Machine learning in trading: theory, models, practice and algo-trading - page 112
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
If that's the case, I'm glad, it's much better.
Anyway, fronttest shows much better result. I split your file into 2 parts (without shuffling, just in order), the first part has 50 lines, the second 19. So jPrediction has no access to examples from the second file, and it will be really new data for the model.
In the end, JPrediction gave the answer in only 9 cases in the second file. Right in 5 cases, wrong in 4. The accuracy is about 50%, there is nothing good in this result.
All is well except for a trifle: there is no comparison with other models.
I offer my services in comparison
1. You prepare an input Excel file containing predictors and target variable
2. You do the calculations
3. You send the input file to me.
4. I do the calculations using randomforest, ada, SVM
Comparison.
What are you going to compare?
Randomforest, Ada and SVM are binary classifiers, while jPrediction is ternary.
Do you have experience and adequate criterion to compare binary classifiers with ternary classifiers?
If you feel like it, jPrediction is in the public domain, including the source code. I've already posted instructions to it. There models are created trivially (just don't tell anyone about it): by pressing just one key - F8. You can play with numbers at your leisure with a clever face. But without my participation. After all, to press one key is unlikely to need my help, because I hope very much that you parents have already allowed to use the computer yourself?
But only without my involvement.
Pity.
Good luck, cyclist.
What are you going to compare?
Randomforest, Ada and SVM are binary classifiers, while jPrediction is ternary.
Do you have experience and an adequate criterion for comparing binary classifiers with ternary classifiers?
If you're so eager, jPrediction is in the public domain, including the source code. I've already posted instructions to it. There models are created trivially (just don't tell anyone about it): by pressing just one key - F8. You can play with numbers at your leisure with a clever face. But without my participation. After all, to press one key, my help is unlikely to be needed, because I hope very much, that you parents have already allowed to use the computer yourself?
If you want an unambiguous answer without looking at the data and algorithms....
Thank you for your comprehensive and understandable answer.
I am talking about the same thing. They all can't understand that NS training is an art, and the choice of the right model comes from many years of experience and a certain feeling that this is it. But no, our corefiends want unambiguous results. By the way, about definiteness. It's quite an interesting thing when the model will be trained unambiguously and unequivocally. And here is what idea I have. I'll write to LS....
We are not coryphaei - we join the majority, whose members want to get the same results in the future as they do today. That is what this thread is about.
And the artists, the priests of art, they are with the mash-ups.
We are not coryphaei - we join the majority, whose members want to get the same results in the future as they do today. That is what this thread is about.
And the artists, the priests of art, they are with the mash-ups.
I'm going to show you the script in which I created an indicator with self-tuning of the period and how I trained the model,
The script is divided into two parts, the first creates a sheet with the objective periods and the data, the second script opens this created sheet and trains the model...
I am not a programmer, if anyone is interested, I will answer the questions,
1) all the paths in the scripts must be replaced by your own paths...
2) Do not try to check the model in the usual way by comparing the target with the new data with predict, this approach does not work, you need to put the data into a program of technical analysis and there simulate trading with stops, so far I am deeply convinced that this is the only objective way to test the model
3) I trained the model about 15 times of which only 3 or 4 times the model showed a small loss on the new data, I consider it in quotes "stable" result and worthy of attention and further development.
In total 50 000 data were used for training on 20 000 and 30 000 was a test on new data
4) the quotes are the same as mine you can download from finam websitehttps://www.finam.ru/profile/mosbirzha-fyuchersy/rts/export/?market=14&em=17455&code=SPFB.RTS&apply=0&df=18&mf=7&yf=2016&from=18.08.2016&dt=18&mt=7&yt=2016&to=18.08.2016&p=7&f=SPFB.RTS_160818_160818&e=.txt&cn=SPFB.RTS&dtf=1&tmf=1&MSOR=1&mstime=on&mstimever=1&sep=1&sep2=1&datf=1&at=1
Because of the awesomeness of this forum I cannot attach my quotes
as it lookshttp://prntscr.com/c776d3
Or you can paste your quotes from metatrader by tweaking the code a bit.
In 111 it looks like a date is created.
222 is already training a model on this date
I kind of figured out why Michael's example might work. The strength of jPrediction is not in the classification (50% accuracy on new data is like flipping a coin), but in the fact that this program does not make a prediction at all if the models in the committee give different answers. Michael's strategy works only because he doesn't trade in such obscure market moments when both models can't come to the same answer. It doesn't matter which prediction jPrediction makes, there will be 50% error anyway, the main thing is not to trade when jPrediction refuses to make a prediction. It then chooses the direction of the trade (buy/sell) according to the sequent. It turns out that jPredictor works as a detector of too chaotic situation in the market when it is better not to trade, a very interesting application.
I tried a similar approach with my data. I have more than 2 models in my committee. I couldn't get a unanimous answer from them, but if I trade when at least 80% of models agree with the answer I get better trading results, I advise to try.
I kind of figured out why Michael's example might work. The strength of jPrediction is not in the classification (50% accuracy on new data is like flipping a coin), but in the fact that this program does not make a prediction at all if the models in the committee give different answers. Michael's strategy works only because he doesn't trade in such unclear market moments when both models can't come to the same answer. It doesn't matter which prediction jPrediction makes, there will be 50% error anyway, the main thing is not to trade when jPrediction refuses to make a prediction. It then chooses the direction of the trade (buy/sell) according to the sequent. It turns out that jPredictor works as a detector of too chaotic situation in the market when it is better not to trade, a very interesting application.
I tried a similar approach with my data. I have more than 2 models in my committee. I couldn't get a unanimous answer from them but if I trade when at least 80% of models agree with the answer I get better trading results.