There is no guarantee that these codes works. I have no responsibility on any of the consequences arising from modifying,running, and distributing this code.
There is nothing amazing happening right now. It is under development, and there is no actual profits or real-time signaling properties so far. It will happen.
There is no GUI because I have no idea how to do that.
This is not for beginners because I'm not expert enough to make it a beginner friendly... I'm so sorry...
It requires python 2.7
and packages that can be installed with pip
(I'll make an official list).
I'm a python noob, so I don't know when it will be. It would be nice if I can make some setup.py
.
If you backtest using the AI in Master node, you can see that the winrate is around 65% (plus or minus 3%)!
- 日本語のドキュメンテーションの作成
Implimentation of RMSPropGravesThis will never happen.
Please report any issues or questions on issues
tab, so that I can ask all the questions at once,
and your question might already be answered there.
Go to
http://www.histdata.com/download-free-forex-historical-data/?/ninjatrader/tick-last-quotes/usdjpy/
Download the zip files for all the month from January of 2014 up to now, extract, and put the csv files in
./data/tickdata/rawdata/
Go in to the folder shapedata
run
python truncatenight.py
and
python shapetwosec.py
This will generate a csv files with all the USD/JPY data in an interval of 1 second (note that the linear extrapolation is taken for missing data points).
Go to tensorflow folder,
and run makeAI-tf.py
. make sure to turn cont_learn
to = False
at line 23
if you are running it for the first time.
Once saved!
is displayed on a terminal (the command window), run backtest.py
and see how it goes.
tweak checkQ
to be = True
to see how Q-values changes over transactions as a graph.
you need to have this trained at least 100,000 times to see a good progress (still not enough, I think).
Go to tensorflow folder,
and run makeAI-tf.py
, but this time make sure to turn cont_learn
to = True
at line 23
if you want to continue from the AI that is made previously.
Backtest can be done using tensorflow/backtest.py
. Just run it and it should output how much you have profitted
if you traded in whatever years you specified.
The key to the success is a corrrect Qthresh value! right now it is 1.45, but if you train further you might want to increase it!
You can run demotrade_tf.py
to test the AI on the real trading website (it is a demo trade from high-low Australia).
Enjoy.
run
make-AI-skl.py
this will create a base AI
then run
prob_clas.py
this takes a probability distribution of the base AI and make further predictions.
then run
backtest.py
to conduct backtesting.
If you have any questions or bugs, either submit in the issues tab or directly contact me via e-mail (t.hashizume.glasgow@gmail.com).