def run_live(self, agent, test=True):
raise_refactor()
def xform_data(self, df):
# TODO here was autoencoder, talib indicators, price-anchoring
raise_refactor()
def execute(self, action):
acc = self.acc[self.mode.value]
totals = acc.step.totals
h = self.hypers
act_pct = {0: -.02, 1: 0, 2: .02}[action]
act_btc = act_pct * (acc.step.cash if act_pct > 0 else acc.step.value)
fee = {
Exchange.cbpro:
0.0025, # https://support.cbpro.com/customer/en/portal/articles/2425097-what-are-the-fees-on-cbpro-
Exchange.KRAKEN: 0.0026 # https://www.kraken.com/en-us/help/fees
}[EXCHANGE]
# Perform the trade. In training mode, we'll let it dip into negative here, but then kill and punish below.
# In testing/live, we'll just block the trade if they can't afford it
if act_pct > 0:
if acc.step.cash < self.min_trade:
act_btc = -(self.start_cash + self.start_value)
elif act_btc < self.min_trade:
act_btc = 0
else:
acc.step.value += act_btc - act_btc * fee
acc.step.cash -= act_btc
elif act_pct < 0:
if acc.step.value < self.min_trade:
act_btc = -(self.start_cash + self.start_value)
elif abs(act_btc) < self.min_trade:
act_btc = 0
else:
acc.step.cash += abs(act_btc) - abs(act_btc) * fee
acc.step.value -= abs(act_btc)
acc.step.signals.append(float(act_btc)) # clipped signal
# acc.step.signals.append(np.sign(act_pct)) # indicates an attempted trade
# next delta. [1,2,2].pct_change() == [NaN, 1, 0]
# pct_change = self.prices_diff[acc.step.i + 1]
_, y = self.data.get_data(acc.ep.i, acc.step.i) # TODO verify
pct_change = y[self.data.target]
acc.step.value += pct_change * acc.step.value
total_now = acc.step.value + acc.step.cash
totals.trade.append(total_now)
# calculate what the reward would be "if I held", to calculate the actual reward's _advantage_ over holding
hold_before = acc.step.hold_value
acc.step.hold_value += pct_change * hold_before
totals.hold.append(acc.step.hold_value + self.start_cash)
reward = 0
acc.step.i += 1
self.data.set_cash_val(acc.ep.i, acc.step.i,
acc.step.cash / self.start_cash,
acc.step.value / self.start_value)
next_state = self.get_next_state()
terminal = int(acc.step.i + 1 >= self.EPISODE_LEN)
if acc.step.value < 0 or acc.step.cash < 0:
terminal = True
if terminal and self.mode in (Mode.TRAIN, Mode.TEST):
# We're done.
acc.step.signals.append(0) # Add one last signal (to match length)
reward = self.get_return()
if np.unique(acc.step.signals).shape[0] == 1:
reward = -(self.start_cash + self.start_value
) # slam if you don't do anything
if terminal and self.mode in (Mode.LIVE, Mode.TEST_LIVE):
raise_refactor()
# if acc.step.value <= 0 or acc.step.cash <= 0: terminal = 1
return next_state, terminal, reward
def fetch_more(self):
raise_refactor()