periods. So, it would be good to introduce some additional conditions.

I have not yet managed to find an Ns that would provide good profit within a year. The conclusion is obvious - the quarterly cycles. There are ideas now how to overcome this, we'll see.

 
Maxim Dmitrievsky:

Is that a forward in the video?

 
Grail:

Is that a forward in the video?

Half back half forward

If you change medium-term cycle, it stops earning anyway, even if you do forward or no forward. I just need more frequent retraining. How to make it work without re-training - that's a problem :) Scalper system, does not learn on global trends.

 
Maxim Dmitrievsky:

The conclusion is obvious - the quarter cycles.


If we take log(price_0/price_-1) then ACF of this increment shows cycles very well, if they exist. The problem is that the period is variable, not quarterly.

 
SanSanych Fomenko:

If we take log(price_0/price_-1), the ACF of this increment shows cycles very well, if they exist. The problem is that the period is variable, not quarterly.


Yes, it's variable, i.e. approximately 3 months (well, contracts of futures expire + seasonality), plus I can't guess when these cycles will meet... I wanted to experiment with Hearst's cycles, but they are very complex: what to screw up and how to do it all)

Plus yes, the autocorrelation is also interesting to try.

 
Vizard_:

http://radikal.ru/video/Rv6hGGmBVSZ


It's no big deal, we're not in the dark either, and it's no problem to make a model of this quality. For example. Starting with 05.01 OOS.

* Sensitivity of generalization abiliy: 93.54838709677419%
* Specificity of generalization ability: 96.55172413793103%
* Generalization ability: 95.0%
* TruePositives: 58
* FalsePositives: 4
* TrueNegatives: 56
* FalseNegatives: 2
* Total patterns in out of samples with statistics: 120

And here is the work from 05.01 to 06.19 (change of liquidity) Parameters is not so good of course, but this TS earns ONLY on 0.00100 points better than the signal on pending orders.


 
Maxim Dmitrievsky:

Well, your ns is retraining, it means adjusting to the story. It takes a lot of testing to see if it will make money or not :)

Mine was 30+% the week before last, everything was perfect. Last week it was -7%, volatility was absurd and there was no directional movement on nonfarm. I will re-train again on Monday :) I train in 2 months on 15-minute symbols.

I have found the basic one - 3-month market cycles. That's why I teach us in more than 2 months. The market changes approximately every 3 months and it stops working, it is clearly visible in the tester.

It is also useful to visualize the state of all predictors during trading. For example, I found one trend - in flat conditions signals of NS often change from buy to sell and vice versa, while it works almost perfectly at trend periods. So it would be nice to introduce some additional conditions.

Ns that would give good profits in a year, when taught in a couple of months has not yet worked out. The conclusion is obvious - quarterly cycles. There are ideas now how to overcome this, we'll see.


Most likely the futures contract is traded for 3 months and then the liquidity flows to the next contract. In the process of flowing over, the rules of the game most likely change....

 
Maxim Dmitrievsky:


I wanted to experiment with Hearst cycles, but they are very complex: what to screw in and how to make it work )

There is no problem with Hearst at all.

The package rugarch::ARFIMA. Fractional differentiation is Hearst. Extremely rare. You can spit. You don't need fractional differentiation at all for the specified increment.

 
Mihail Marchukajtes:

Most likely a futures contract trades for 3 months, then the liquidity flows to the next contract. In the process of overflowing, most likely, the rules of the game change....

That's right, the futures are traded for the last 3 months - just before and after the expiration of the previous one . It is useless to look earlier and try to do something with it.

Maxim (or his system) noticed the obvious.

1...436437438439440441442443444445446447448449450...3498
New comment
Reason: