Machine learning in trading: theory, models, practice and algo-trading - page 738
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
Michael's foundation is fine. A good example for those who are looking for a place to start.
He has some good indicators, I think even from the stock exchange, which is already better than the standard mashka and rsishka. Then he does not just shove all this stuff into the neural network but uses different tools to evaluate the importance of indicators and select the most suitable ones for prediction. Then the neuronka is trained, but not a simple one, but one with built-in overfit control. And this entire strategy is tested in the tester using walk-forward.
SanSanych has already described a similar scheme at the beginning of this thread, finally someone decided to work with it. I think that many traders would like to put more indicators in their neuronics and not think.
I would also pre-process indicators through vtreat or caret. And a more adequate neuronka with kfold. And I would take more examples for training, I agree with the other critics.
I know, brothers!!! I'll answer you all at once... Everyone dances to the extent of his education, but our dispute will be resolved by the end result. The result, so to speak. I'm winning so far, because the deposit is growing. I know my approach is one of the correct one.
By the way, I'm just vtreat and select important predictors + tricks with shuffling.
This will be discussed in detail in the article about BOO.
I know, brothers! I'll answer you all at once... Everyone dances to the extent of his education, but our dispute will be resolved by the end result. The result, so to speak. I'm winning so far, because the deposit is growing. I know my approach is one of the correct one.
By the way, it's vtreat and I'm selecting important predictors + tricks with shuffling.
This will be discussed in detail in the article about BO.
Above you wrote about mutual information. What does it have to do with it?
You wrote above about mutual information. What does it have to do with it?
The decisive criterion in choosing which model to use. It prepares the ditset. We train 10 times. We obtain 10 models with the same training results (as an example), where the number of training errors is the same for all models. Then we save the results of the models in the training section not in binary form, but in a dable. The result of polynomial calculation is initially considered in the dable......
Next, we count the VI of polynomials with respect to the output. Working models are those which have VI above the output entropy or more than 0.7 approximately....
Choose a model with high VI, but not higher than 0.95. Although I have not exactly checked it yet.... IMHO
As for increasing the training sample. Yes, it can be increased, but the quality of the network will decrease, which leads to a decrease in transactions. The network begins to work longer in time, but the number of "Don't know" states increases and it turns out that the network runs less often. With this approach it is necessary to train several models and run them in parallel.
Especially since the training period is now dictated to me by vtreat.R. Which with my input data offers me sets of inputs for different sample lengths. I choose the maximum learning period with the maximum number of input variables it chooses....And strangely enough, it spins from 20 to 50 samples. Apparently my inputs are like that. What can I do...
Mihail Marchukajtes:
...
The decisive criterion for choosing which model to use
...
I'll try to translate it into MO:
Although Mikhail's model is a classification model, it can return the result not only binary 0 and 1, but also the probability of belonging to a class (a fractional number from 0 to 1). Therefore, the model can be estimated not only with classification metrics, but also with regression metrics. I don't know why I chose mutual information, but it correlates well with MNC, so it's OK.
You can calculate it in R like this
library(entropy)
mi.plugin(rbind(target,prediction))
Here the smaller the result the better.
Of course, but in statistics error (variation) is proportional to 1/sqrt(N) where N is the number of observations, here everything depends on the initial variation, the more it is, the more observations are needed to fit error into 5% quantile
"Look for someone who benefits from it" ((c) Lenin)
Only useless or even harmful tools are put and sold in public, for obvious reasons, this trivial truth should always be kept in mind, algotrading is not a webdesign or some other franchise, it's a direct functional theft of money from the market, everything that is open and easily accessible, much less advertised, a priori can not be useful in this context, junk or scam, such an aggressive and competitive environment we have.
Well, yes, the 5 percentile of winnings and an infinite abundance of publicly available tools is Klondike.
And I recently had to make a custom neural network component, which by three orders of magnitude surpassed a smart DNN from TensorFlow.
And most importantly, there was no complicated math or advanced technology, rather a simple, true original solution.
Against this background, Michael's commitment to the author's development Reshetov, personally, I am not at all surprised.
That's what I mean, just that. If I take 40 samples through my infrastructure, I can't vouch for the result either, and 4000 points is very little. If of course we are talking about combat strategies, not article strategies.
I adhere to the following rule of thumb - an EA on new data will not work for more than 10% of bars trained. If I have trained on 4000 bars, it means that it will work on no more than 400. For example, if I want to trade a week on m15, I need at least 4800 (4*24*5*10, 10 weeks) bars of history for training.
Some day MT5 will introduce parallel threads in experts to be able to optimize the Expert Advisor in the background, then it will be possible to conduct interesting experiments with an extremely low number of bars and perhaps I will be able to do it with 40 bars. Although I probably won't be able to. Even if I have less than 1000 samples for training the result of neuronics skews in favor of trend and as soon as it changes - everything is bad.
As for increasing the training sample. Yes, it can be increased, but the quality of the network will decrease, which leads to a decrease in transactions. The network begins to work longer in time, but the number of "Don't know" states increases and it turns out that the network runs less often. With this approach it is necessary to train several models and run them in parallel.
Especially since the training period is now dictated to me by vtreat.R. Which with my input data offers me sets of inputs for different sample lengths. I choose the maximum learning period with the maximum number of input variables it chooses....And strangely enough, it spins from 20 to 50 samples. Apparently my inputs are like that. What can I do...
hmm... good afternoon!)
and the word "curvafitter" is something neural?