Machine learning in trading: theory, models, practice and algo-trading - page 2006

New comment
 
Farkhat Guzairov:
Max, do you observe the same conditions in the demo as in the test?

I have the same exact conditions in the demo, but the spread is bigger in the demo. There is a pitfall somewhere

 
I find it hard to believe that in the test you get a perfect result on the data of the test period, personally I have not managed to achieve it and it can not be otherwise
 
Farkhat Guzairov:
I find it hard to believe that you will get a perfect result on the test period data, I personally did not succeed

There is no test / train there. It just trades in the tester (good). Betting on the real - bad.

 
I did not even bother to check the bot on the test period, after training sent it into battle, and then once a day ran it in the tester and checked the results obtained in combat, when the results coincided, I realized that now you can switch to cosmetics ))))
 
Maxim Dmitrievsky:

there is no test\train. He just trades in the tester (good). Betting on the real - bad.

Maxim, you wanted to post your code. Put it on your blog, as the founder of this thread did. Maybe someone will want to look into it at some point. But it will get lost here...
 
Maxim Dmitrievsky:
So, did you figure out Python? It doesn't work in Real, I don't know what's the problem. Or can I throw the source code here? Maybe someone can poke around some more.

Go to

 
Maxim Dmitrievsky:

There are 2 programs in the archive. The tester is just for system tuning, i.e. it tests the neural network on historical data and displays the result as a balance graph (in the form of accumulated amount of pips).

The trader trades on the account, pre-starting the tester. By changing the seed, you can achieve the best results of the tester. Also, by changing the configuration of the NS.

Problem description: Results of the tester and trader do not coincide, the trader does not show profit for some reason. The error is either in the trader or in the tester. Or there is some unobvious bug/pitfall related to the NS library. Haven't managed to catch it (no time).

The library itself with NS. You can use it for personal use. If you find a bug, contact me (there are improvement options if the TC starts working)

What you can check:

  1. the correctness of the tester
  2. compare the logic of the tester with the trader, find a discrepancy
  3. Figure out the particularities of the TC, think about why there might be a difference. Perhaps due to some randomization, the influence of seed in the real trading (this is a suspicion)
By changing the Seed can you get a draining result? It is possible that it is one of the parameters for fitting.
 
elibrarius:
By changing Seed can you get a draining result? It is possible that it is one of the parameters for fitting.

You can't, but you can improve the curve

At least I didn't get plum
 



Preparation of the data in training and work is different line

prices = prices.reindex(pd.date_range(start, end, freq='15min')).dropna()

I don't have python and I have no experience with it, I don't know what these commands are. So I'll ask.

Assumption - maybe the result will be inverted data relative to what we get in the real trade?

 
elibrarius:



Preparation of the data in training and work is different line

prices = prices.reindex(pd.date_range(start, end, freq='15min')).dropna()

I don't have python and I have no experience with it, I don't know what these commands are. So I'll ask.

May be the result will be inverted with respect to what we get in real trading?

The reindex of bars by date time is implemented here because there may be some missed bars in the history to avoid holes. Then the empty values are thrown out and then the detrend by the MA.

There are no omissions in the trader, since the last n bars are taken. They should not be reversed.

I don't think it would affect anything. But we can redo it, see... thanks

1...199920002001200220032004200520062007200820092010201120122013...3497
New comment
Reason: