Generative Adversarial Networks on Stock Bars

1.General

Use GAN technology on stock bars (Chinese A stocks, stocks of USA and HK) to train two models: generator and discriminator.

1.1 Discriminator

which can be used as an indicator on a stock, similar with KDJ.

1.2 Generator

which can be used to select stocks or group them.

2.File

2.1 Discriminator.py

build discriminator model;

2.2 Generator.py

build generator model;

2.3 Generate_Bar.py

generate faked bars data from random seed by Generator and plot bars figure;

2.4 Sample.py

select sequential real bars data from real stocks trading data which are saved in ‘data’ Dir or would be downloaded by ‘tushare’ / ‘pandas_datareader’ if not in this Dir;

2.5 Gan_Bar.py

train GAN model on a certain stock data;

2.6 Train_A.py

train GAN model on all Chinese A stocks one by one(all needed data had been saved in ‘data’ Dir); It runs on Colab to train GAN on 7 accounts simultaneously and it has two mode: Master ans slave; ‘Master’ distributes task for each account(recorded in ‘task’ Dir); ‘Slave’ train GAN on own distributed stocks one by one.

3.Dir

3.1 Data

save stocks data which are downloaded by ‘tushare’ / ‘pandas_datareader’;

3.2 Results

each model weights, train loss, generate figs and discriminator indicator after complete GAN model training;

3.2.1 Train Loss

3.2.2 Generate Fig

3.2.3 Discriminator Indicator

3.3 Task

each task which is distributed by Train_A.py when it run on mode ‘Master’ and modified on mode ‘Slave’ on Colab for multi-training.