musiclicn / deep_finance Goto Github PK

Upper shadow, body including color, lower shadow

https://github.com/yandexdataschool/Practical_RL/tree/master

df_sub = df[['close', 'cur_lable']]
df_sub = df_sub.dropna()

• 126 day change% as input

• change% in 10 class as output

缠论解析
http://product.dangdang.com/22808614.html

http://gouthamanbalaraman.com/blog/calculating-stock-beta.html

https://github.com/tensorflow/tensor2tensor

Bi looks strange in some cases, for example at the end of this QCOM example, the gap down has one up bar, which break the definition of down Bi because it does not have 3 consecutive down bar.
One proposed solution to this is add one more case which confirms a Bi, in 5 consecutive bars, if 4 out 5 are same trend, this confirms a Bi.
https://drive.google.com/file/d/0B4wqpXzYVibXV2dpcWJXOVNWUTA/view?usp=sharing

1. leave utils for downloading data in datasource
2. move process_bar.py, process_yahoo_data.py, generate_training_data.py to model folder
3. create folder for each model

to run export_png
conda install -c bokeh selenium
conda install phantomjs

input:
state: past N days prices including OHLC, 5 MA, 10 MA, 20 MA, 50 MA and 200 MA (including position)
dollar amount
action: NA (no action), buy 1/3, buy 2/3, buy all, sell 1/3, sell 2/3, sell all
reward: discounted future 5 day PNL

Notes:
use CNN for policy function/ network without pooling

http://chuansong.me/n/1925754342342

https://www.zhihu.com/question/52589498/answer/131396559

https://www.zhihu.com/question/20874888

http://scikit-learn.org/stable/modules/learning_curve.html

Reader is quite slow. Try construct x, y pair from iterate through rows instead of using Dataframe range slicing.

https://github.com/vnpy/vnpy

http://www.thealgoengineer.com/2014/download_sp500_data/

https://github.com/tensorflow/models/blob/master/slim/nets/inception_resnet_v2.py

use dates out side of dates used in training data
create separate folder for validation data and construct Reader from this folder

http://dataaspirant.com/2015/04/11/five-most-popular-similarity-measures-implementation-in-python/
How to deal with zoom in, zoom out?

https://startupsventurecapital.com/essential-cheat-sheets-for-machine-learning-and-deep-learning-researchers-efb6a8ebd2e5

apply LSTM according to tensorflow ptb example

https://journal.thriveglobal.com/if-youre-too-busy-for-these-5-things-your-life-is-more-off-course-than-you-think-b54f102598b0

input seq: past prices including OHLC
output seq: future prices mostly close prices

try to convert OHLC candlestick bar into some meaningful vector similar to word2vec

http://r2rt.com/recurrent-neural-networks-in-tensorflow-i.html

http://neuro.cs.ut.ee/demystifying-deep-reinforcement-learning/

Chan bi diff of start time, end time and slope.

http://pbpython.com/visualization-tools-1.html

Gravity 0

Basic concepts:
gravity = (high + low)/2
log change = log( price t/ price t-1)

Features:
log change of gravity
if is bar3 of fenxin, shape of fenxin- sum of previous three gravity log change
days of current trend
accumulated strength of current trend- sum log change of gravity belong to this trend
previous opposite trend strength, days
previous same trend strength, days
Target:
gravity direction (Up or Down)

https://numer.ai/rules

https://www.joinquant.com/post/425

https://github.com/proactiveSam/chan_lun_py/blob/master/chan_lun_util.py
https://www.joinquant.com/post/4948

pick best perform stock(s) in certain period time

http://book.fx110.com/book/read/75-15223#top_q

DXY, CDX
crude, weather

http://stooq.com/db/h/
http://www.quantshare.com/sa-426-6-ways-to-download-free-intraday-and-tick-data-for-the-us-stock-market