1. Introduction to Tushare
1. About Tushare
Tushare is a free and open-source Python financial data interface package. The data content includes stocks, indices, funds, futures, bonds, foreign exchange, industry big data, etc., and also includes blockchain data such as digital currency market data, providing financial investment for various types of It has been fully upgraded to Tushare pro, which is very suitable for data analysis and visualization.
Official website link: Tushare big data community
2. Tushare installation and initialization
Github URL: https://github.com/waditu/Tushare
Install the tushare package using pip
pip install tushare lxmlimport tushare
import tushare as tsset token
ts.set_token('your token here')You can get the token by registering an account on the official website. The above method only needs to be called for the first time or after the token expires to complete the setting of calling the tushare data credential. Under normal circumstances, there is no need to repeat the setting. You can also ignore this step and complete the initialization directly with pro_api('your token').
Initialize the pro interface
pro = ts.pro_api()3. Tushare interface call
Tushare's data interface is divided into eleven categories including Shanghai and Shenzhen stocks, indexes, funds, futures, options, bonds, foreign exchange, Hong Kong stocks, industry economy, macro economy, and characteristic big data. Shanghai and Shenzhen stocks are divided into basic data and market data , financial data, market reference data four types of interface. Tushare also provides blockchain-related basic data, market data, and information announcements, as well as macroeconomics, foreign exchange, A-shares, blockchain, etc. , US stocks, oil, gold, gold foreign exchange, Hong Kong stocks, commodities, bonds, companies, markets, focus, central bank and other financial information.
The Tushare API interface needs to obtain corresponding access rights based on the number of points in the registered account. Insufficient points level may lead to unauthorized access to the API interface, and blockchain-related interfaces need to donate to obtain corresponding permissions.
Official website permission description: https://tushare.pro/document/1?doc_id=108
stock ticker data
Obtain stock basic information data, including stock code, high opening and bottom closing, etc.
pro = ts.pro_api()
df = pro.daily(ts_code='000001.SZ', start_date='20180701', end_date='20180718')
#多个股票
df = pro.daily(ts_code='000001.SZ,600000.SH', start_date='20180701', end_date='20180718')data sample
ts_code trade_date open high low close pre_close change pct_chg vol amount
0 000001.SZ 20180718 8.75 8.85 8.69 8.70 8.72 -0.02 -0.23 525152.77 460697.377
1 000001.SZ 20180717 8.74 8.75 8.66 8.72 8.73 -0.01 -0.11 375356.33 326396.994
2 000001.SZ 20180716 8.85 8.90 8.69 8.73 8.88 -0.15 -1.69 689845.58 603427.713
3 000001.SZ 20180713 8.92 8.94 8.82 8.88 8.88 0.00 0.00 603378.21 535401.175
4 000001.SZ 20180712 8.60 8.97 8.58 8.88 8.64 0.24 2.78 1140492.31 1008658.828
5 000001.SZ 20180711 8.76 8.83 8.68 8.78 8.98 -0.20 -2.23 851296.70 744765.824
6 000001.SZ 20180710 9.02 9.02 8.89 8.98 9.03 -0.05 -0.55 896862.02 803038.965
7 000001.SZ 20180709 8.69 9.03 8.68 9.03 8.66 0.37 4.27 1409954.60 1255007.609
8 000001.SZ 20180706 8.61 8.78 8.45 8.66 8.60 0.06 0.70 988282.69 852071.526
9 000001.SZ 20180705 8.62 8.73 8.55 8.60 8.61 -0.01 -0.12 835768.77 722169.579Financial Indicator Data
Obtain various financial indicator data of listed companies, such as ROA, ROE, etc.
pro = ts.pro_api()
df = pro.fina_indicator(ts_code='600000.SH')'20180718')data sample
ts_code ann_date end_date eps dt_eps total_revenue_ps revenue_ps \
0 600000.SH 20180830 20180630 0.95 0.95 2.8024 2.8024
1 600000.SH 20180428 20180331 0.46 0.46 1.3501 1.3501
2 600000.SH 20180428 20171231 1.84 1.84 5.7447 5.7447
3 600000.SH 20180428 20171231 1.84 1.84 5.7447 5.7447
4 600000.SH 20171028 20170930 1.45 1.45 4.2507 4.2507
5 600000.SH 20171028 20170930 1.45 1.45 4.2507 4.2507
6 600000.SH 20170830 20170630 0.97 0.97 2.9659 2.9659
7 600000.SH 20170427 20170331 0.63 0.63 1.9595 1.9595
8 600000.SH 20170427 20170331 0.63 0.63 1.9595 1.9595 2. Quantitative trading robot based on ElegantRL
1. Download data and preprocess
Use Tushare to download 1635 days of China A-share data, including basic market data, various technical indicators and financial indicators.
financial data
Write a function to download and process it into its own dataframe format
def get_fund_data(list_ticker):
fund_data = pd.DataFrame()
pro = ts.pro_api()
for i in list_ticker:
data = pro.fina_indicator(ts_code=i, start_date='20150118', end_date='20190731').drop_duplicates(subset='end_date')
fund_data = pd.concat([fund_data,data], axis=0)
date = pd.to_datetime(fund_data['end_date'],format='%Y%m%d').to_frame('date')
tic = fund_data['ts_code'].to_frame('tic')
OPM = fund_data['grossprofit_margin'].to_frame('OPM')
NPM = fund_data['netprofit_margin'].to_frame('NPM')
ROA = fund_data['roa'].to_frame('ROA')
ROE = fund_data['roe'].to_frame('ROE')
EPS = fund_data['eps'].to_frame('EPS')
cur_ratio = fund_data['current_ratio'].to_frame('cur_ratio')
quick_ratio = fund_data['quick_ratio'].to_frame('quick_ratio')
cash_ratio = fund_data['cash_ratio'].to_frame('cash_ratio')
acc_rec_turnover = fund_data['ar_turn'].to_frame('acc_rec_turnover')
debt_ratio = fund_data['debt_to_assets'].to_frame('debt_ratio')
debt_to_equity = fund_data['debt_to_eqt'].to_frame('debt_to_equity')
ratios = pd.concat([date,tic,OPM,NPM,ROA,ROE,EPS,
cur_ratio,quick_ratio,cash_ratio,acc_rec_turnover,
debt_ratio,debt_to_equity], axis=1)
ratios = ratios.sort_values(by=['date','tic']).reset_index(drop=True)
return ratiosThe results show
The same is true for basic market data and technical indicators.
Data connection and preprocessing
Remove illegal data
final_ratios = ratios.copy()
final_ratios = final_ratios.fillna(0)
final_ratios = final_ratios.replace(np.inf,0)Use date and tic to connect data
import itertools
list_date = list(pd.date_range(df['date'].min(),df['date'].max()))
combination = list(itertools.product(list_date,list_ticker))
# Merge stock price data and ratios into one dataframe
processed_full = pd.DataFrame(combination,columns=["date","tic"]).merge(df,on=["date","tic"],how="left")
processed_full = processed_full.merge(final_ratios,how='left',on=['date','tic'])
processed_full = processed_full.sort_values(['tic','date'])
Backfill financial indicator data to ensure data integrity
# Backfill the ratio data to make them daily
processed_full = processed_full.bfill(axis='rows')The results show
2. Training DRL model
The DRL model uses the open-source ElegantRL library as a framework.
MDP model of transaction task: We model the transaction task as a Markov decision process (MDP) (S, A, P, r, γ) , where S and A denote state space and action space respectively, P(s'| s, A) represents the transition probability, r(s, a) is the reward function, and γ∈(0, 1) is the discount factor. Specifically, the state represents an observation of the market environment by the DRL agent, the action space consists of actions that the agent can take in a certain state, and r (s, a, s') motivates the agent to learn better policies. The purpose of a trading agent is to learn a policy 
that maximizes the expected return 
.
Relevant codes are from FinRL-Meta Demo_China_A_share_market
Run function hyperparameter configuration
def run(gpu_id=0):
env = StockTradingEnv()
env_func = StockTradingEnv
env_args = get_gym_env_args(env=env, if_print=False)
env_args['beg_idx'] = 10 # training set
env_args['end_idx'] = 1296 # training set
args = Arguments(AgentPPO, env_func=env_func, env_args=env_args)
args.target_step = args.max_step * 4
args.reward_scale = 2 ** -7
args.learning_rate = 1.5 ** -14
args.break_step = int(10e5)
args.learner_gpus = gpu_id
args.random_seed += gpu_id + 1943
train_agent(args)Link ElegantRL library code
run()The function will use the following:
# env.py
class StockTradingEnv() # 来自FinRL库的交易仿真环境
def build_env() # 创建训练仿真环境
def get_gym_env_args() # 获得仿真环境的参数
# net.py
class ActorPPO # PPO算法的策略网络
class CriticPPO # PPO算法的价值网络
# agent.py
class AgentPPO # PPO算法的主体
class ReplayBufferList # 经验回放缓存（存放强化学习的训练数据）
# run.py
class Arguments # 强化学习的超参数（可以看这个类的注释了解超参数的作用）
def train_agent() # 训练强化学习智能体的函数In order to meet the needs of training, some modifications need to be made to the StockTradingEnv class.
Train and evaluate the model
Use the check_env() function to test the data of the test set
- random action indicates that the transaction action is random
action=rd.uniform(-1, 1, action_dim), and the agent trained by reinforcement learning should not be lower than this score. - buy all share means that the largest amount has been used to buy all stocks
action=np.ones(action_dim), which can reflect the trend of the market to a certain extent
For the convenience of demonstration and learning, less data is selected to compress the training time to within 1000 seconds. In fact, the training data needs to be an order of magnitude more.
| StockTradingEnv: close_ary.shape (339, 15)
| StockTradingEnv: tech_ary.shape (339, 120)
| StockTradingEnv: fund_ary.shape (339, 165)
cumulative_returns of random action: 1.21
cumulative_returns of random action: 1.20
cumulative_returns of random action: 1.18
cumulative_returns of random action: 1.43
cumulative_returns of buy all share: 1.36
cumulative_returns of buy all share: 1.36
cumulative_returns of buy all share: 1.36
cumulative_returns of buy all share: 1.36
Use the run() function for training
| StockTradingEnv: close_ary.shape (1625, 15)
| StockTradingEnv: tech_ary.shape (1625, 120)
| StockTradingEnv: fund_ary.shape (1625, 165)
| Arguments Remove cwd: ./StockTradingEnv-v2_PPO_0
| StockTradingEnv: close_ary.shape (1286, 15)
| StockTradingEnv: tech_ary.shape (1286, 120)
| StockTradingEnv: fund_ary.shape (1286, 165)
Step:7.71e+03 ExpR: -0.01 Returns: 0.84 ObjC: 0.66 ObjA: -0.02
Step:6.17e+04 ExpR: 0.27 Returns: 3.42 ObjC: 0.49 ObjA: -0.29
Step:1.16e+05 ExpR: 0.23 Returns: 3.18 ObjC: 0.26 ObjA: -0.31
Step:1.70e+05 ExpR: 0.07 Returns: 1.81 ObjC: 0.21 ObjA: -0.23
Step:2.24e+05 ExpR: 0.23 Returns: 3.24 ObjC: 0.17 ObjA: -0.24
Step:2.78e+05 ExpR: 0.00 Returns: 1.15 ObjC: 0.17 ObjA: -0.25
Step:3.32e+05 ExpR: 0.04 Returns: 1.42 ObjC: 0.11 ObjA: -0.25
Step:3.86e+05 ExpR: 0.05 Returns: 1.36 ObjC: 0.07 ObjA: -0.22
Step:4.39e+05 ExpR: 0.05 Returns: 1.51 ObjC: 0.06 ObjA: -0.28
Step:4.93e+05 ExpR: 0.05 Returns: 1.50 ObjC: 0.03 ObjA: -0.24
Step:5.47e+05 ExpR: 0.04 Returns: 1.48 ObjC: 0.15 ObjA: -0.32
Step:6.01e+05 ExpR: 0.05 Returns: 1.54 ObjC: 0.09 ObjA: -0.20
Step:6.55e+05 ExpR: 0.05 Returns: 1.54 ObjC: 0.06 ObjA: -0.25
Step:7.09e+05 ExpR: 0.07 Returns: 1.74 ObjC: 0.06 ObjA: -0.29
Step:7.63e+05 ExpR: 0.05 Returns: 1.40 ObjC: 0.03 ObjA: -0.25
Step:8.17e+05 ExpR: 0.05 Returns: 1.46 ObjC: 0.02 ObjA: -0.36
Step:8.71e+05 ExpR: 0.21 Returns: 2.96 ObjC: 0.25 ObjA: -0.21
Step:9.25e+05 ExpR: 0.21 Returns: 2.86 ObjC: 0.16 ObjA: -0.33
Step:9.79e+05 ExpR: 0.21 Returns: 2.97 ObjC: 0.10 ObjA: -0.29
| UsedTime: 1118 | SavedDir: ./StockTradingEnv-v2_PPO_0Model evaluation using the evaluate_models_in_directory() function
| StockTradingEnv: close_ary.shape (339, 15)
| StockTradingEnv: tech_ary.shape (339, 120)
| StockTradingEnv: fund_ary.shape (339, 165)
cumulative_returns 1.084 actor_00000000007710_00000011_-0000.01.pth
cumulative_returns 1.411 actor_00000000061680_00000073_00000.27.pth
cumulative_returns 1.381 actor_00000000115650_00000132_00000.23.pth
cumulative_returns 1.167 actor_00000000169620_00000192_00000.07.pth
cumulative_returns 1.445 actor_00000000223590_00000252_00000.23.pth
cumulative_returns 1.210 actor_00000000277560_00000311_00000.00.pth
cumulative_returns 1.126 actor_00000000331530_00000371_00000.04.pth
cumulative_returns 0.965 actor_00000000385500_00000430_00000.05.pth
cumulative_returns 0.965 actor_00000000439470_00000490_00000.05.pth
cumulative_returns 0.965 actor_00000000493440_00000552_00000.05.pth
cumulative_returns 1.025 actor_00000000547410_00000612_00000.04.pth
cumulative_returns 1.339 actor_00000000601380_00000671_00000.05.pth
cumulative_returns 1.265 actor_00000000655350_00000728_00000.05.pth
cumulative_returns 1.258 actor_00000000709320_00000788_00000.07.pth
cumulative_returns 1.258 actor_00000000763290_00000847_00000.05.pth
cumulative_returns 1.258 actor_00000000817260_00000909_00000.05.pth
cumulative_returns 1.421 actor_00000000871230_00000970_00000.21.pth
cumulative_returns 1.421 actor_00000000925200_00001031_00000.21.pth
cumulative_returns 1.421 actor_00000000979170_00001091_00000.21.pthcumulative_returns represents the rewards from the start to the end of the agent's transaction. For convenience, we directly display the "principal growth multiple".
actor_00000000004998_00000003_00000.08.pthIt is the strategy network model file saved during training actor, which means that it samples 4998 steps in the environment, trains for 3 seconds, and explores the environment with a score of 0.08. Gold increased by 1.093 times.
Connect Tensorboard visualization and parameter adjustment
Just keep it as a souvenir for yourself, please point out if you have any questions