In one sentence, what is your edge for 2019?

[ph1l]

How have the results been for both models to date?

The k-nearest neighbor strategy for 02/15/2019 through 03/19/2019 gained 1.12 percent (traded with real money and includes commissions which were zero and all fees).

 

[Visaria]

The federal reserve bank of the usa is my edge.

 

[userque]

The k-nearest neighbor strategy for 02/15/2019 through 03/19/2019 gained 1.12 percent (traded with real money and includes commissions which were zero and all fees).

Thanks. Do you have any additional ideas to test out with your models, or are they pretty much complete, or something else?

 

[ph1l]

Thanks. Do you have any additional ideas to test out with your models, or are they pretty much complete, or something else?

I was thinking about trying a multi-dimensional or not strictly time series inputs to my k-nearest neighbor strategy. Code for a fast dynamic time warping algorithm from
https://github.com/melode11/FastDTW-x
supports finding similar time series that have multiple dimensions. So I might look at how that feature works and try it.



Another application I developed was a recurrent neural network using the minimal gated unit architecture as described in
https://arxiv.org/pdf/1603.09420.pdf
and
https://arxiv.org/ftp/arxiv/papers/1701/1701.03452.pdf

This generated rules for potential trade entries like my genetic programming strategy. The rules weren't human-readable because they were sets of floating point weights used inside the neural network. And it took a long time to train because it was doing more calculations than the genetic programming strategy. The rules seemed to work outside of the training data, so I might try to work with this again with the same data as my genetic programming strategy and/or k-nearest neighbor strategy.

 

[userque]

I was thinking about trying a multi-dimensional or not strictly time series inputs to my k-nearest neighbor strategy. Code for a fast dynamic time warping algorithm from
https://github.com/melode11/FastDTW-x
supports finding similar time series that have multiple dimensions. So I might look at how that feature works and try it.



Another application I developed was a recurrent neural network using the minimal gated unit architecture as described in
https://arxiv.org/pdf/1603.09420.pdf
and
https://arxiv.org/ftp/arxiv/papers/1701/1701.03452.pdf

This generated rules for potential trade entries like my genetic programming strategy. The rules weren't human-readable because they were sets of floating point weights used inside the neural network. And it took a long time to train because it was doing more calculations than the genetic programming strategy. The rules seemed to work outside of the training data, so I might try to work with this again with the same data as my genetic programming strategy and/or k-nearest neighbor strategy.

Ok. Thanks for the links. Sounds like you'll be busy for a while.

 

[ironchef]

I was thinking about trying a multi-dimensional or not strictly time series inputs to my k-nearest neighbor strategy. Code for a fast dynamic time warping algorithm from
https://github.com/melode11/FastDTW-x
supports finding similar time series that have multiple dimensions. So I might look at how that feature works and try it.



Another application I developed was a recurrent neural network using the minimal gated unit architecture as described in
https://arxiv.org/pdf/1603.09420.pdf
and
https://arxiv.org/ftp/arxiv/papers/1701/1701.03452.pdf

This generated rules for potential trade entries like my genetic programming strategy. The rules weren't human-readable because they were sets of floating point weights used inside the neural network. And it took a long time to train because it was doing more calculations than the genetic programming strategy. The rules seemed to work outside of the training data, so I might try to work with this again with the same data as my genetic programming strategy and/or k-nearest neighbor strategy.

OK I have to go back to Coursera to sign up on how to use neural network to trade, after I complete my introductory course on how to use Excel and how to code VBA. It will take me 20 more years. In the meantime I just have to rely on my poor overwork brain.

 

[ph1l]

Ok. Thanks for the links. Sounds like you'll be busy for a while.

In recent testing of recurrent neural networks, I concluded they are less accurate for future predictions than genetic programming. I'm guessing this is because a rule from a recurrent neural network applies operations on all the current input indicators while a rule from genetic programming applies operations on the subset of current input indicators found to be relevant.

 

[userque]

In recent testing of recurrent neural networks, I concluded they are less accurate for future predictions than genetic programming. I'm guessing this is because a rule from a recurrent neural network applies operations on all the current input indicators while a rule from genetic programming applies operations on the subset of current input indicators found to be relevant.

Thanks! Appreciate the update.

 

[heavenskrow]

The same edge as last year. Which was the same edge as the year before. Which was the same edge as the year before. Which was the same edge as the year before. Which was the same edge as the year before. Which was the same edge as the year before. Which was the same edge as the year before. Which was the same edge as the year before...

Agreed. One thing I might add is perfecting the edge...that being my state of mind, my discipline, my emotions, my breathing, my risk management, etc.
Technical edge=always same edge, just developing the skill level.

 

[apo99]

trading illiquid tsx stocks just like the last 7 years