Machine learning in trading: theory, models, practice and algo-trading - page 77
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
Very good score for R, given that all other languages are used for a lot of different purposes, and R is only for data manipulation.
Have you tried training on my data?
I ran the experiment for 5 days. I'm only torturing EURUSD!
Made 99 training samples of about the size I posted here. All of them are unique because they contain time-thinned observations. I'm training models on them and will make a committee from GBM forests. Already have interesting results on validation, although I only passed 18 samples in 24 hours. At the same time I did 2.5 times as much validation and I'll show later why!
As you can see, in the very first training sample the model trained so that for 400 K observations in validation I got MO 2.7 points (taking into account the spread).
Well, there are still 1, 1.5 points results, okay. All this will be stacked in the committee then and see what the picture will finally be on the validation! I anticipate a grade of 4 or 4 + for homework.
Go for it, gentlemen.
PS: training data - 99 samples for diversification and committee: https://drive.google.com/open?id=0B_Au3ANgcG7CNG5JT3R2UzI0UzQ
Yes, I tried it, it did not work the first time. To estimate the model I took the accuracy of the regression result (the required result scaled in [0;0.5;1] minus the obtained result, modulo). Genetics has found such a global maximum with a pair of predictors, at which the neuronke does not have enough data for training, so it returns 0.5 on almost all data (corresponding to the "do not trade" class). In general such a fitness function is not suitable at all, the network does not trade at all.
Now the second experiment is still in the learning process. I still use 3 classes but I immediately round the regression result to the [0;0.5;1] levels. Assessment - accuracy of classification ([number of correct answers] / [total number]). From the graph of the best fitness values in genetics I can judge that the result will be close to 33%, essentially the same as random. Maybe a little higher if you're lucky. I need to wait another day or two until the genetics will not reach its maximum, then I can fronttest.
On my predictors I usually have better results, I think you need to add more indicators to this data. If your algorithm can evaluate and discard predictors, the more indicators you add initially the better.
Yes, I tried it, it did not work the first time. To estimate the model I took the accuracy of the regression result (the required result scaled in [0;0.5;1] minus the obtained result, modulo). Genetics has found such a global maximum with a pair of predictors, at which the neuronke does not have enough data for training, so it returns 0.5 on almost all data (corresponding to "do not trade" class). In general such a fitness function is not suitable at all, the network does not trade at all.
Now the second experiment is still in the learning process. I still use 3 classes but I immediately round the regression result to the [0;0.5;1] levels. Assessment - accuracy of classification ([number of correct answers] / [total number]). From the graph of the best fitness values in genetics I can judge that the result will be close to 33%, essentially the same as random. Maybe a little higher if you're lucky. I need to wait another day or two until the genetics will not reach its maximum, then I can fronttest.
On my predictors I usually have better results, I think you need to add more indicators to this data. If your algorithm can evaluate and discard predictors, the more indicators you initially add the better.
Still trying to build the grail???? Come on......
Still trying to build a grail???? Come on......
No, no guys, the system is gone, there is no point in showing it. I apologize..... Concerning the treatise on classification, it is yes ... I can ... but time is not present yet, when I will be more free and creativity will come to me, I will write about it..... But otherwise.....
Interesting... You wrote, that you successfully use your algorithm for a year? A few hours ago you wrote "Still trying to build a grail???? Come on......" which implies that the algorithm was working a couple of hours ago and then the algorithm went down???????????????????? your story doesn't make any sense at all....
Listen, if you have something to tell me, then send me a message, because I did the same thing you wrote about, I'm interested in this topic
There were posts in the thread about RNeat, which is a neuron with adaptive topology, where weights and connections of neurons are formed using genetics.
Judging by the description and simple tests, the model is quite good. But it will be pretty weak for forex. For more than a week I was running training, predicting to buy/sell on the next bar with 400 predictors.
Learning model is too slow, in a week model's fitness almost grew to only 2%. Prediction accuracy even on training data is only 55%, the model has not had time to evolve to good logic to start generating profits.
Accuracy on validation data (random bars taken from training data) is a little better, but it's more by chance than the model's merit.
Fronttest accuracy fluctuates between 50% and 53% and judging by the chart it is also accidental but not due to the model.
Miracle did not happen, I guess the model will de-evolve to the desired logic during months of work, but it will retrain with bad results on validation and by that time will be out of date and I will have to start all over again.
I'm stopping the experiment; I don't see the point in continuing.