Machine learning in trading: theory, models, practice and algo-trading - page 42
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Hello!
I have two questions for you
1) with regard to recurrent networks from the previous page, I threw the package "rnn" it really is some kind of strange and ill-conceived and switched to "RSNNS" took the network "Elman" threw the data all workedI really don't know where and how to set "how far the network remembers itself in the past" in "rnn", but it was immediately clear that each variable turned into a matrix with a number of columns that corresponded to the size of the network. Columns corresponded as if to network memory size, but I haven't found such parameter in"RSNNS", but it's impossible that it wasn't there as the essence of recurrent network is exactly in it.
2) What should I look for in"rminer"? If you mean something like "arima" then it is not suitable already discussed.
I've tried to write such a classifier target to make several step forecasts, but I failed to forecast it, the results were strange, in the first place the quality fell down (it's normal because the forecast is not on the 5th or 10th candle), anti-correlation was still present even if less expressed, but the trick is in the outrunning effect of the rminerIt's not clear why I got so many indicator reversals, but they were the same as market reversals...
My calculations will help me to understand the effect of the indicator and the market reversal.
1. Elman's network as well as Jordan's network remember only the previous step. To account for many previous steps, you need to connect RNNs in a chain, this is the so-called LSTM. This article describes quite lucidly about such networks. Unfortunately they are implemented in Python. But that is not a problem is it? Python and R are perfectly integrated.
2. There is lforecast function in rminer - Performs multi-step forecasts by iteratively using 1-ahead predictions as inputs . Speaking of multi-step predictions, are you referring to regression, of course?
Good luck
Conclusion: It is necessary to take each trait and to extract from it useful, I have ideas how to do it but before to sound it I would like to hear your thoughts, ideas and suggestions about it
Something can be deducted from a trait only on historical data. When a new bar comes, the sign must predict something and in order for it to predict, it must have predictive ability. Predictive ability is some potency of a trait, which is when some values of a trait predict one class and other values of a trait predict another class. I have already given an example of such predictive ability. Target: "men/women." Trait: "clothing." If a trait has only two values: pants/skirts, then in a Muslim society such a trait with these values unambiguously predicts class. But in non-Muslim society there is unisex clothing, in addition to a huge number of other names.
So the problem of determining the predictive ability of the attribute "clothing" for the target variable with two values of the class "male/female" is formulated as follows: what percentage of values of the attribute clothing will uniquely predict male and female? If this is a Western society and all clothing is "unisex," then the attribute "clothing" has no predictive ability. In a Muslim society, the trait "clothing" would have a very good predictive ability. If we make our example more real by introducing age...., then we will get a more real predictive ability. It will be concrete, and it is this predictive ability that will determine the prediction error.
That is, from the predictive ability of the predictor comes the prediction error, and if the chosen model fits the problem at hand, this error depends little on the choice of model.
From ideas and suggestions.
I have expressed them on this thread and on this forum many times. The main difficulty is that my point about "predictive ability" is not yet our understanding.
Of the tools I gave a link to an article andDr.Trader tried to apply, but not successfully. I attribute the negativity of its result to the specificity of its feature set: a large number of features that have little value. It is a very specific set of attributes for Forex. In Forex any attribute may have thousands of values, while he has dozens.
I have expressed them many times in this thread and on this forum. The main difficulty is that my point of view on the "predictive ability" is not yet our understanding.
Maybe because this point of view is not supported by either testing or testing results on your part? :)
And in general, why these kilometer-long enlightenment, the question was how to get the useful out of the features, and not how to select signs, these are different things, and here your reference to an article generally to place ...
Maybe because this point of view is not supported by either testing or testing results on your part? :)
And in general, why these kilometer-long encyclopedias, the question was how to get the useful out of the features, but not how to select signs, these are different things, and here your reference to an article generally to place ...
I have to write a kilometer-long libes. In a nutshell: a sign is a whole and nothing can be taken out of it. You can determine whether or not the whole trait fits.
PS.
I do a custom selection of traits that have predictive power. When using traits that are selected using my algorithm, I get models without retraining.
I have to write a kilometer-long libretto. In a nutshell: a trait is a whole, and you can't take anything out of it. You can determine whether or not the whole trait fits.
PS.
I perform a custom selection of features that have predictive power. When using the features selected according to my algorithm, I get models without retraining.
Wow, cool... Can you tell me any specific results of your untrained model? Or this topic will also be "skipped over" like with your robot, which you "seemingly have" and "seemingly earns"?
And in general, Sanych, can you stop it already?!?
Talking about something that does not really exist, everything became clear to me a long time ago...
This is not good, to put it mildly, to other participants who read you and then spend time on the road to nowhere
Wow, cool... Maybe you will tell us some specific results of your untrained model? Or are you going to "skip this topic" just like with your robot, which you "seem to have" and "seem to be making money"?
And in general, Sanych, can you stop it already?!?
Talking about something that does not really exist, everything became clear to me a long time ago...
This is not good, to put it mildly, with respect to other participants who read you and then spend time on the road to nowhere
Let's distract and think how a professional trader works(remember, I'm still exaggerating :)) who has only two signs - levels and "RSI" indicator And there is a trading system in which a sell trade sounds like this: if the level is broken through upwards and RSI is greater than 0.9, then the sale...
What is a trading system? Trading system in this case acts as a data filter, a filter that does not let the trader enter in noise, and the share of noise in this example with RSI is not a joke 95% because RSI range from -1 to 1, and the trader needs only what is >0.9 those 5% ...
This will work if we take a dozen indicators with a dozen different lags. But, which indicators to take and with what lags it should be determined. To begin with take a large set of them, select some of them by some rules and the totality of this final hundred predictors will give you a real chance to predict the future. You take a random forest model, feed it with data, and build a decision tree. For example if rsi[20] > 0.4, and ma(16)[20] > 1.2, etc. - then it's a buy. And if rsi < 0.1 then it is selling. Here you can see some examples of what the forest looks like: http://www.intuit.ru/studies/courses/6/6/lecture/174 . In general, the forest will give you the attributes with useful values and thresholds, just as you want for dozens of indicators at once.
There is a nuance that the forest tends to overlearn. If you give the forest all kinds of garbage along with useful predictors, then the forest will add it to its logic. And according to rule "garbage in leads to garbage out" any prediction of model built on garbage - on fronttest data will be random and useless. You should always consider this when selecting predictors, and do cross-validations to validate the model.
This is not good, to say the least, to other participants who read you and then waste time walking on the road to nowhere
I totally agree with what SanSanych has ever written here, I recommend you listen. I learned a lot from him, checked it out, wrote it down.