def doTrade(self, portfolio: Portfolio, current_portfolio_value: float,
stock_market_data: StockMarketData) -> OrderList:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this trader
current_portfolio_value : value of Portfolio at given Momemnt
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
y_a = stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_A)
y_b = stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_B)
p_a = self.stock_a_predictor.doPredict(
stock_market_data[CompanyEnum.COMPANY_A])
p_b = self.stock_b_predictor.doPredict(
stock_market_data[CompanyEnum.COMPANY_B])
r_a = (p_a - y_a) / y_a * 100
r_b = (p_b - y_b) / y_b * 100
result = OrderList()
if (r_a < 0 and portfolio.get_amount(CompanyEnum.COMPANY_A) > 0):
result.sell(CompanyEnum.COMPANY_A,
portfolio.get_amount(CompanyEnum.COMPANY_A))
if (r_b < 0 and portfolio.get_amount(CompanyEnum.COMPANY_B) > 0):
result.sell(CompanyEnum.COMPANY_B,
portfolio.get_amount(CompanyEnum.COMPANY_B))
if (r_a <= 0 and r_b <= 0):
return result
company = CompanyEnum.COMPANY_A
if (r_b > r_a):
company = CompanyEnum.COMPANY_B
buy_amount = math.floor(
portfolio.cash / stock_market_data.get_most_recent_price(company))
if (portfolio.cash > 0 and buy_amount > 0):
result.buy(company, buy_amount)
print("Portfolio: " + str(
portfolio.total_value(
stock_market_data.get_most_recent_trade_day(),
stock_market_data)))
# TODO: implement trading logic
return result
def doTrade(self, portfolio: Portfolio, current_portfolio_value: float,
stock_market_data: StockMarketData) -> OrderList:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this trader
current_portfolio_value : value of Portfolio at given Momemnt
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
result = OrderList()
predictions = {
CompanyEnum.COMPANY_A: self.stock_a_predictor.doPredict(stock_market_data[CompanyEnum.COMPANY_A]),
CompanyEnum.COMPANY_B: self.stock_b_predictor.doPredict(stock_market_data[CompanyEnum.COMPANY_B])
}
zuwachs = {}
# sell companies which get worse
for company in CompanyEnum:
prediction = predictions[company]
anzahl = portfolio.get_amount(company)
current = stock_market_data.get_most_recent_price(company)
zuwachs[company] = prediction / current
# sell if getting worse
if anzahl > 0:
if current > prediction:
result.sell(company, anzahl)
best = sorted(zuwachs, key=zuwachs.__getitem__)[::-1]
currentCash = portfolio.cash
for b in best:
prediction = predictions[b]
current = stock_market_data.get_most_recent_price(b)
wachstum = (prediction / current)
vola = self.isVolatile(stock_market_data[b].get_values())
if wachstum > 1:
if vola or wachstum > 1.001:
count = math.floor(currentCash / current)
result.buy(b, count)
currentCash -= count * current
else:
result.sell(b, portfolio.get_amount(b))
return result
def doTrade(self, portfolio: Portfolio, current_portfolio_value: float,
stock_market_data: StockMarketData) -> OrderList:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this trader
current_portfolio_value : value of Portfolio at given Momemnt
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
order_list = OrderList()
pred_a_value = self.stock_a_predictor.doPredict(
stock_market_data[CompanyEnum.COMPANY_A])
pred_b_value = self.stock_b_predictor.doPredict(
stock_market_data[CompanyEnum.COMPANY_B])
stock_a = portfolio.get_amount(CompanyEnum.COMPANY_A)
stock_a_value = stock_market_data.get_most_recent_price(
CompanyEnum.COMPANY_A)
stock_b = portfolio.get_amount(CompanyEnum.COMPANY_B)
stock_b_value = stock_market_data.get_most_recent_price(
CompanyEnum.COMPANY_B)
increase_a = pred_a_value - stock_a_value
increase_b = pred_b_value - stock_b_value
new_cash = 0.0
if stock_a > 0 and increase_a < 0.0:
order_list.sell(CompanyEnum.COMPANY_A, stock_a)
if stock_b > 0 and increase_b < 0.0:
order_list.sell(CompanyEnum.COMPANY_B, stock_b)
if increase_a > increase_b and increase_a > 0.0:
count_a = portfolio.cash / stock_a_value
if count_a > 0:
order_list.buy(CompanyEnum.COMPANY_A, int(count_a))
elif increase_b > increase_a and increase_b > 0.0:
count_b = portfolio.cash / stock_b_value
if count_b > 0:
order_list.buy(CompanyEnum.COMPANY_B, int(count_b))
# TODO: implement trading logic
return order_list
def doTrade(self, portfolio: Portfolio, current_portfolio_value: float,
stock_market_data: StockMarketData) -> OrderList:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this trader
current_portfolio_value : value of Portfolio at given Momemnt
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
result = OrderList()
predictions = {
CompanyEnum.COMPANY_A: self.stock_a_predictor.doPredict(stock_market_data[CompanyEnum.COMPANY_A]),
CompanyEnum.COMPANY_B: self.stock_b_predictor.doPredict(stock_market_data[CompanyEnum.COMPANY_B])
}
zuwachs = {}
# sell companies which get worse
for company in CompanyEnum:
prediction = predictions[company]
anzahl = portfolio.get_amount(company)
current = stock_market_data.get_most_recent_price(company)
zuwachs[company] = prediction / current
# sell if getting worse
if anzahl > 0:
if current > prediction:
result.sell(company, anzahl)
bestCompany = None
bestZuwachs = 1
for company in CompanyEnum:
if zuwachs[company] > bestZuwachs:
bestZuwachs = zuwachs[company]
bestCompany = company
if bestCompany:
current = stock_market_data.get_most_recent_price(company)
count = math.floor(portfolio.cash / current)
result.buy(bestCompany, count)
# TODO: implement trading logic
return result
def create_order_list(self, action_a: float, action_b: float,
portfolio: Portfolio, stock_market_data: StockMarketData) -> OrderList:
"""
Take two floats between -1.0 and +1.0 (one for stock A and one for stock B) and convert them into corresponding
orders.
Args:
action_a: float between -1.0 and 1.0, representing buy(positive) / sell(negative) for Company A
action_b: float between -1.0 and 1.0, representing buy(positive) / sell(negative) for Company B
portfolio: current portfolio of this trader
stock_market_data: current stock market data
Returns:
List of corresponding orders
"""
assert -1.0 <= action_a <= +1.0 and -1.0 <= action_b <= +1.0
assert portfolio is not None and stock_market_data is not None
order_list = OrderList()
# Create orders for stock A
owned_amount_a = portfolio.get_amount(CompanyEnum.COMPANY_A)
if action_a > 0.0:
current_price = stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_A)
amount_to_buy = int(action_a * (portfolio.cash // current_price))
order_list.buy(CompanyEnum.COMPANY_A, amount_to_buy)
if action_a < 0.0 and owned_amount_a > 0:
amount_to_sell = int(abs(action_a) * owned_amount_a)
order_list.sell(CompanyEnum.COMPANY_A, amount_to_sell)
# Create orders for stock B
owned_amount_b = portfolio.get_amount(CompanyEnum.COMPANY_B)
if action_b > 0.0:
current_price = stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_B)
amount_to_buy = int(action_b * (portfolio.cash // current_price))
order_list.buy(CompanyEnum.COMPANY_B, amount_to_buy)
if action_b < 0.0 and owned_amount_b > 0:
amount_to_sell = int(abs(action_b) * owned_amount_b)
order_list.sell(CompanyEnum.COMPANY_B, amount_to_sell)
return order_list
def doTrade(self, portfolio: Portfolio, current_portfolio_value: float,
stock_market_data: StockMarketData) -> OrderList:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this trader
current_portfolio_value : value of Portfolio at given moment
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
# build current state object
stock_a_data = stock_market_data[CompanyEnum.COMPANY_A]
stock_b_data = stock_market_data[CompanyEnum.COMPANY_B]
predicted_stock_a = self.stock_a_predictor.doPredict(stock_a_data)
predicted_stock_b = self.stock_b_predictor.doPredict(stock_b_data)
current_state = State(portfolio.cash,
portfolio.get_amount(CompanyEnum.COMPANY_A),
portfolio.get_amount(CompanyEnum.COMPANY_B),
stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_A),
stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_B),
predicted_stock_a,
predicted_stock_b)
logger.debug(f"DQL Trader: Current state: {current_state}")
# Store experience and train the neural network only if doTrade was called before at least once
if self.train_while_trading and self.last_state is not None:
reward = self.calculate_reward(self.last_portfolio_value, current_portfolio_value)
memory_tuple = (self.last_state, self.last_action_a, self.last_action_b, reward, current_state)
self.memory.append(memory_tuple)
if len(self.memory) > self.batch_size + self.min_size_of_memory_before_training:
self.train_model()
# Create actions for current state and decrease epsilon for fewer random actions
(action_a, action_b) = self.get_action(current_state)
self.epsilon = max([self.epsilon_min, self.epsilon * self.epsilon_decay])
logger.debug(f"DQL Trader: Computed orders {action_a} and {action_b} with epsilon {self.epsilon}")
# Save created state, actions and portfolio value for the next call of doTrade
self.last_state, self.last_action_a, self.last_action_b = current_state, action_a, action_b
self.last_portfolio_value = current_portfolio_value
return self.create_order_list(action_a, action_b, portfolio, stock_market_data)
def doTrade(self, portfolio: Portfolio, current_portfolio_value: float,
stock_market_data: StockMarketData) -> OrderList:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this trader
current_portfolio_value : value of Portfolio at given moment
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
# TODO: Build and store current state object
s_a_current = stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_A)
s_b_current = stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_B)
pct_a = round((s_a_current * portfolio.get_amount(CompanyEnum.COMPANY_A))/portfolio.total_value(stock_market_data.get_most_recent_trade_day(), stock_market_data), 2)
pct_b = round((s_b_current * portfolio.get_amount(CompanyEnum.COMPANY_B))/portfolio.total_value(stock_market_data.get_most_recent_trade_day(), stock_market_data), 2)
pct_cash = 1 - pct_a - pct_b;
s_a_next = self.stock_a_predictor.doPredict(stock_market_data[CompanyEnum.COMPANY_A]);
s_b_next = self.stock_b_predictor.doPredict(stock_market_data[CompanyEnum.COMPANY_B]);
pct_a_diff = round((s_a_next - s_a_current) / s_a_current, 2);
pct_b_diff = round((s_b_next - s_b_current) / s_b_current, 2);
portfolio_value = portfolio.total_value(stock_market_data.get_most_recent_trade_day(), stock_market_data);
portfolio_diff = round((self.portfolio_value_prev - portfolio_value) / self.portfolio_value_prev, 2);
#r = -1;
#if (portfolio_diff < 0):
# r = 1;
#elif (portfolio_diff == 0):
# r = 0;
if (portfolio_diff < 0):
self.y_prev[0][self.idx_prev] = 1;
elif (portfolio_diff >= 0):
self.y_prev[0][self.idx_prev] = -1;
# TODO: Store experience and train the neural network only if doTrade was called before at least once
if (self.loop_value > 0):
self.model.fit(np.array([[self.pct_a_prev, self.pct_b_prev, self.pct_cash_prev, self.diff_a_prev, self.diff_b_prev]]), self.y_prev, epochs=1, batch_size=1)
# TODO: Create actions for current state and decrease epsilon for fewer random actions
res = self.model.predict(np.array([[pct_a, pct_b, pct_cash, pct_a_diff, pct_b_diff]])) # [0, 0, 1, 0]
idx = np.argmax(res);
ret = OrderList()
if (idx == 0):
ret.buy(CompanyEnum.COMPANY_A, math.floor(portfolio.cash / stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_A)))
elif (idx == 1):
ret.buy(CompanyEnum.COMPANY_B, math.floor(portfolio.cash / stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_B)))
elif (idx == 2):
ret.sell(CompanyEnum.COMPANY_A, portfolio.get_amount(CompanyEnum.COMPANY_A))
else:
ret.sell(CompanyEnum.COMPANY_B, portfolio.get_amount(CompanyEnum.COMPANY_B))
# TODO: Save created state, actions and portfolio value for the next call of doTrade
self.portfolio_value_prev = portfolio_value;
self.pct_a_prev = pct_a;
self.pct_b_prev = pct_b;
self.pct_cash_prev = pct_cash;
self.diff_a_prev = pct_a_diff;
self.diff_b_prev = pct_a_diff;
self.idx_prev = idx;
self.y_prev = res;
self.loop_value = self.loop_value + 1;
return ret
def doTrade(self, portfolio: Portfolio, current_portfolio_value: float,
stock_market_data: StockMarketData) -> OrderList:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this trader
current_portfolio_value : value of Portfolio at given moment
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
# TODO: Store experience and train the neural network only if doTrade was called before at least once
# TODO: Create actions for current state and decrease epsilon for fewer random actions
# TODO: Save created state, actions and portfolio value for the next call of doTrade
deltaA = self.stock_a_predictor.doPredict(
stock_market_data[CompanyEnum.COMPANY_A]) / stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_A)
deltaB = self.stock_b_predictor.doPredict(
stock_market_data[CompanyEnum.COMPANY_B]) / stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_B)
INPUT = numpy.asarray([[
(deltaA - 1.0) / 0.04,
(deltaB - 1.0) / 0.04,
]])
qualities = self.model.predict(INPUT)[0]
qmax = max(qualities[0], qualities[1], qualities[2])
currentValue = portfolio.total_value(stock_market_data.get_most_recent_trade_day(), stock_market_data)
if self.lastValue and self.train_while_trading:
lastReward = min(1, max(-1, (currentValue / self.lastValue - 1) / 0.04))
shouldBeQ = lastReward + GAMMA * qmax
self.lastOutput[self.lastAmax] = shouldBeQ
xtrain = [self.lastInput[0]]
ytrain = [self.lastOutput]
for m in self.memory:
xtrain.append(m[0][0])
qs = self.model.predict(m[0])[0]
qs[m[1]] = m[2] + GAMMA * qs[m[1]]
ytrain.append(qs)
self.model.fit(numpy.asarray(xtrain), numpy.asarray(ytrain))
self.memory.append([self.lastInput, self.lastAmax, lastReward])
if len(self.memory) > MEMOMRY_SIZE:
self.memory.pop(0)
self.lastValue = currentValue
self.lastInput = INPUT
self.lastOutput = qualities
result = OrderList()
actions = ["BUY_A__SELL_B", "BUY_A", "BUY_B__SELL_A", "BUY_B", "SELL_ALL"]
nextAction = None
if random.random() < self.epsilon and self.train_while_trading:
nextAction = actions[random.randint(0, self.action_size - 1)]
else:
i = 0 if qualities[0] > qualities[1] else 1
i = 2 if qualities[2] > qualities[i] else i
i = 3 if qualities[3] > qualities[i] else i
i = 4 if qualities[4] > qualities[i] else i
nextAction = actions[i]
self.epsilon = max(self.epsilon_decay * self.epsilon, self.epsilon_min)
if nextAction == "BUY_A__SELL_B":
result.sell(CompanyEnum.COMPANY_B, portfolio.get_amount(CompanyEnum.COMPANY_B))
count = math.floor(portfolio.cash / stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_A))
result.buy(CompanyEnum.COMPANY_A, count)
self.lastAmax = 0
elif nextAction == "BUY_A":
count = math.floor(portfolio.cash / stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_A))
result.buy(CompanyEnum.COMPANY_A, count)
self.lastAmax = 1
elif nextAction == "BUY_B__SELL_A":
result.sell(CompanyEnum.COMPANY_A, portfolio.get_amount(CompanyEnum.COMPANY_A))
count = math.floor(portfolio.cash / stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_B))
result.buy(CompanyEnum.COMPANY_B, count)
self.lastAmax = 2
elif nextAction == "BUY_B":
count = math.floor(portfolio.cash / stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_B))
result.buy(CompanyEnum.COMPANY_B, count)
self.lastAmax = 3
elif nextAction == "SELL_ALL":
result.sell(CompanyEnum.COMPANY_A, portfolio.get_amount(CompanyEnum.COMPANY_A))
result.sell(CompanyEnum.COMPANY_B, portfolio.get_amount(CompanyEnum.COMPANY_B))
self.lastAmax = 4
return result
def doTrade(self, portfolio: Portfolio, current_portfolio_value: float,
stock_market_data: StockMarketData) -> OrderList:
"""
Generate action to be taken on the "stock market"
Args:
portfolio : current Portfolio of this trader
current_portfolio_value : value of Portfolio at given moment
stock_market_data : StockMarketData for evaluation
Returns:
A OrderList instance, may be empty never None
"""
# TODO: Build and store current state object
## cash %, portfolio a value %, portfolio b %, pred a, pred b
account_value = portfolio.cash + current_portfolio_value
a_value = portfolio.get_amount(CompanyEnum.COMPANY_A) * stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_A)
b_value = portfolio.get_amount(CompanyEnum.COMPANY_B) * stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_B)
a_value_percent = a_value / account_value
b_value_percent = b_value / account_value
cash_percent = portfolio.cash / account_value
pred_a_value = self.stock_a_predictor.doPredict(stock_market_data[CompanyEnum.COMPANY_A])
pred_b_value = self.stock_b_predictor.doPredict(stock_market_data[CompanyEnum.COMPANY_B])
stock_a_value = stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_A)
stock_b_value = stock_market_data.get_most_recent_price(CompanyEnum.COMPANY_B)
increase_a = (pred_a_value - stock_a_value) / stock_a_value
increase_b = (pred_b_value - stock_b_value) / stock_b_value
current_status = [[cash_percent, a_value_percent, b_value_percent, increase_a, increase_b]]
np_current_status = np.array(current_status)
# TODO: Store experience and train the neural network only if doTrade was called before at least once
## calc rewards = was cash + portfolio - (old cash + old portfolio), map auf 1 0 -1, now simple:
if self.stored_action >= 0:
if current_portfolio_value > self.stored_portfolio_value:
reward = 1.0
elif current_portfolio_value < self.stored_portfolio_value:
reward = -1.0
else:
reward = 0
reward_array = [0, 0, 0]
reward_array[self.stored_action] = reward
np_reward_array = np.array([reward_array])
## train again
# print(np_reward_array)
if self.train_while_trading:
self.model.fit(self.np_previous_status, np_reward_array, epochs=1, batch_size=1, verbose=0)
# TODO: Create actions for current state and decrease epsilon for fewer random actions
action = -1
random_value = uniform(0.0, 1.0)
if random_value > self.epsilon:
action = randint(0, 2)
else:
pred = self.model.predict(np_current_status)
prediction = pred[0]
if len(prediction) != 3:
print("komische prediction")
action = 2
max = prediction[action]
for i in range(3):
if prediction[i] > max:
action = i
self.count[action] = self.count[action] + 1
# print("actions")
# print(self.count)
self.epsilon = self.epsilon * self.epsilon_decay
if self.epsilon < self.epsilon_min:
self.epsilon = self.epsilon_min
order_list = OrderList()
if action < 0 or action > 2:
print("komische action")
print(action)
else:
stock_a = portfolio.get_amount(CompanyEnum.COMPANY_A)
stock_b = portfolio.get_amount(CompanyEnum.COMPANY_B)
if action == 0:
# sell a
if stock_a > 0:
order_list.sell(CompanyEnum.COMPANY_A, stock_a)
# buy b
count_b = portfolio.cash / stock_b_value
if count_b > 0:
order_list.buy(CompanyEnum.COMPANY_B, int(count_b))
if action == 1:
# sell b
if stock_b > 0:
order_list.sell(CompanyEnum.COMPANY_B, stock_b)
# boy a
count_a = portfolio.cash / stock_a_value
if count_a > 0:
order_list.buy(CompanyEnum.COMPANY_A, int(count_a))
# TODO: Save created state, actions and portfolio value for the next call of
self.stored_action = action
self.stored_portfolio_value = current_portfolio_value
self.np_previous_status = np_current_status
## save current state as laststate, current portfolio and cash
return order_list