Machine learning in trading: theory, models, practice and algo-trading - page 936
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
I see. The tree couldn't learn how to filter properly, so the result was not noticeably better with the filter, just less deals. Basically filtered out randomly some of the good deals and some of the bad deals,
I trained the tree on 2015 only for malovhodov.
I used filter_02 and mnogovhodov_02 in 2016, it would be better to compare 2016 and 2017 in the tester (2017 - all new data, which was not in the archive, that's the most interesting to see).
Yeah, and I thought the 2015 was a learning curve, then things are like this - blue with no filter and green with a filter
I have to say that 2015 is more of an up trend, 2016 is a down trend, and 2017 is almost sideways on the days. I.e. the entities of the three are somewhat different, and I think global trends play a role.
Also, my buy entry is generated from 5 to 9 by the arr_DonProc predictor - so part of the tree is automatically cut.
But in general the result is not bad, do you think?
Further branching in my case led to an overfit. For better accuracy, we should move to more complex models - forest or neuronics.
It is possible to branch to 100% accuracy on training data, but what is the point in it if such a tree will only fail on new data. We need to train such a model, which would be able to show almost the same result on new data as on training.
Up to 100% is possible, with different sets of predictors, but obviously we haven't used all the potential here.
By the way, I think we should provide more information about the past - now we can get it from Regressor and iDelta, and a couple more predictors, but there is no such thing as the number of bullish and bearish candles in a row - their relation to each other - it may be useful, too.
What does this have to do with a question of faith? I see squiggles on the graph - and I don't understand how to interpret them - that's it.
The random forest is calculated on every tick. If the results are gathered by bars, like a normal price flow, then you get such a graph. Interpretation is needed when there is a formula, and here is just the result of the forest for clarity.
Then I can only reply to the screenshot "Interesting picture! For it is not clear what they wanted to show, everyone, if the essence is clear only to you ...
The forest error percentage is calculated for a certain period of time. And on the graph you can see the difference between reality and the data from the forest in a particular minute (I have M5 there).
Of course, the graph from another forest will be completely different from mine.
The forest error percentage is calculated for a certain period of time. And on the graph you can see the divergence of reality with the data from the forest in a particular minute (I have M5 there).
Of course, a graph from another forest will be completely different from mine.
Now it is clearer, but it is not clear what is predicted on each tick - the next tick?
How in reality you will calculate at every tick - only OpenCL with a top graphics card will probably help here.
I only gave an example of my forest. And I didn't ask to deal with the results of my model.
If you want a recommendation, what exactly is wrong in your model, then instead of strange tables show the results on the real price chart.
And the overall result is not bad, what do you think?
2017 is on the plus side, so that's a little encouraging.
I will try one more time. I took file mnogovhodov_02 and made a new targeting:
"1" class where arr_Buy = 1
"-1" where arr_Sell = -1
"0" for the other cases
For your strategy this targeting seems more appropriate.
I only gave an example of my forest. And I did not ask to deal with the results of my model.
If you want a recommendation, what exactly is wrong with your model, then instead of strange tables show already the result on the real price chart.
While there is no model, as such, in search. The table shows the change in the results, while more is not required - it is moving, it means it is alive.
2017 on the plus side, that's a little gratifying.
I'll give it one more try. I took file mnogovhodov_02, made a new target:
"1" class where arr_Buy = 1
"-1" where arr_Sell = -1
"0" for the other cases
For your strategy this targeting seems more appropriate.
Does that mean you can build more than 3 target tree outputs?