Machine learning in trading: theory, models, practice and algo-trading - page 2810

New comment
 
Maxim Dmitrievsky #:

There's a whole zoo of approaches, you can find implementations on github, I saw one for python

There may be many solutions, but what is the quality of these solutions and how well they work in real tasks?

 
mytarmailS #:
If Alexei wanted to think and listen, it would take two minutes with my script....

Interesting value judgement. You didn't hear what I wrote to you. Do you need an example?

 
Aleksey Vyazmikin #:

Interesting value judgement. You didn't hear what I wrote to you. You want an example?

yes, I read ... so many errors in thinking on not so many lines of text I have not seen for a long time ...

Give me an example.)

 
mytarmailS #:
Do you know what correlation is? You're talking such rubbish that I'm starting to shake.

Here's an example of what I'm talking about - a yellow trimmed sample and a whole green one - think about it.


 
Aleksey Vyazmikin #:

Here's an example of what I'm talking about - a yellow trimmed sample and a whole green one - ponder.

It still hasn't occurred to you that no one understands what your signs mean, maybe you need to signify the meaning of the doodles? ))
 
mytarmailS #:
It still hasn't occurred to you that no one understands what your signs mean, maybe you should signify the meaning of the doodles? ))

I think you should be more restrained. Leave the frustrations of life to a drinking buddy or a priest.

Obviously, Pn are predictors and K is the correlation coefficient. The example shows how the coefficient can change if the other 50% of the data changes its consistency. Your approach can only be applied to stationary data, with unchanging correlation, which is not the case in our case. So, I recommend you to think before you snicker.

 
Aleksey Vyazmikin #:

So, I recommend thinking before you snide.

About thinking....

The question is why to go through all variants of correlation of signs from 0,1...0,2...0,7...0,8 thereby increasing the already huge array of work for the algorithm?

Even generated random noise can correlate with the predictor by 0.6 and 0.8 , rarely more ...

It follows that it's stupid to take correlation threshold less than 0.8, but you decided to do a complete oversample, which shows that you don't think about what you're doing....

So I have the right to chuckle, I think so...

Aleksey Vyazmikin #:

Obviously, Pn are predictors and K is the correlation coefficient.

Obviously... ))))) this is a scribe.... What is obvious, I'm ashamed to ask? What decoding algorithm should I apply to make it obvious?

Do you even think what you say?

Aleksey Vyazmikin #:

Your approach can be applied only to stationary data with unchanging dependence, which is not the case here.

This is not my approach, but yours, and it was initially bastardised, and three people told you about it.

 
mytarmailS #:

About Thinking....

The question is why to search all variants of correlation of signs from 0,1...0,2...0,7...0,8 thus increasing the already huge array of work for the algorithm?

Even generated random noise can correlate with the predictor by 0.6 and 0.8 , rarely more ...

It follows that it is stupid to take a correlation threshold less than 0.8, but you decided to do a complete oversample, which shows that you do not think about what you are doing....

All I'm saying is that I'm experimenting and drawing conclusions.

mytarmailS #:

Obviously. )))))))) that's the beejeezus... What is the obviousness, I'm embarrassed to ask? What decoding algorithm should I apply to make it obvious?

From the context of the discussion it is obvious.

mytarmailS #:

It's not my approach, it's yours, and it was ghoulish to begin with, and you've been told that by three people.

What is your approach? What is my approach? I don't understand what you're talking about.

You, in particular, claimed that a model cannot be trained on a sample - you were wrong.

 
Maxim Dmitrievsky #:

memory is a NS with weights trained, you retrain it at each step, you move the weights a little bit... not much, so lag.

and you can't transfer that to the terminal.

I've tried it, it doesn't work, it's all down to the signs again.


If you are interested, I'm throwing the design of a multicurrency tester with spread, primitive lot and a hint of opening closing positions with fractional lot.

For the tester to work, you need to prepare a dataframe with ['open', 'spread] columns, and also throw a numpy array of format x (n,2) with forecasts of buy/sell probabilities for each new bar into signal. The tester works from a loop, below is an example of initialisation of using the tester

symbols = ['EURUSD','GBPUSD']
points = [1 / mt5.symbol_info(s).point for s in symbols]

trade_agent = Agent()
for s, p in zip(symbols, points):
        trade_agent.create_symbol(s, p,max_positions=6)

for index, row in dataset.iterrows():
        sig = signal[index]
        prices = list(row.filter(like='open'))
        spreads = list(row.filter(like='spread'))

        if index != len(dataset) - 1:
            trade_agent.refresh(prices, [sig], spreads)
        else:
            trade_agent.stop = True

trade logic and lot can be adjusted in the transcript_sig method of the Symbol object.

    def transcript_sig(self,sig):
        act = 0
        lot = 1
        revers = 0
        if len(sig)==4:
            pass
        else:
            act = sig

        if revers<=0.5:
            if np.argmax(act) == 0:
                return 'buy',lot
            if np.argmax(act) == 1:
                return 'sell',lot
        else:
            if np.argmax(act) == 0:
                return 'sell',lot
            if np.argmax(act) == 1:
                return 'buy',lot


        return self.last_action,lot


The results of the test are in the dictionary trade_history_data, for the overall test and trade_symbol_data of each symbol.

        self.trade_history_data = {'step': [],
                                   'acc_profit': [],
                                   'acc_balance': [],
                                   'acc_equity': [],
                                   'lots_open': [],
                                   'R2_balance': [0,0,0,0]}

there are lists, if anyone wants to optimise or change something - welcome)

Files:
Tester.py  12 kb
 

Вы, в частности, утверждали, что модель на выборке нельзя будет обучить - Вы ошиблись.

?? Where did I say that
1...280328042805280628072808280928102811281228132814281528162817...3512
New comment
Reason: