def test_predict():
print("STARTING PREDICT")
start_date = date(2017, 10, 1)
end_date = date(2018, 10, 1)
# setup environment
num_training_time = 200
env = PortfolioEnv(
start_date,
end_date,
window_length,
tickers_list,
features_list,
trading_cost=0.0025,
# trading_cost=0.0025,
batch_size=num_training_time,
buffer_bias_ratio=0.0)
#variable_scope ="ddpg"
action_dim = env.action_space.shape[0]
actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(action_dim))
#with tf.variable_scope(variable_scope):
tf.reset_default_graph()
sess = tf.Session()
with sess.as_default():
# Missing this was the source of one of the most challenging an insidious bugs that I've ever encountered.
# it was impossible to save the model.
K.set_session(sess)
ddpg_model = DDPG(env,
sess,
actor_noise,
obs_normalizer='history',
predictor_type=predictor_type,
use_batch_norm=use_batch_norm,
load_root_model=True,
config=CONFIG)
ddpg_model.initialize(load_weights=True)
observation = env.reset()
done = False
print(
'date, portfolio_change/y_return, portfolio_value, market_value, weight_cash, weight_AAPL'
)
while not done:
action = ddpg_model.predict_single(observation)
observation, reward, done, infos = env.step(action)
print(
datetime.fromtimestamp(infos['date']).strftime('%Y-%m-%d'),
infos['portfolio_change'] / infos['y_return'],
infos['portfolio_value'], infos['market_value'],
infos['weight_cash'], infos['weight_AAPL'])
df = env.df_info()
print(df.iloc[-1, :])
print(df.describe())
env.render(mode='notebook')
sess.close()
def main2():
print("STRATING")
start_date = date(2008, 1, 1)
end_date = date(2017, 1, 1)
features_list = ['open', 'high', 'low', 'close']
tickers_list = [
'AAPL', 'ATVI', 'CMCSA', 'COST', 'CSX', 'DISH', 'EA', 'EBAY', 'FB',
'GOOGL', 'HAS', 'ILMN', 'INTC', 'MAR', 'REGN', 'SBUX'
]
predictor_list = ['lstm', 'cnn', 'dense']
batch_norm_list = [False, True]
window_length_list = [5, 10, 15, 20]
for predictor_type in predictor_list:
for use_batch_norm in batch_norm_list:
for window_length in window_length_list:
# setup environment
num_training_time = 2000
env = PortfolioEnv(start_date,
end_date,
window_length,
tickers_list,
features_list,
batch_size=num_training_time)
# variable_scope ="ddpg"
action_dim = env.action_space.shape[0]
actor_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(action_dim))
model_save_path = get_model_path(window_length, predictor_type,
use_batch_norm)
summary_path = get_result_path(window_length, predictor_type,
use_batch_norm)
tf.reset_default_graph()
sess = tf.Session()
with sess.as_default():
# Missing this was the source of one of the most challenging an insidious bugs that I've ever encountered.
# it was impossible to save the model.
K.set_session(sess)
ddpg_model = DDPG2(env,
sess,
actor_noise,
action_dim=action_dim,
obs_normalizer="history",
predictor_type=predictor_type,
use_batch_norm=use_batch_norm,
model_save_path=model_save_path,
summary_path=summary_path)
ddpg_model.initialize(load_weights=False)
ddpg_model.train()
def test_main():
print("STARTING MAIN")
start_date = date(2008, 1, 1)
end_date = date(2017, 1, 1)
# setup environment
num_training_time = 2000
env = PortfolioEnv(start_date,
end_date,
window_length,
tickers_list,
features_list,
batch_size=num_training_time)
nb_assets = len(tickers_list) + 1
nb_features = len(features_list)
action_dim = env.action_space.shape[0]
root_model_dir_path = get_root_model_dir_path(window_length,
predictor_type,
use_batch_norm)
root_model_path = get_root_model_path(window_length, predictor_type,
use_batch_norm)
# create the network
state_input = Input(shape=(nb_assets, window_length, nb_features),
name="state_input")
root_model, imitation_model = create_network_given_future(
state_input,
predictor_type="cnn",
use_batch_norm=False,
weight_path=root_model_path,
load_weights=True)
history = env.src.full_history()
# create the dataset
(X_train, y_train), (X_validation,
y_validation) = create_optimal_imitation_dataset(
history, training_data_ratio=0.8)
# train
history = imitation_model.fit(X_train,
y_train,
batch_size=128,
epochs=100,
validation_data=(X_validation, y_validation),
shuffle=True)
os.makedirs(root_model_dir_path, exist_ok=True)
# remove last layer before saving
root_model.save_weights(root_model_path)
def test_main():
print("STARTING MAIN")
start_date = date(2008, 1, 1)
end_date = date(2017, 1, 1)
# setup environment
num_training_time = 2000
env = PortfolioEnv(start_date,
end_date,
window_length,
tickers_list,
features_list,
trading_cost=0.01,
batch_size=num_training_time)
#variable_scope ="ddpg"
nb_assets = len(tickers_list) + 1
action_dim = env.action_space.shape[0]
actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.ones(action_dim) /
nb_assets)
actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(action_dim))
#actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(action_dim)/nb_assets, sigma=0.1)
#with tf.variable_scope(variable_scope):
tf.reset_default_graph()
sess = tf.Session()
with sess.as_default():
# Missing this was the source of one of the most challenging an insidious bugs that I've ever encountered.
# it was impossible to save the model.
K.set_session(sess)
ddpg_model = DDPG(env,
sess,
actor_noise,
obs_normalizer="history",
predictor_type=predictor_type,
use_batch_norm=use_batch_norm,
load_root_model=True,
config=CONFIG)
ddpg_model.initialize(load_weights=True)
ddpg_model.train()
sess.close()
window_length = int(args['window_length'])
nb_classes = len(target_stocks) + 1
nb_features = len(features_list)
# get target history
target_history = np.empty(shape=(len(target_stocks), num_training_time,
history.shape[2]))
for i, stock in enumerate(target_stocks):
target_history[i] = history[
abbreviation.index(stock), :num_training_time, :]
# setup environment
env = PortfolioEnv(start_date,
end_date,
window_length,
tickers_list,
features_list,
batch_size=num_training_time)
action_dim = [nb_classes]
state_dim = [nb_classes, window_length]
batch_size = 64
action_bound = 1.
tau = 1e-3
assert args['predictor_type'] in ['cnn', 'lstm'
], 'Predictor must be either cnn or lstm'
predictor_type = args['predictor_type']
if args['batch_norm'] == 'True':
use_batch_norm = True
elif args['batch_norm'] == 'False':
use_batch_norm = False
def main():
start_date = date(2008, 1, 1)
end_date = date(2018, 1, 1)
features_list = ['open', 'high', 'low', 'close']
tickers_list = [
'AAPL', 'ATVI', 'CMCSA', 'COST', 'CSX', 'DISH', 'EA', 'EBAY', 'FB',
'GOOGL', 'HAS', 'ILMN', 'INTC', 'MAR', 'REGN', 'SBUX'
]
tickers_list = [
'AAPL', 'ATVI', 'CMCSA', 'FB', 'HAS', 'ILMN', 'INTC', 'MAR', 'REGN',
'SBUX'
]
historyManager = gdm.HistoryManager(tickers=tickers_list, online=True)
df = historyManager.historical_data(start=start_date,
end=end_date,
tickers=tickers_list,
features=features_list,
adjusted=True)
history = prepare_dataframe(df)
history = history[:, :, :4]
num_training_time = 1095
# get target history
target_history = np.empty(shape=(len(tickers_list), num_training_time,
history.shape[2]))
for i, stock in enumerate(tickers_list):
target_history[i] = history[
tickers_list.index(stock), :num_training_time, :]
predictor_list = ['cnn', 'lstm']
batch_norm_list = [False, True]
window_length_list = [5, 7, 10, 15, 20]
for predictor_type in predictor_list:
for use_batch_norm in batch_norm_list:
for window_length in window_length_list:
# setup environment
env = PortfolioEnv(start_date,
end_date,
window_length,
tickers_list,
features_list,
batch_size=num_training_time)
nb_classes = len(tickers_list) + 1
nb_features = len(features_list)
action_dim = [nb_classes]
state_dim = [nb_classes, window_length]
action_bound = 1.
tau = 1e-3
actor_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(action_dim))
model_save_path = get_model_path(window_length, predictor_type,
use_batch_norm)
summary_path = get_result_path(window_length, predictor_type,
use_batch_norm)
variable_scope = get_variable_scope(window_length,
predictor_type,
use_batch_norm)
tf.reset_default_graph()
sess = tf.Session()
with sess.as_default():
with tf.variable_scope(variable_scope):
root_net = None
state_inputs = Input(
shape=([nb_classes, window_length, nb_features]))
root_net = model_predictor(state_inputs,
predictor_type,
use_batch_norm)
actor = StockActor(sess=sess,
root_net=root_net,
inputs=state_inputs,
state_dim=state_dim,
action_dim=action_dim,
action_bound=action_bound,
learning_rate=1e-4,
tau=tau,
batch_size=window_length,
predictor_type=predictor_type,
use_batch_norm=use_batch_norm)
critic = StockCritic(sess=sess,
root_net=root_net,
inputs=state_inputs,
state_dim=state_dim,
action_dim=action_dim,
tau=1e-3,
learning_rate=1e-3,
predictor_type=predictor_type,
use_batch_norm=use_batch_norm)
ddpg_model = DDPG(env=env,
sess=sess,
actor=actor,
critic=critic,
actor_noise=actor_noise,
obs_normalizer="history",
model_save_path=model_save_path,
summary_path=summary_path)
ddpg_model.initialize(load_weights=True)
ddpg_model.train(debug=True)
sess.close()
def predict():
print("STRATING PreDICT")
start_date = date(2017, 1, 1)
end_date = date(2018, 1, 1)
features_list = ['open', 'high', 'low', 'close']
tickers_list = [
'AAPL', 'ATVI', 'CMCSA', 'COST', 'CSX', 'DISH', 'EA', 'EBAY', 'FB',
'GOOGL', 'HAS', 'ILMN', 'INTC', 'MAR', 'REGN', 'SBUX'
]
predictor_list = ['lstm', 'cnn', 'dense']
batch_norm_list = [False, True]
window_length_list = [5, 10, 15, 20]
for predictor_type in predictor_list:
for use_batch_norm in batch_norm_list:
for window_length in window_length_list:
# setup environment
num_training_time = 200
env = PortfolioEnv(start_date,
end_date,
window_length,
tickers_list,
features_list,
trading_cost=0.00,
batch_size=num_training_time)
# variable_scope ="ddpg"
action_dim = env.action_space.shape[0]
actor_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(action_dim))
model_save_path = get_model_path(window_length, predictor_type,
use_batch_norm)
summary_path = get_result_path(window_length, predictor_type,
use_batch_norm)
print("MODEL :", model_save_path)
tf.reset_default_graph()
sess = tf.Session()
with sess.as_default():
# Missing this was the source of one of the most challenging an insidious bugs that I've ever encountered.
# it was impossible to save the model.
K.set_session(sess)
ddpg_model = DDPG2(env,
sess,
actor_noise,
action_dim=action_dim,
obs_normalizer="history",
predictor_type=predictor_type,
use_batch_norm=use_batch_norm,
model_save_path=model_save_path,
summary_path=summary_path)
ddpg_model.initialize(load_weights=True)
observation = env.reset()
done = False
while not done:
action = ddpg_model.predict_single(observation)
observation, reward, done, infos = env.step(action)
print(
datetime.fromtimestamp(
infos['date']).strftime('%Y-%m-%d'),
infos['portfolio_change'] / infos['y_return'],
infos['portfolio_value'], infos['market_value'],
infos['weight_cash'], infos['weight_AAPL'])
df = env.df_info()
print(df.iloc[-1, :])
print(df.describe())
sess.close()
"""
:param config: config dictionary
:param fake_data: if True will use data generated randomly
:param restore_dir: path to the model trained before
:param save_path: path to save the model
:param device: the device used to train the network
:param agent: the nnagent object. If this is provides, the trainer will not
create a new agent by itself. Therefore the restore_dir will not affect anything.
"""
env = PortfolioEnv(
df=df_train,
steps=40,
scale=True,
trading_cost=0.0025,
window_length=window_length,
output_mode='EIIE',
# random_reset=False,
)
# wrap it in a few wrappers
env = ConcatStates(env)
env = SoftmaxActions(env)
environment = TFOpenAIGymCust('CryptoPortfolioEIIE-v0', env)
# sanity check out environment is working
state = environment.reset()
state, reward, done = environment.execute(env.action_space.sample())
state.shape
# sanity check out environment is working