1

My challenged article is below.

FX system trading code with Deep Learning on Python - Qiita http://qiita.com/ryo_grid/items/8dd84d2df1fcb7265ae1

However, I want know many other approches.

Ryo Kanbayashi
  • 11
  • 2
  • Three questions, if you permit: **[1]** What makes you believe to expect an ANN model, with a topology of just inputLayer[17]-hiddenLayer[50]-hiddenLayer[25]-outputLayer[2] to be called a Deep Learning candidate? **[2]**: What makes you believe any professional grade AI/ML-trading engine ( in wider sense ) or a QuantModel ( in narrow sense ) would ever arrive into public domain ( TAlib et al are community-sponsored efforts, but that does not mean it contains all advanced gems, does it? )? **[3]** How many M5 bars were needed to achieve +60% ~ +80% profits demonstrated for an epoch of 8 years? – user3666197 Aug 07 '17 at 01:33
  • One more: **[4]** How does a term *( cit. )* "**trends of portfolio**" ( referred to in the above cited article ) match the declaration, that the ANN was trained just on a single instrument ( the USDJPY ) - so what were the other components of the said portfolio and how was the other parts traded, besides the USDJPY instrument trading? – user3666197 Aug 07 '17 at 01:43
  • [1] sorry, I don't know ANN model [2] I believe AI/ML trading will overperome professional trader. but, these programs won't be open sourced. [3] you can calculate divide 8 years with 5min :-) – Ryo Kanbayashi Aug 07 '17 at 01:48
  • [4] I tried other pairs (not USD/JPY), but I can't get good performance on these pars. – Ryo Kanbayashi Aug 07 '17 at 01:52
  • [1] in[17]:hid1[50]:hid2[25]:out[2] ANN was in the source code. What do you mean by *(cit.)* "I don't know ANN model" -- the >>> ryogrid/fx_systrade is not your own code? – user3666197 Aug 07 '17 at 01:54
  • I understand. ryogrid/fx_systrade is my code :) – Ryo Kanbayashi Aug 07 '17 at 02:06
  • Ok, so back to the square [1]: **What makes you believe** to expect an ANN model, with a topology of just in[17]-hid[50]-hid[25]-out[2‌​] to **be called a Deep Learning**? – user3666197 Aug 07 '17 at 02:10
  • I think there is no well-established definition of DNN. And I think my topology is deep enough. So I called my topology DNN. – Ryo Kanbayashi Aug 07 '17 at 02:22
  • With all due respect, the only less-"deep" architecture would be a single hidden layer - which elementary algebra nomenclature calls a linear combination of factors X[17], does this sound as any reasonable FX-market model? No, not at all +just another 50:25 reduction layer will not help that much. This all was presented in a will to understand the concept & compare this to AI/ML-models with both TA+FA factors, having small thousands of features & having way deeper ML-prediction engines' architectures, with distributed-computing infra. Hope that could inspire your further research directions. – user3666197 Aug 07 '17 at 02:53

0 Answers0