class Test_Environment(unittest.TestCase):
def setUp(self):
self.env = Environment()
#def tearDown(self):
#self.widget.dispose()
def test_create_environment(self):
self.assertNotEqual(self.env, None)
def test_get_data(self):
with open('config.json') as f:
config = json.load(f)
start_date = config['session']['start_date']
end_date = config['session']['end_date']
codes_num = config['session']['codes']
market = config['session']['market_types']
features = config['session']['features']
train_start_date, train_end_date, test_start_date, test_end_date, codes = self.env.get_repo(
start_date, end_date, codes_num, market)
window_length = 10
self.env.get_data(train_start_date, train_end_date, features,
window_length, market, codes)
self.assertTrue(
len(self.env.states) > 0
) # states has shape (1,6,10,2) 1,codes_num+1,window_length, features
self.assertTrue(len(self.env.price_history) >
0) #price_history has shape (6,1) codes_num + 1 ;
#First element in price_history is always 1, means cash
#print (self.env.states[0].shape)
#print (self.env.price_history[0].shape)
#print (self.env.price_history[0])
def test_get_repo(self):
with open('config.json') as f:
config = json.load(f)
start_date = config['session']['start_date']
end_date = config['session']['end_date']
codes_num = config['session']['codes']
market = config['session']['market_types']
self.train_start_date, self.train_end_date, test_start_date, test_end_date, self.codes = self.env.get_repo(
start_date, end_date, codes_num, market)
self.assertTrue(len(self.env.data) > 0)
self.assertTrue(len(self.env.date_set) > 0)
# step requires get_data to have been called first to fill the environment.
def test_step(self):
self.test_get_data()
self.env.reset()
noise_flag = False
info = self.env.step(None, None, noise_flag)
# dict_keys(['reward', 'continue', 'next state', 'weight vector', 'price', 'risk'])
#print (info.keys())
#print (info['reward']) # Reward is an integer
#print (info['continue']) # continue is True/False
#print (info['next state'].shape) # Shape for next state is (1,6,10,2)
#print (info['weight vector'].shape) # Shape for weight vector is (1,6)
#print (info['risk']) #Risk is an integer
#print (info['price'].shape) #Shape for price is 6,1)
self.assertEqual(len(info.keys(), 6))
def experiment_cost():
parser = build_parser()
args = vars(parser.parse_args())
with open('config.json') as f:
config = json.load(f)
with open("result/PG/" + str(args['num']) + '/config.json', 'r') as f:
dict_data = json.load(f)
codes, start_date, end_date, features, agent_config, market, predictor, framework, window_length, noise_flag, record_flag, plot_flag, reload_flag, trainable, method = parse_config(
config, args)
env = Environment()
test_start_date, test_end_date, codes = datetime.datetime.strptime(
dict_data['test_start_date'],
'%Y-%m-%d'), datetime.datetime.strptime(dict_data['test_end_date'],
'%Y-%m-%d'), dict_data['codes']
env.get_data(test_start_date, test_end_date, features, window_length,
market, codes)
costs = [0, 0.001, 0.005, 0.01, 0.02]
agent = PG(
len(codes) + 1, int(window_length), len(features),
'-'.join(agent_config), 'True', 'False', 'True', args['num'])
wealths_result = []
rs_result = []
for i in range(len(costs)):
stocktrader = StockTrader()
env.cost = costs[i]
info = env.step(None, None, 'False')
r, contin, s, w1, p, risk = parse_info(info)
contin = 1
wealth = 10000
wealths = [wealth]
rs = [1]
while contin:
w2 = agent.predict(s, w1)
env_info = env.step(w1, w2, 'False')
r, contin, s_next, w1, p, risk = parse_info(env_info)
wealth = wealth * math.exp(r)
rs.append(math.exp(r) - 1)
wealths.append(wealth)
s = s_next
stocktrader.update_summary(0, r, 0, 0, w2, p)
print('finish one agent')
wealths_result.append(wealths)
rs_result.append(rs)
print('agent', ' ', 'cumulative return', ' ', 'average daily return',
' ', 'sharpe ratio', ' ', 'maximum drawback')
plt.figure(figsize=(8, 6), dpi=100)
for i in range(len(costs)):
plt.plot(wealths_result[i], label=costs[i])
cumr = float((wealths_result[i][-1] - 10000) / 10000 * 100)
mrr = float(np.mean(rs_result[i]) * 100)
sharpe = float(
np.mean(rs_result[i]) / np.std(rs_result[i]) * np.sqrt(252))
maxdrawdown = float(
max(1 - min(wealths_result[i]) /
np.maximum.accumulate(wealths_result[i])))
print(costs[i], ' ', round(cumr, 3), '%', ' ', round(mrr, 3), '%',
' ', round(sharpe, 3), ' ', round(maxdrawdown, 3), '%')
plt.legend()
plt.xlabel('time')
plt.ylabel('wealth')
plt.savefig(PATH_prefix + 'backtest_differntcost.png')
plt.show()
