def get_reward(self, portfolio: Portfolio, stock_market_data: StockMarketData):
current_portfolio_value = portfolio.get_value(stock_market_data)
if self.last_portfolio_value < current_portfolio_value:
return 100 * (current_portfolio_value / self.last_portfolio_value)
elif self.last_portfolio_value > portfolio.get_value(stock_market_data):
return -100
return -20
def test_get_value_with_date(self):
stock_market_data = StockMarketData([Company.A, Company.B], [Period.TESTING])
date = Date(2012, 1, 3)
portfolio = Portfolio(100.0)
self.assertEqual(portfolio.get_value(stock_market_data, date), 100.0)
portfolio = Portfolio(100.0, {Company.A: 10})
self.assertEqual(portfolio.get_value(stock_market_data, date), 455.54107999999997)
portfolio = Portfolio(100.0, {Company.A: 10, Company.B: 10})
self.assertEqual(portfolio.get_value(stock_market_data, date), 2046.9924999999998)
def test_get_value_without_date(self):
stock_market_data = StockMarketData([Company.A, Company.B], [Period.TESTING])
portfolio = Portfolio()
self.assertEqual(portfolio.get_value(stock_market_data), 0)
portfolio = Portfolio(100.0)
self.assertEqual(portfolio.get_value(stock_market_data), 100.0)
portfolio = Portfolio(100.0, {Company.A: 10})
self.assertEqual(portfolio.get_value(stock_market_data), 1127.59895)
portfolio = Portfolio(100.0, {Company.A: 10, Company.B: 10})
self.assertEqual(portfolio.get_value(stock_market_data), 2416.5398400000004)
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
order_list = []
vote_a = self.expert_a.vote(stock_market_data[Company.A])
vote_b = self.expert_b.vote(stock_market_data[Company.B])
# convert votes to state
state = np.array([self.vote2num[vote_a], self.vote2num[vote_b]])
if self.train_while_trading:
if len(self.memory) > self.min_size_of_memory_before_training:
# helper function which executes experience replay
self._replay()
# act
action = self._act(state, stock_market_data, portfolio, order_list)
if self.last_portfolio_value is not None:
# Reward function R1
if self.last_portfolio_value <= portfolio.get_value(
stock_market_data):
reward = 1
else:
reward = 0
# Reward function R2
#reward = (portfolio.get_value(
# stock_market_data) - self.last_portfolio_value) / self.last_portfolio_value
# helper to fill up the queue for performance replay
self._remember(self.last_state, action, reward, state)
# save state and portfolio value
self.last_portfolio_value = portfolio.get_value(stock_market_data)
self.last_state = state
return order_list
def gen_reward(self, portfolio: Portfolio,
stock_market_data: StockMarketData):
print('gen_reward')
if self.last_portfolio_value < portfolio.get_value(stock_market_data):
return self.reward_factor * (portfolio.get_value(stock_market_data)
/ self.last_portfolio_value)
elif self.last_portfolio_value > portfolio.get_value(
stock_market_data):
return -self.reward_factor * (
portfolio.get_value(stock_market_data) /
self.last_portfolio_value)
else:
return -self.reward_factor / 5
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
current_state = get_state(self, stock_market_data)
if self.train_while_trading and self.last_state is not None:
reward = get_reward(self, portfolio, stock_market_data)
self.memory.append(
(self.last_state, self.last_action, reward, current_state))
train_neural_net(self)
action_index = get_index_for_action_to_execute(self, current_state)
self.last_state = current_state
self.last_action = action_index
self.last_portfolio_value = portfolio.get_value(stock_market_data)
return get_order_list(self, portfolio, stock_market_data)
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# calculate current state
state_ = state_maker(self.expert_a, self.expert_b, stock_market_data,
portfolio)
curr_state = state_.create_numerical_state(self.state_size)
# calculate current portfolio value
curr_portfolio_value = portfolio.get_value(stock_market_data)
# train or testing mode
if self.train_while_trading == True:
# Store state as experience (memory) and train the neural network only if trade() was called before at least once
if self.last_state is not None:
reward = self.get_rewards(self.last_portfolio_value,
curr_portfolio_value)
self.remember(self.last_state, self.last_action_a, reward,
curr_state)
if len(self.memory) > self.min_size_of_memory_before_training:
self.replay()
# Create actions for current state and decrease epsilon for fewer random actions
curr_action_a = self.get_best_action(curr_state)
final_order = self.create_order_list(curr_action_a, portfolio,
stock_market_data)
self.decrease_epsilon()
else:
# predict best action from neuronal net
curr_action_a = self.model.predict(curr_state)
curr_action_a = np.argmax(curr_action_a[0])
final_order = self.create_order_list(curr_action_a, portfolio,
stock_market_data)
# Save created state, actions and portfolio value for the next call of trade() --> erledigt
self.last_state = curr_state
self.last_action_a = curr_action_a
self.last_portfolio_value = curr_portfolio_value
return final_order
def create_reward(self, portfolio: Portfolio,
stock_market_data: StockMarketData, state_now: State):
new_portfolio_value = portfolio.get_value(stock_market_data)
index_of_action = Action.get_id_from_action(self.last_action)
reward = -1 if (self.last_portfolio_value -
new_portfolio_value) > 0 else 1
reward = 0 if (self.last_portfolio_value -
new_portfolio_value) == 0 else reward
reward_array = np.zeros([9])
reward_array[index_of_action] = reward
q_next = self.run_model(state_now)
weighted_q_next = q_next * self.gamma
reward_array = np.sum([reward_array, weighted_q_next], axis=0)
#reward_array[index_of_action] += self.gamma * q_next[index_of_action]
return reward_array
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
state = State(portfolio,
self.expert_a.vote(stock_market_data[Company.A]),
self.expert_b.vote(stock_market_data[Company.B]),
stock_market_data[Company.A],
stock_market_data[Company.B])
param = state.get_nn_input_state()
if self.train_while_trading and self.last_state is not None:
current_reward = self.gen_reward(portfolio, stock_market_data)
self.update_memory(current_reward, param)
self.train_model()
self.last_action = self.get_action_idx(param)
self.last_state = param
self.last_portfolio_value = portfolio.get_value(stock_market_data)
orders = self.get_orders(stock_market_data, portfolio)
return orders
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# Compute the current state
current_state = self.current_state(stock_market_data)
current_portfolio_value = portfolio.get_value(stock_market_data)
# Store state as experience (memory) and train the neural network only if trade() was called before at least once
if self.last_action is not None and self.train_while_trading:
reward = self.reward(current_portfolio_value)
self.memory.append(
(self.last_state, self.last_action, reward, current_state))
if len(self.memory) > self.min_size_of_memory_before_training:
self.experience_replay()
# Create actions for current state and decrease epsilon for fewer random actions
action = self.get_action(current_state)
self.epsilon = max(
[self.epsilon_min, self.epsilon * self.epsilon_decay])
# Save created state, actions and portfolio value for the next call of trade()
self.last_state = current_state
self.last_action = action
self.last_portfolio_value = current_portfolio_value
# convert action to orderlist
return self.mapping_action_order(action, portfolio, stock_market_data)
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock marketf"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# TODO Compute the current state
stock_data_a = None
stock_data_b = None
last_stock_data_a = None
last_stock_data_b = None
company_list = stock_market_data.get_companies()
for company in company_list:
if company == Company.A:
stock_data_a = stock_market_data[Company.A]
last_stock_data_a = stock_data_a.get_from_offset(-2)
elif company == Company.B:
stock_data_b = stock_market_data[Company.B]
last_stock_data_b = stock_data_b.get_from_offset(-2)
else:
assert False
vote_a = self.expert_a.vote(stock_data_a)
vote_b = self.expert_b.vote(stock_data_b)
state = State(last_stock_data_a, last_stock_data_b, vote_a, vote_b)
# TODO Q-Learning
nn_input = np.array(
[np.array([state.aDiff, state.vote_a, state.bDiff, state.vote_b])])
action_vals = self.model.predict(nn_input)
# TODO Store state as experience (memory) and train the neural network only if trade() was called before at least once
# TODO Create actions for current state and decrease epsilon for fewer random actions
actions = [[
Order(OrderType.BUY, Company.A,
int((portfolio.cash / 2) // stock_data_a.get_last()[-1])),
Order(OrderType.BUY, Company.B,
int((portfolio.cash / 2) // stock_data_b.get_last()[-1]))
],
[
Order(
OrderType.BUY, Company.A,
int((portfolio.cash) //
stock_data_a.get_last()[-1])),
Order(OrderType.SELL, Company.B,
portfolio.get_stock(Company.B))
],
[
Order(OrderType.SELL, Company.A,
portfolio.get_stock(Company.A)),
Order(
OrderType.BUY, Company.B,
int(portfolio.cash // stock_data_b.get_last()[-1]))
],
[
Order(OrderType.SELL, Company.A,
portfolio.get_stock(Company.A)),
Order(OrderType.SELL, Company.B,
portfolio.get_stock(Company.B))
],
[
Order(OrderType.SELL, Company.A, 0),
Order(OrderType.SELL, Company.B, 0)
]]
if not self.train_while_trading:
self.epsilon = 0.0
else:
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
else:
self.epsilon = self.epsilon_min
# randomize action
if random.random() < self.epsilon:
next_action = random.choice(list(range(self.action_size)))
else:
next_action = np.argmax(action_vals[0])
order_list = actions[next_action]
portfolio_value = portfolio.get_value(
stock_market_data, stock_market_data.get_most_recent_trade_day())
if (self.last_state != None and self.train_while_trading):
def reward(oldVal, newVal):
neg = -100.0
pos = 100.0
q = newVal / oldVal
if q < 1:
return neg
elif q == 1:
return -10
else:
print("Q: ", q)
return pos / 2 * oldVal / newVal
r = reward(self.last_portfolio_value, portfolio_value)
action_vals[0][self.last_order] = r
self.memory.append([self.last_input, action_vals])
if (len(self.memory) > self.min_size_of_memory_before_training):
sample = random.sample(self.memory, self.batch_size)
trainSample = list()
testSample = list()
for [sampleIn, sampleOut] in sample:
trainSample.append(sampleIn[0])
testSample.append(sampleOut[0])
self.model.fit(np.array(trainSample), np.array(testSample),
self.batch_size)
# Save created state, actions and portfolio value for the next call of trade()
self.last_input = nn_input
self.last_state = state
self.last_order = next_action
self.last_portfolio_value = portfolio_value
print(next_action, action_vals, portfolio.cash,
portfolio.get_stock(Company.A), portfolio.get_stock(Company.B))
return order_list
def trade(self, portfolio: Portfolio, stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# INPUT layer: 1 (buy or sell A?)
# output layer: 2 ([buy_A, sell_A])
# TODO Compute the current state
stock_data_A = stock_market_data[Company.A]
expertA_voteA = self.expert_a.vote(stock_data_A)
expertB_voteA = self.expert_b.vote(stock_data_A)
stock_data_B = stock_market_data[Company.B]
expertA_voteB = self.expert_a.vote(stock_data_B)
expertB_voteB = self.expert_b.vote(stock_data_B)
state = np.array([[
self.vote_map[expertA_voteA] + self.vote_map[expertB_voteA],
self.vote_map[expertA_voteB] + self.vote_map[expertB_voteB],
]])
# do action 0 or 1?
predictions = self.model.predict(state)
'''
if random.random() < self.epsilon:
# use random actions for A and B
action_A = random.randrange(2)
action_B = random.randrange(2)
else:
# use prediction actions
action_A = np.argmax(predictions[0][0:2])
action_B = np.argmax(predictions[0][2:4])
'''
action_A = np.argmax(predictions[0][0:2])
action_B = np.argmax(predictions[0][2:4])
current_price_a = stock_market_data.get_most_recent_price(Company.A)
current_price_b = stock_market_data.get_most_recent_price(Company.B)
money_to_spend = portfolio.cash
order_list = []
# do stuff for A
if action_A == 0:
# buy all A
amount_to_buy = money_to_spend // current_price_a
if amount_to_buy > 0:
money_to_spend -= amount_to_buy * current_price_a
order_list.append(Order(OrderType.BUY, Company.A, amount_to_buy))
elif action_A == 1:
# sell all A
amount_to_sell = portfolio.get_stock(Company.A)
if amount_to_sell > 0:
order_list.append(Order(OrderType.SELL, Company.A, amount_to_sell))
else:
assert False
# do stuff for B
if action_B == 0:
# buy all B
amount_to_buy = money_to_spend // current_price_b
if amount_to_buy > 0:
order_list.append(Order(OrderType.BUY, Company.B, amount_to_buy))
elif action_B == 1:
# sell all B
amount_to_sell = portfolio.get_stock(Company.B)
if amount_to_sell > 0:
order_list.append(Order(OrderType.SELL, Company.B, amount_to_sell))
else:
assert False
if self.last_state is not None:
# train
diff_a = (current_price_a / self.last_price_a - 1)
diff_b = (current_price_b / self.last_price_b - 1)
fut_reward_a_buy = np.max(predictions[0][0])
fut_reward_a_buy = fut_reward_a_buy if fut_reward_a_buy > 0 else 0
fut_reward_a_sell = np.max(predictions[0][1])
fut_reward_a_sell = fut_reward_a_sell if fut_reward_a_sell > 0 else 0
fut_reward_b_buy = np.max(predictions[0][2])
fut_reward_b_buy = fut_reward_b_buy if fut_reward_b_buy > 0 else 0
fut_reward_b_sell = np.max(predictions[0][3])
fut_reward_b_sell = fut_reward_b_sell if fut_reward_b_sell > 0 else 0
reward_vec = np.array([[
diff_a + self.gamma * fut_reward_a_buy,
-diff_a + self.gamma * fut_reward_a_sell,
diff_b + self.gamma * fut_reward_b_buy,
-diff_b + self.gamma * fut_reward_b_sell
]])
#reward_vec = np.array([[portfolio.get_value(stock_market_data)]])
self.model.fit(self.last_state, reward_vec, verbose=0)
self.last_state = state
self.last_action_a = action_A
self.last_action_b = action_B
self.last_portfolio_value = portfolio.get_value(stock_market_data)
self.last_price_a = current_price_a
self.last_price_b = current_price_b
return order_list
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# 1. Compute current state
state = self.compute_state(stock_market_data)
# 1.2 If training is turned off, just predict the next action and return orders
if not self.train_while_trading:
self.last_state = state
actionSpace = self.model.predict(state)
action = np.argmax(actionSpace[0])
orders = self.action_to_order(action, portfolio, stock_market_data)
return orders
# 2. Get a random action with the probability of epsilon, otherwise predict the action via the ANN
if np.random.rand() <= self.epsilon and self.train_while_trading:
action = np.random.randint(self.action_size, size=1)[0]
else:
actionSpace = self.model.predict(state)
action = np.argmax(actionSpace[0])
# 3. Reduce Epsilon if it is bigger than epsilon min
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
# 4. Training of the ANN
if self.train_while_trading and self.last_state is not None:
# 4.1 Get reward
reward = self.get_reward(portfolio.get_value(stock_market_data),
self.last_portfolio_value)
# 4.2 Store memory
self.memory.append(
[self.last_state, self.last_action, reward, state])
# 4.3 Actual training via Experience Replay
if len(self.memory) > self.min_size_of_memory_before_training:
self.experienceReplay(self.batch_size)
# 5. Map Action + Create Order
orders = self.action_to_order(action, portfolio, stock_market_data)
# 6. Save the values
self.last_state = state
self.last_action = action
self.last_portfolio_value = portfolio.get_value(stock_market_data)
return orders
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# INPUT layer: 1 (buy or sell A?)
# output layer: 2 ([buy_A, sell_A])
# TODO Compute the current state
stock_data_A = stock_market_data[Company.A]
expertA_voteA = self.expert_a.vote(stock_data_A)
expertB_voteA = self.expert_b.vote(stock_data_A)
stock_data_B = stock_market_data[Company.B]
expertA_voteB = self.expert_a.vote(stock_data_B)
expertB_voteB = self.expert_b.vote(stock_data_B)
state = np.array([[
self.vote_map[expertA_voteA] + self.vote_map[expertB_voteA],
self.vote_map[expertA_voteB] + self.vote_map[expertB_voteB],
]])
# do action 0 or 1?
predictions = self.model.predict(state)
#print(f'predictions:{predictions}')
#input()
action_A = np.argmax(predictions[0][0:2])
action_B = np.argmax(predictions[0][2:4])
most_recent_price_A = stock_market_data.get_most_recent_price(
Company.A)
most_recent_price_B = stock_market_data.get_most_recent_price(
Company.B)
order_list = []
money_to_spend = portfolio.cash
# do stuff for A
if action_A == 0:
# buy all A
amount_to_buy = money_to_spend // most_recent_price_A
if amount_to_buy > 0:
money_to_spend -= amount_to_buy * most_recent_price_A
order_list.append(
Order(OrderType.BUY, Company.A, amount_to_buy))
elif action_A == 1:
# sell all A
amount_to_sell = portfolio.get_stock(Company.A)
if amount_to_sell > 0:
order_list.append(
Order(OrderType.SELL, Company.A, amount_to_sell))
else:
assert False
# do stuff for B
if action_B == 0:
# buy all B
amount_to_buy = money_to_spend // most_recent_price_B
if amount_to_buy > 0:
order_list.append(
Order(OrderType.BUY, Company.B, amount_to_buy))
elif action_B == 1:
# sell all B
amount_to_sell = portfolio.get_stock(Company.B)
if amount_to_sell > 0:
order_list.append(
Order(OrderType.SELL, Company.B, amount_to_sell))
else:
assert False
if self.last_state is not None:
# train
diff = (portfolio.get_value(stock_market_data) /
self.last_portfolio_value - 1)
rec_vec = np.array([[-diff, -diff, -diff, -diff]])
rec_vec[0][self.last_action_a] = diff
rec_vec[0][2 + self.last_action_b] = diff
#reward_vec = np.array([[portfolio.get_value(stock_market_data)]])
self.model.fit(self.last_state, rec_vec)
self.last_state = state
self.last_action_a = action_A
self.last_action_b = action_B
self.last_portfolio_value = portfolio.get_value(stock_market_data)
return order_list
def trade(self, portfolio: Portfolio, stock_market_data: StockMarketData) -> List[Order]:
"""
Generate actions to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# Compute the current state
expert_votes = [
self.experts[i].vote(stock_market_data[company])
for i, company in enumerate(stock_market_data.get_companies())
]
shares_owned = [
portfolio.get_stock(company)
for company in stock_market_data.get_companies()
]
if State is StateExpertsOnly:
state = StateExpertsOnly(expert_votes, portfolio.get_value(stock_market_data))
elif State is StateExpertsCashShares:
state = StateExpertsCashShares(expert_votes, portfolio.cash, shares_owned, portfolio.get_value(stock_market_data))
else:
raise RuntimeError
if self.train_while_trading:
# store state as experience in memory
if len(self.memory) > 0:
self.memory[-1].state2 = state
experience = Experience(
state1=state
)
self.memory.append(experience)
# train
if len(self.memory) >= self.min_size_of_memory_before_training:
if self.days_passed % self.train_each_n_days == 0:
self.train()
# determine probability for random actions
if not self.is_evolved_model:
# first training episode
random_action_probability = (
(self.epsilon ** self.days_passed) * (1.0 - self.random_action_min_probability) +
self.random_action_min_probability
)
else:
# subsequent training episode
random_action_probability = self.random_action_min_probability
if self.training_occasions == 0 or random.random() < random_action_probability:
actions = [Action.get_random(), Action.get_random()]
else:
# choose actions by querying network
x = state.to_input()
y = self.model.predict(numpy.array([x]))
assert y.shape == (1, self.action_size)
actions = Action.from_model_prediction(y[0])
experience.actions = actions
else:
# not training -> always choose actions by querying network
actions = Action.from_model_prediction(self.model.predict(numpy.array([state.to_input()]))[0])
# translate actions into orders
orders: typing.List[Order] = []
companies_with_actions_and_magnitudes = list(zip(list(Company), actions, Action.get_action_magnitudes(actions)))
for comp, action, mag in companies_with_actions_and_magnitudes:
if action.is_buy():
cash_limit = portfolio.cash * mag
date, stock_price = stock_market_data[comp].get_last()
shares_amount = cash_limit / stock_price
if shares_amount > 0:
orders.append(Order(OrderType.BUY, comp, shares_amount))
elif action.is_sell():
shares_amount = portfolio.get_stock(comp) * mag
if shares_amount > 0:
orders.append(Order(OrderType.SELL, comp, shares_amount))
self.days_passed += 1
return orders
def trade(self, portfolio: Portfolio, stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
self.day += 1
# TODO Compute the current state
order_list = []
stock_data_a = stock_market_data[Company.A]
stock_data_b = stock_market_data[Company.B]
# Expert A
expert_a = self.expert_a.vote(stock_data_a)
# Expert B
expert_b = self.expert_b.vote(stock_data_b)
state = np.array([[
self.vote_num[expert_a], self.vote_num[expert_b], ]])
action = self.decide_action(state)
# TODO Store state as experience (memory) and train the neural network only if trade() was called before at least once
if self.last_state is not None:
reward = (portfolio.get_value(stock_market_data) - self.last_portfolio_value) / self.last_portfolio_value
self.memory.append((self.last_state, self.last_action, reward, state))
self.train_network(self.batch_size)
# TODO Create actions for current state and decrease epsilon for fewer random actions
act0 = 0
act1 = 0
act2 = 0
act3 = 0
act4 = 0
act5 = 0
act6 = 0
act7 = 0
act8 = 0
# What amount of the stocks should be bought or sold
percent_buy = 1
percent_sell = 1
if action == 0:
# Buy A
stock_price_a = stock_market_data.get_most_recent_price(Company.A)
amount_to_buy_a = int((portfolio.cash*percent_buy/2) // stock_price_a)
if amount_to_buy_a > 0:
order_list.append(Order(OrderType.BUY, Company.A, amount_to_buy_a))
# Buy B
stock_price_b = stock_market_data.get_most_recent_price(Company.B)
amount_to_buy_b = int((portfolio.cash*percent_buy/2) // stock_price_b)
if amount_to_buy_b > 0:
order_list.append(Order(OrderType.BUY, Company.B, amount_to_buy_b))
act0 += 1
elif action == 1:
# Buy A
stock_price_a = stock_market_data.get_most_recent_price(Company.A)
amount_to_buy_a = int(portfolio.cash *percent_buy// stock_price_a)
if amount_to_buy_a > 0:
order_list.append(Order(OrderType.BUY, Company.A, amount_to_buy_a))
# Sell B
amount_to_sell_b = int(portfolio.get_stock(Company.B)*percent_sell)
if amount_to_sell_b > 0:
order_list.append(Order(OrderType.SELL, Company.B, amount_to_sell_b))
act1 += 1
elif action == 2:
# Sell A
amount_to_sell_a = int(portfolio.get_stock(Company.A)*percent_sell)
if amount_to_sell_a > 0:
order_list.append(Order(OrderType.SELL, Company.A, amount_to_sell_a))
# Buy B
stock_price_b = stock_market_data.get_most_recent_price(Company.B)
amount_to_buy_b = int(portfolio.cash*percent_buy // stock_price_b)
if amount_to_buy_b > 0:
order_list.append(Order(OrderType.BUY, Company.B, amount_to_buy_b))
act2 += 1
elif action == 3:
# Sell A
amount_to_sell_a = int(portfolio.get_stock(Company.A)*percent_sell)
if amount_to_sell_a > 0:
order_list.append(Order(OrderType.SELL, Company.A, amount_to_sell_a))
# Sell B
amount_to_sell_b = int(portfolio.get_stock(Company.B)*percent_sell)
if amount_to_sell_b > 0:
order_list.append(Order(OrderType.SELL, Company.B, amount_to_sell_b))
act3 += 1
elif action == 4:
# Sell A
amount_to_sell_a = int(portfolio.get_stock(Company.A)*percent_sell)
if amount_to_sell_a > 0:
order_list.append(Order(OrderType.SELL, Company.A, amount_to_sell_a))
# Hold B
act4 += 1
elif action == 5:
# Hold A
# Sell B
amount_to_sell_b = int(portfolio.get_stock(Company.B)*percent_sell)
if amount_to_sell_b > 0:
order_list.append(Order(OrderType.SELL, Company.B, amount_to_sell_b))
act5 += 1
elif action == 6:
# Buy A
stock_price_a = stock_market_data.get_most_recent_price(Company.A)
amount_to_buy_a = int((portfolio.cash*percent_buy) // stock_price_a)
if amount_to_buy_a > 0:
order_list.append(Order(OrderType.BUY, Company.A, amount_to_buy_a))
# Hold B
act6 += 1
elif action == 7:
# Hold A
# Buy B
stock_price_b = stock_market_data.get_most_recent_price(Company.B)
amount_to_buy_b = int((portfolio.cash*percent_buy) // stock_price_b)
if amount_to_buy_b > 0:
order_list.append(Order(OrderType.BUY, Company.B, amount_to_buy_b))
act7 += 1
elif action == 8:
# Hold A
# Hold B
order_list.append(Order(OrderType.BUY, Company.B, 0))
act8 += 1
else:
print("undefined action called"+str(action))
# Decrease the epsilon for fewer random actions
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
# TODO Save created state, actions and portfolio value for the next call of trade()
self.last_state = state
self.last_action = action
self.last_portfolio_value = portfolio.get_value(stock_market_data)
return order_list
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
#Compute current State
state_now = State(stock_market_data, self.expert_a, self.expert_b,
self.last_state)
if (self.last_state == None):
state_now.prev_state = state_now
#Create actions and let probability decide
rand_action = Action.create_random_action()
model_action = self.action_by_model(state_now)
action = np.random.choice([rand_action, model_action],
1,
p=[self.epsilon, 1 - self.epsilon])[0]
if (self.epsilon > self.epsilon_min):
self.epsilon -= self.epsilon_decay
#if training is deactivated
if not self.train_while_trading:
self.last_state = state_now
return model_action.create_order_list(portfolio, stock_market_data)
#for the first call of trade()
if (self.last_state == None):
self.epsilon = 1.0
self.last_state = state_now
self.last_portfolio_value = portfolio.get_value(stock_market_data)
self.last_action = action
return action.create_order_list(portfolio, stock_market_data)
#calculate reward and create tuple for memory
reward = self.create_reward(portfolio, stock_market_data, state_now)
memory_unit = (self.last_state, self.last_action, reward, state_now)
self.memory.append(memory_unit)
#train if there is enough experience
if len(self.memory) > self.min_size_of_memory_before_training:
#start training with random batch
batch = random.sample(self.memory, self.batch_size)
x = np.empty((self.batch_size, self.state_size))
y = np.empty((self.batch_size, self.action_size))
i = 0
for (s1, a, r, s2) in batch:
x[i] = s1.create_array_diff()
y[i] = r
i += 1
self.model.train_on_batch(x, y)
#save old values for next run
self.last_state = state_now
self.last_portfolio_value = portfolio.get_value(stock_market_data)
self.last_action = action
return action.create_order_list(portfolio, stock_market_data)
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# TODO Compute the current state
stock_data_a = stock_market_data[Company.A]
vote_a = self.expert_a.vote(stock_data_a)
stock_data_b = stock_market_data[Company.B]
vote_b = self.expert_a.vote(stock_data_b)
curr_state = np.array([[
#portfolio.cash,
#portfolio.stocks[Company.A],
#portfolio.stocks[Company.B],
#stock_market_data.get_most_recent_price(Company.A),
#stock_market_data.get_most_recent_price(Company.B),
self.vote_map[vote_a],
self.vote_map[vote_b],
#vote_a,
#vote_b
]])
# TODO Store state as experience (memory) and train the neural network only if trade() was called before at least once
# <s, a, r, s'>
trade_called_once_before = (self.last_state is not None)
experience = (self.last_state, self.last_action_a, self.last_action_b,
self.last_portfolio_value, curr_state)
self.memory.append(experience)
if trade_called_once_before and self.min_size_of_memory_before_training <= len(
self.memory):
# create self.batch_size-times random numbers
# between 0 and length of queue
selected_mems_ind = random.sample(range(0, len(self.memory)),
self.batch_size)
selected_mems = [self.memory[i] for i in selected_mems_ind]
# reverse the list, so del indices works out
selected_mems_ind = sorted(selected_mems_ind)[::-1]
for i in selected_mems_ind:
print(f'len(self.memory): {len(self.memory)}')
print(f'deleting: {i}')
del self.memory[i]
# mem[:2] -> s, a
# mem[2] -> r
X = [np.asarray(mem[:3]) for mem in selected_mems]
Y = [np.asarray(mem[3]) for mem in selected_mems]
self.model.fit(X, Y, batch_size=self.batch_size)
# TODO Create actions for current state and decrease epsilon for fewer random actions
# Order(OrderType.SELL, company, amount_to_sell)
# model get suggested action
predicted_actions_matrix = self.model.predict(curr_state)
model_choice = np.argmax(predicted_actions_matrix)
order_list = []
curr_action_a, curr_action_b, order_list = self.choose_actions(
stock_data_a,
stock_data_b,
portfolio,
order_list,
epsilon=self.epsilon,
model_choice=model_choice)
curr_portfolio_value = portfolio.get_value(stock_market_data)
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
# TODO Save created state, actions and portfolio value for the next call of trade()
self.last_state = curr_state
self.last_action_a = curr_action_a
self.last_action_b = curr_action_b
self.last_portfolio_value = curr_portfolio_value
return order_list
def trade(self, portfolio: Portfolio, stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# TODO Compute the current state
def get_state():
expert_vot_dict = {"BUY": 1, "SELL": 2, "HOLD": 0}
stock_data_a = stock_market_data[Company.A]
stock_data_b = stock_market_data[Company.B]
vote_a = self.expert_a.vote(stock_data_a)
vote_a = str(vote_a).split(".")[1]
vote_b = self.expert_b.vote(stock_data_b)
vote_b = str(vote_b).split(".")[1]
vote_a = expert_vot_dict[vote_a]
vote_b = expert_vot_dict[vote_b]
state = np.array([portfolio.get_stock(Company.A), portfolio.get_stock(Company.B), portfolio.cash,
portfolio.get_value(stock_market_data), vote_a, vote_b])
state = state.reshape((1, self.state_size))
return state
# TODO Store state as experience (memory) and train the neural network only if trade() was called before at least once
def train_without_experience_replay(current_state, next_state, reward, action_index):
expected_value = reward + self.gamma * np.amax(self.model.predict(next_state))
expected_value_array = self.model.predict(current_state)
expected_value_array[0][action_index] = expected_value
self.model.fit(current_state, expected_value_array, epochs=1, verbose=0)
# TODO Save created state, actions and portfolio value for the next call of trade()
def experience_replay():
batch = random.sample(self.memory, self.batch_size)
for state, action_index, reward, next_state in batch:
expected_value_array = self.model.predict(state)
expected_value = reward + self.gamma * np.amax(self.model.predict(next_state))
expected_value_array[0][action_index] = expected_value
self.model.fit(state, expected_value_array, epochs=1, verbose=0)
# TODO Create actions for current state and decrease epsilon for fewer random actions
def get_order_index(state):
self.epsilon *= self.epsilon_decay
self.epsilon = max(self.epsilon_min, self.epsilon)
if np.random.random() < self.epsilon:
return random.randrange(self.action_size)
return np.argmax(self.model.predict(state)[0])
def get_orders(order_index):
order_combination_dict = {0: ["BUY", "BUY"], 1: ["SELL", "SELL"], 2: ["BUY", "SELL"], 3: ["SELL", "BUY"]}
order_dict = {"BUY": 1, "SELL": 2}
action_combination = order_combination_dict[order_index]
order_a = order_dict[action_combination[0]]
order_b = order_dict[action_combination[1]]
order_a = Order(OrderType(order_a), Company.A, portfolio.get_stock(Company.A))
order_b = Order(OrderType(order_b), Company.B, portfolio.get_stock(Company.B))
order_list = [order_a, order_b]
return order_list
def follow_orders(orders):
order_list = []
company_list = stock_market_data.get_companies()
for company, order in zip(company_list, orders):
stock_data = stock_market_data[company]
if order.type == OrderType.BUY:
# buy as many stocks as possible
stock_price = stock_data.get_last()[-1]
amount_to_buy = int(portfolio.cash // stock_price)
logger.debug(f"{self.get_name()}: Got order to buy {company}: {amount_to_buy} shares a {stock_price}")
if amount_to_buy > 0:
order_list.append(Order(OrderType.BUY, company, amount_to_buy))
elif order.type == OrderType.SELL:
# sell as many stocks as possible
amount_to_sell = portfolio.get_stock(company)
logger.debug(f"{self.get_name()}: Got order to sell {company}: {amount_to_sell} shares available")
if amount_to_sell > 0:
order_list.append(Order(OrderType.SELL, Company.A, amount_to_sell))
return order_list
def get_reward(current_portfolio_value, next_portfolio_value):
if next_portfolio_value > current_portfolio_value:
reward = 1
return reward
elif next_portfolio_value < current_portfolio_value:
reward = -2
return reward
else:
reward = -1
return reward
self.last_state = get_state()
self.memory_state.append(self.last_state)
self.last_portfolio_value = portfolio.get_value(stock_market_data)
self.memory_portfolio.append(self.last_portfolio_value)
if self.called_once:
current_state = self.memory_state[0]
next_state = self.memory_state[1]
reward = get_reward(self.memory_portfolio[0], self.memory_portfolio[1])
self.memory.append((current_state, self.memory_action, reward, next_state))
if len(self.memory) >= self.min_size_of_memory_before_training:
experience_replay()
self.last_state = next_state
action_index = get_order_index(self.last_state)
action_list = get_orders(action_index)
action_list = follow_orders(action_list)
self.memory_action = action_index
self.called_once = True
return action_list
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# INPUT layer: 1 (buy or sell A?)
# output layer: 2 ([buy_A, sell_A])
# TODO Compute the current state
stock_data_A = stock_market_data[Company.A]
expertA_voteA = self.expert_a.vote(stock_data_A)
stock_data_B = stock_market_data[Company.B]
expertB_voteB = self.expert_b.vote(stock_data_B)
state = np.array([[
self.vote_map[expertA_voteA],
self.vote_map[expertB_voteB],
]])
# do action 0 or 1?
predictions = self.model.predict(state)
# TODO Create actions for current state and decrease epsilon for fewer random actions
if random.random() < self.epsilon:
# use random actions for A and B
action_A = random.randrange(2)
action_B = random.randrange(2)
else:
# use prediction actions
action_A = np.argmax(predictions[0][0:2])
action_B = np.argmax(predictions[0][2:4])
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
current_price_a = stock_market_data.get_most_recent_price(Company.A)
current_price_b = stock_market_data.get_most_recent_price(Company.B)
money_to_spend = portfolio.cash
order_list = []
# do stuff for A
if action_A == 0:
# buy all A
amount_to_buy = money_to_spend // current_price_a
if amount_to_buy > 0:
money_to_spend -= amount_to_buy * current_price_a
order_list.append(
Order(OrderType.BUY, Company.A, amount_to_buy))
elif action_A == 1:
# sell all A
amount_to_sell = portfolio.get_stock(Company.A)
if amount_to_sell > 0:
order_list.append(
Order(OrderType.SELL, Company.A, amount_to_sell))
else:
assert False
# do stuff for B
if action_B == 0:
# buy all B
amount_to_buy = money_to_spend // current_price_b
if amount_to_buy > 0:
order_list.append(
Order(OrderType.BUY, Company.B, amount_to_buy))
elif action_B == 1:
# sell all B
amount_to_sell = portfolio.get_stock(Company.B)
if amount_to_sell > 0:
order_list.append(
Order(OrderType.SELL, Company.B, amount_to_sell))
else:
assert False
# TODO train the neural network only if trade() was called before at least once
if self.last_state is not None:
# train
diff_a = (current_price_a / self.last_price_a - 1)
diff_b = (current_price_b / self.last_price_b - 1)
fut_reward_a = np.max(predictions[0][0:2])
fut_reward_b = np.max(predictions[0][2:4])
reward_vec = np.array([[
diff_a + self.gamma * fut_reward_a,
-diff_a + self.gamma * fut_reward_a,
diff_b + self.gamma * fut_reward_b,
-diff_b + self.gamma * fut_reward_b
]])
# TODO Store state as experience (memory) and replay
# slides: <s, a, r, s'>
# mine: <s, r>
if self.min_size_of_memory_before_training <= len(self.memory):
# take self.batch_size - 1 from memory
batch = random.sample(self.memory, self.batch_size - 1)
# append current state, reward
batch.append((self.last_state, reward_vec))
for x, y in batch:
self.model.fit(x, y, batch_size=self.batch_size, verbose=0)
else:
# only train with current (state, reward)
self.model.fit(self.last_state,
reward_vec,
batch_size=1,
verbose=0)
self.memory.append((self.last_state, reward_vec))
# TODO Save created state, actions and portfolio value for the next call of trade()
self.last_state = state
self.last_action_a = action_A
self.last_action_b = action_B
self.last_portfolio_value = portfolio.get_value(stock_market_data)
self.last_price_a = current_price_a
self.last_price_b = current_price_b
return order_list
def trade(self, portfolio: Portfolio,
stock_market_data: StockMarketData) -> List[Order]:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this traders
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
assert portfolio is not None
assert stock_market_data is not None
assert stock_market_data.get_companies() == [Company.A, Company.B]
# INPUT layer: 1 (buy or sell A?)
# output layer: 2 ([buy_A, sell_A])
# TODO Compute the current state
stock_data_a = stock_market_data[Company.A]
vote_a_for_a = self.expert_a.vote(stock_data_a)
vote_b_for_a = self.expert_b.vote(stock_data_a)
#stock_data_b = stock_market_data[Company.B]
#vote_b = self.expert_a.vote(stock_data_b)
state = np.array([[
self.vote_map[vote_a_for_a] + self.vote_map[vote_b_for_a]
]]) #, self.vote_map[vote_b]])
# do action 0 or 1?
predictions = self.model.predict(state)
action = np.argmax(predictions)
current_price_a = stock_market_data.get_most_recent_price(Company.A)
order_list = []
if action == 0:
# buy all A
amount_to_buy = portfolio.cash // current_price_a
if amount_to_buy > 0:
order_list.append(
Order(OrderType.BUY, Company.A, amount_to_buy))
elif action == 1:
# sell all A
amount_to_sell = portfolio.get_stock(Company.A)
if amount_to_sell > 0:
order_list.append(
Order(OrderType.SELL, Company.A, amount_to_sell))
else:
assert False
if self.last_state is not None:
# train
diff_a = (current_price_a / self.last_price_a - 1)
if self.last_action_a == 0:
rec_vec = np.array([[diff_a, -diff_a]])
elif self.last_action_a == 1:
rec_vec = np.array([[diff_a, -diff_a]])
else:
assert False # wtf
#reward_vec = np.array([[portfolio.get_value(stock_market_data)]])
self.model.fit(self.last_state, rec_vec)
self.last_state = state
self.last_action_a = action
self.last_portfolio_value = portfolio.get_value(stock_market_data)
self.last_price_a = current_price_a
return order_list