class PriceJump(Env):
metadata = {'render.modes': ['human']}
id = 'long-short-v0'
action_repeats = 4
inventory_features = [
'long_inventory', 'short_inventory',
'total_unrealized_and_realized_pnl', 'long_unrealized_pnl',
'short_unrealized_pnl'
]
# Turn to true if Bitifinex is in the dataset (e.g., include_bitfinex=True)
features = Sim.get_feature_labels(include_system_time=False,
include_bitfinex=False)
indicator_features = ['tns', 'rsi']
best_bid_index = features.index('coinbase-bid-distance-0')
best_ask_index = features.index('coinbase-ask-distance-0')
notional_bid_index = features.index('coinbase-bid-notional-0')
notional_ask_index = features.index('coinbase-ask-notional-0')
buy_trade_index = features.index('coinbase-buys')
sell_trade_index = features.index('coinbase-sells')
instance_count = 0
def __init__(self,
*,
training=True,
fitting_file='ETH-USD_2018-12-31.xz',
testing_file='ETH-USD_2019-01-01.xz',
step_size=1,
max_position=5,
window_size=4,
frame_stack=False):
# properties required for instantiation
PriceJump.instance_count += 1
self._seed = int(PriceJump.instance_count) # seed
self._random_state = np.random.RandomState(seed=self._seed)
self.training = training
self.step_size = step_size
self.fee = BROKER_FEE
self.max_position = max_position
self.window_size = window_size
self.frame_stack = frame_stack
self.frames_to_add = 3 if self.frame_stack else 0
self.action = 0
# derive gym.env properties
self.actions = np.eye(3)
self.sym = testing_file[:7] # slice the CCY from the filename
# properties that get reset()
self.reward = 0.0
self.done = False
self.local_step_number = 0
self.midpoint = 0.0
self.observation = None
# get Broker class to keep track of PnL and orders
self.broker = Broker(max_position=max_position)
# get historical data for simulations
self.sim = Sim(use_arctic=False)
fitting_data_filepath = '{}/data_exports/{}'.format(
self.sim.cwd, fitting_file)
data_used_in_environment = '{}/data_exports/{}'.format(
self.sim.cwd, testing_file)
# print('Fitting data: {}\nTesting Data: {}'.format(fitting_data_filepath,
# data_used_in_environment))
fitting_data = self.sim.import_csv(filename=fitting_data_filepath)
fitting_data['coinbase_midpoint'] = np.log(
fitting_data['coinbase_midpoint'].values)
fitting_data['coinbase_midpoint'] = fitting_data['coinbase_midpoint']. \
pct_change().fillna(method='bfill')
self.sim.fit_scaler(fitting_data)
del fitting_data
self.data = self.sim.import_csv(filename=data_used_in_environment)
self.prices_ = self.data[
'coinbase_midpoint'].values # used to calculate PnL
self.normalized_data = self.data.copy()
self.data = self.data.values
self.normalized_data['coinbase_midpoint'] = np.log(
self.normalized_data['coinbase_midpoint'].values)
self.normalized_data['coinbase_midpoint'] = (
self.normalized_data['coinbase_midpoint'] -
self.normalized_data['coinbase_midpoint'].shift(1)).fillna(
method='bfill')
self.tns = TnS()
self.rsi = RSI()
logger.info("Pre-scaling {}-{} data...".format(self.sym, self._seed))
self.normalized_data = self.normalized_data.apply(self.sim.z_score,
axis=1).values
logger.info("...{}-{} pre-scaling complete.".format(
self.sym, self._seed))
# rendering class
self._render = TradingGraph(sym=self.sym)
# graph midpoint prices
self._render.reset_render_data(
y_vec=self.prices_[:np.shape(self._render.x_vec)[0]])
self.data_buffer, self.frame_stacker = list(), list()
self.action_space = spaces.Discrete(len(self.actions))
variable_features_count = len(self.inventory_features) + len(self.actions) + 1 + \
len(PriceJump.indicator_features)
if self.frame_stack:
shape = (4, len(PriceJump.features) + variable_features_count,
self.window_size)
else:
shape = (self.window_size,
len(PriceJump.features) + variable_features_count)
self.observation_space = spaces.Box(low=self.data.min(),
high=self.data.max(),
shape=shape,
dtype=np.int)
print('PriceJump #{} instantiated.\nself.observation_space.shape : {}'.
format(PriceJump.instance_count, self.observation_space.shape))
def __str__(self):
return '{} | {}-{}'.format(PriceJump.id, self.sym, self._seed)
def step(self, action):
for current_step in range(PriceJump.action_repeats):
if self.done:
self.reset()
return self.observation, self.reward, self.done
# reset the reward if there ARE action repeats
if current_step == 0:
self.reward = 0.
step_action = action
else:
step_action = 0
# Get current step's midpoint
self.midpoint = self.prices_[self.local_step_number]
# Pass current time step midpoint to broker to calculate PnL,
# or if any open orders are to be filled
buy_volume = self._get_book_data(PriceJump.buy_trade_index)
sell_volume = self._get_book_data(PriceJump.sell_trade_index)
self.tns.step(buys=buy_volume, sells=sell_volume)
self.rsi.step(price=self.midpoint)
self.broker.step(midpoint=self.midpoint)
self.reward += self._send_to_broker_and_get_reward(
action=step_action)
step_position_features = self._create_position_features()
step_action_features = self._create_action_features(
action=step_action)
step_indicator_features = self._create_indicator_features()
step_observation = np.concatenate(
(self.process_data(
self.normalized_data[self.local_step_number]),
step_indicator_features, step_position_features,
step_action_features, np.array([self.reward],
dtype=np.float32)),
axis=None)
self.data_buffer.append(step_observation)
if len(self.data_buffer) >= self.window_size:
self.frame_stacker.append(
np.array(self.data_buffer, dtype=np.float32))
del self.data_buffer[0]
if len(self.frame_stacker) > self.frames_to_add + 1:
del self.frame_stacker[0]
self.local_step_number += self.step_size
self.observation = np.array(self.frame_stacker, dtype=np.float32)
# This removes a dimension to be compatible with the Keras-rl module
# because Keras-rl uses its own frame-stacker. There are future
# plans to integrate this repository with more reinforcement learning
# packages, such as baselines.
if self.frame_stack is False:
self.observation = np.squeeze(self.observation, axis=0)
if self.local_step_number > self.data.shape[0] - 40:
self.done = True
order = Order(ccy=self.sym,
side=None,
price=self.midpoint,
step=self.local_step_number)
self.reward = self.broker.flatten_inventory(order=order)
return self.observation, self.reward, self.done, {}
def reset(self):
if self.training:
self.local_step_number = self._random_state.randint(
low=1, high=self.data.shape[0] // 4)
else:
self.local_step_number = 0
logger.info(' {}-{} reset. Episode pnl: {} | First step: {}'.format(
self.sym, self._seed,
self.broker.get_total_pnl(midpoint=self.midpoint),
self.local_step_number))
self.reward = 0.0
self.done = False
self.broker.reset()
self.data_buffer.clear()
self.frame_stacker.clear()
self.rsi.reset()
self.tns.reset()
for step in range(self.window_size + self.frames_to_add +
self.tns.window):
self.midpoint = self.prices_[self.local_step_number]
step_buy_volume = self._get_book_data(PriceJump.buy_trade_index)
step_sell_volume = self._get_book_data(PriceJump.sell_trade_index)
self.tns.step(buys=step_buy_volume, sells=step_sell_volume)
self.rsi.step(price=self.midpoint)
step_position_features = self._create_position_features()
step_action_features = self._create_action_features(action=0)
step_indicator_features = self._create_indicator_features()
step_observation = np.concatenate(
(self.process_data(
self.normalized_data[self.local_step_number]),
step_indicator_features, step_position_features,
step_action_features, np.array([self.reward])),
axis=None)
self.data_buffer.append(step_observation)
self.local_step_number += self.step_size
if step >= self.window_size - 1:
self.frame_stacker.append(
np.array(self.data_buffer, dtype=np.float32))
del self.data_buffer[0]
if len(self.frame_stacker) > self.frames_to_add + 1:
del self.frame_stacker[0]
self.observation = np.array(self.frame_stacker, dtype=np.float32)
# This removes a dimension to be compatible with the Keras-rl module
# because Keras-rl uses its own frame-stacker. There are future plans
# to integrate this repository with more reinforcement learning packages,
# such as baselines.
if self.frame_stack is False:
self.observation = np.squeeze(self.observation, axis=0)
return self.observation
def render(self, mode='human'):
self._render.render(midpoint=self.midpoint, mode=mode)
def close(self):
logger.info('{}-{} is being closed.'.format(self.id, self.sym))
self.data = None
self.normalized_data = None
self.prices_ = None
self.broker = None
self.sim = None
self.data_buffer = None
self.tns = None
self.rsi = None
return
def seed(self, seed=1):
self._random_state = np.random.RandomState(seed=seed)
self._seed = seed
logger.info('PriceJump.seed({})'.format(seed))
return [seed]
@staticmethod
def process_data(_next_state):
return np.clip(_next_state.reshape((1, -1)), -10., 10.)
# def process_data(self, _next_state):
# # return self.sim.scale_state(_next_state).values.reshape((1, -1))
# return np.reshape(_next_state, (1, -1))
def _send_to_broker_and_get_reward(self, action):
reward = 0.0
discouragement = 0.000000000001
if action == 0: # do nothing
pass
elif action == 1: # buy
price_fee_adjusted = self.midpoint + (self.fee * self.midpoint)
if self.broker.short_inventory_count > 0:
order = Order(ccy=self.sym,
side='short',
price=price_fee_adjusted,
step=self.local_step_number)
self.broker.remove(order=order)
reward += self.broker.get_reward(side=order.side)
elif self.broker.long_inventory_count >= 0:
order = Order(ccy=self.sym,
side='long',
price=price_fee_adjusted,
step=self.local_step_number)
if self.broker.add(order=order) is False:
reward -= discouragement
else:
logger.info(
('gym_trading.get_reward() ' + 'Error for action #{} - ' +
'unable to place an order with broker').format(action))
elif action == 2: # sell
price_fee_adjusted = self.midpoint - (self.fee * self.midpoint)
if self.broker.long_inventory_count > 0:
order = Order(ccy=self.sym,
side='long',
price=price_fee_adjusted,
step=self.local_step_number)
self.broker.remove(order=order)
reward += self.broker.get_reward(side=order.side)
elif self.broker.short_inventory_count >= 0:
order = Order(ccy=self.sym,
side='short',
price=price_fee_adjusted,
step=self.local_step_number)
if self.broker.add(order=order) is False:
reward -= discouragement
else:
logger.info(
'gym_trading.get_reward() ' + 'Error for action #{} - ' +
'unable to place an order with broker'.format(action))
else:
logger.info(
('Unknown action to take in get_reward(): ' +
'action={} | midpoint={}').format(action, self.midpoint))
return reward
def _create_position_features(self):
return np.array(
(self.broker.long_inventory.position_count / self.max_position,
self.broker.short_inventory.position_count / self.max_position,
self.broker.get_total_pnl(midpoint=self.midpoint),
self.broker.long_inventory.get_unrealized_pnl(self.midpoint),
self.broker.short_inventory.get_unrealized_pnl(self.midpoint)))
def _create_action_features(self, action):
return self.actions[action]
def _create_indicator_features(self):
return np.array((self.tns.get_value(), self.rsi.get_value()),
dtype=np.float32)
def _get_nbbo(self):
best_bid = round(
self.midpoint - self._get_book_data(PriceJump.best_bid_index), 2)
best_ask = round(
self.midpoint + self._get_book_data(PriceJump.best_ask_index), 2)
return best_bid, best_ask
def _get_book_data(self, index=0):
return self.data[self.local_step_number][index]
class PriceJump(Env):
metadata = {'render.modes': ['human']}
id = 'long-short-v0'
# Turn to true if Bitifinex is in the dataset (e.g., include_bitfinex=True)
features = Sim.get_feature_labels(include_system_time=False,
include_bitfinex=False)
best_bid_index = features.index('coinbase-bid-distance-0')
best_ask_index = features.index('coinbase-ask-distance-0')
notional_bid_index = features.index('coinbase-bid-notional-0')
notional_ask_index = features.index('coinbase-ask-notional-0')
buy_trade_index = features.index('coinbase-buys')
sell_trade_index = features.index('coinbase-sells')
target_pnl = BROKER_FEE * 10 * 5 # e.g., 5 for max_positions
fee = BROKER_FEE
def __init__(self,
*,
fitting_file='ETH-USD_2018-12-31.xz',
testing_file='ETH-USD_2019-01-01.xz',
step_size=1,
max_position=5,
window_size=10,
seed=1,
action_repeats=10,
training=True,
format_3d=False,
z_score=True):
# properties required for instantiation
self.action_repeats = action_repeats
self._seed = seed
self._random_state = np.random.RandomState(seed=self._seed)
self.training = training
self.step_size = step_size
self.max_position = max_position
self.window_size = window_size
self.format_3d = format_3d # e.g., [window, features, *NEW_AXIS*]
self.action = 0
# derive gym.env properties
self.actions = np.eye(3)
self.sym = testing_file[:7] # slice the CCY from the filename
# properties that get reset()
self.reward = 0.0
self.done = False
self.local_step_number = 0
self.midpoint = 0.0
self.observation = None
# get Broker class to keep track of PnL and orders
self.broker = Broker(max_position=max_position)
# get historical data for simulations
self.sim = Sim(use_arctic=False)
self.data = self._load_environment_data(fitting_file, testing_file)
self.prices_ = self.data[
'coinbase_midpoint'].values # used to calculate PnL
self.normalized_data = self.data.copy()
self.data = self.data.values
self.max_steps = self.data.shape[0] - self.step_size * \
self.action_repeats - 1
# normalize midpoint data
self.normalized_data['coinbase_midpoint'] = \
np.log(self.normalized_data['coinbase_midpoint'].values)
self.normalized_data['coinbase_midpoint'] = (
self.normalized_data['coinbase_midpoint'] -
self.normalized_data['coinbase_midpoint'].shift(1)).fillna(0.)
# load indicators into the indicator manager
self.tns = IndicatorManager()
self.rsi = IndicatorManager()
for window in INDICATOR_WINDOW:
self.tns.add(('tns_{}'.format(window), TnS(window=window)))
self.rsi.add(('rsi_{}'.format(window), RSI(window=window)))
if z_score:
logger.info("Pre-scaling {}-{} data...".format(
self.sym, self._seed))
self.normalized_data = self.normalized_data.apply(self.sim.z_score,
axis=1).values
logger.info("...{}-{} pre-scaling complete.".format(
self.sym, self._seed))
else:
self.normalized_data = self.normalized_data.values
# rendering class
self._render = TradingGraph(sym=self.sym)
# graph midpoint prices
self._render.reset_render_data(
y_vec=self.prices_[:np.shape(self._render.x_vec)[0]])
# buffer for appending lags
self.data_buffer = list()
self.action_space = spaces.Discrete(len(self.actions))
self.reset() # reset to load observation.shape
self.observation_space = spaces.Box(low=-10,
high=10,
shape=self.observation.shape,
dtype=np.float32)
print(
'{} PriceJump #{} instantiated.\nself.observation_space.shape : {}'
.format(self.sym, self._seed, self.observation_space.shape))
def __str__(self):
return '{} | {}-{}'.format(PriceJump.id, self.sym, self._seed)
def step(self, action: int):
for current_step in range(self.action_repeats):
if self.done:
self.reset()
return self.observation, self.reward, self.done
# reset the reward if there ARE action repeats
if current_step == 0:
self.reward = 0.
step_action = action
else:
step_action = 0
# Get current step's midpoint
self.midpoint = self.prices_[self.local_step_number]
# Pass current time step midpoint to broker to calculate PnL,
# or if any open orders are to be filled
buy_volume = self._get_book_data(PriceJump.buy_trade_index)
sell_volume = self._get_book_data(PriceJump.sell_trade_index)
self.tns.step(buys=buy_volume, sells=sell_volume)
self.rsi.step(price=self.midpoint)
self.broker.step(midpoint=self.midpoint)
self.reward += self._send_to_broker_and_get_reward(
action=step_action)
step_observation = self._get_step_observation(action=action)
self.data_buffer.append(step_observation)
if len(self.data_buffer) > self.window_size:
del self.data_buffer[0]
self.local_step_number += self.step_size
self.observation = self._get_observation()
if self.local_step_number > self.max_steps:
self.done = True
order = Order(ccy=self.sym,
side=None,
price=self.midpoint,
step=self.local_step_number)
self.reward = self.broker.flatten_inventory(order=order)
return self.observation, self.reward, self.done, {}
def reset(self):
if self.training:
self.local_step_number = self._random_state.randint(
low=1, high=self.data.shape[0] // 4)
else:
self.local_step_number = 0
msg = ' {}-{} reset. Episode pnl: {:.4f} with {} trades | First step: {}'.format(
self.sym, self._seed,
self.broker.get_total_pnl(midpoint=self.midpoint),
self.broker.get_total_trade_count(), self.local_step_number)
logger.info(msg)
self.reward = 0.0
self.done = False
self.broker.reset()
self.data_buffer.clear()
self.rsi.reset()
self.tns.reset()
for step in range(self.window_size + INDICATOR_WINDOW_MAX):
self.midpoint = self.prices_[self.local_step_number]
step_buy_volume = self._get_book_data(PriceJump.buy_trade_index)
step_sell_volume = self._get_book_data(PriceJump.sell_trade_index)
self.tns.step(buys=step_buy_volume, sells=step_sell_volume)
self.rsi.step(price=self.midpoint)
step_observation = self._get_step_observation(action=0)
self.data_buffer.append(step_observation)
self.local_step_number += self.step_size
if len(self.data_buffer) > self.window_size:
del self.data_buffer[0]
self.observation = self._get_observation()
return self.observation
def render(self, mode='human'):
self._render.render(midpoint=self.midpoint, mode=mode)
def close(self):
logger.info('{}-{} is being closed.'.format(self.id, self.sym))
self.data = None
self.normalized_data = None
self.prices_ = None
self.broker = None
self.sim = None
self.data_buffer = None
self.tns = None
self.rsi = None
return
def seed(self, seed=1):
self._random_state = np.random.RandomState(seed=seed)
self._seed = seed
logger.info('Setting seed in PriceJump.seed({})'.format(seed))
return [seed]
@staticmethod
def _process_data(_next_state):
return np.clip(_next_state.reshape((1, -1)), -10., 10.)
# def _process_data(self, _next_state):
# # return self.sim.scale_state(_next_state).values.reshape((1, -1))
# return np.reshape(_next_state, (1, -1))
def _send_to_broker_and_get_reward(self, action):
reward = 0.0
discouragement = 0.000000000001
if action == 0: # do nothing
reward += discouragement
elif action == 1: # buy
price_fee_adjusted = self.midpoint + (PriceJump.fee *
self.midpoint)
if self.broker.short_inventory_count > 0:
order = Order(ccy=self.sym,
side='short',
price=price_fee_adjusted,
step=self.local_step_number)
self.broker.remove(order=order)
reward += self.broker.get_reward(side=order.side)
elif self.broker.long_inventory_count >= 0:
order = Order(ccy=self.sym,
side='long',
price=price_fee_adjusted,
step=self.local_step_number)
if self.broker.add(order=order) is False:
reward -= discouragement
else:
logger.info(
('gym_trading.get_reward() ' + 'Error for action #{} - ' +
'unable to place an order with broker').format(action))
elif action == 2: # sell
price_fee_adjusted = self.midpoint - (PriceJump.fee *
self.midpoint)
if self.broker.long_inventory_count > 0:
order = Order(ccy=self.sym,
side='long',
price=price_fee_adjusted,
step=self.local_step_number)
self.broker.remove(order=order)
reward += self.broker.get_reward(side=order.side)
elif self.broker.short_inventory_count >= 0:
order = Order(ccy=self.sym,
side='short',
price=price_fee_adjusted,
step=self.local_step_number)
if self.broker.add(order=order) is False:
reward -= discouragement
else:
logger.info(
'gym_trading.get_reward() ' + 'Error for action #{} - ' +
'unable to place an order with broker'.format(action))
else:
logger.info(
('Unknown action to take in get_reward(): ' +
'action={} | midpoint={}').format(action, self.midpoint))
return reward
def _create_position_features(self):
return np.array(
(self.broker.long_inventory.position_count / self.max_position,
self.broker.short_inventory.position_count / self.max_position,
self.broker.get_total_pnl(midpoint=self.midpoint) /
PriceJump.target_pnl,
self.broker.long_inventory.get_unrealized_pnl(self.midpoint) /
self.broker.reward_scale,
self.broker.short_inventory.get_unrealized_pnl(self.midpoint) /
self.broker.reward_scale),
dtype=np.float32)
def _create_action_features(self, action):
return self.actions[action]
def _create_indicator_features(self):
return np.array((*self.tns.get_value(), *self.rsi.get_value()),
dtype=np.float32)
def _get_nbbo(self):
best_bid = round(
self.midpoint - self._get_book_data(PriceJump.best_bid_index), 2)
best_ask = round(
self.midpoint + self._get_book_data(PriceJump.best_ask_index), 2)
return best_bid, best_ask
def _get_book_data(self, index=0):
return self.data[self.local_step_number][index]
def _get_step_observation(self, action=0):
step_position_features = self._create_position_features()
step_action_features = self._create_action_features(action=action)
step_indicator_features = self._create_indicator_features()
return np.concatenate(
(self._process_data(self.normalized_data[self.local_step_number]),
step_indicator_features, step_position_features,
step_action_features, np.array([self.reward])),
axis=None)
def _get_observation(self):
observation = np.array(self.data_buffer, dtype=np.float32)
# Expand the observation space from 2 to 3 dimensions.
# This is necessary for conv nets in Baselines.
if self.format_3d:
observation = np.expand_dims(observation, axis=-1)
return observation
def _load_environment_data(self, fitting_file, testing_file):
fitting_data_filepath = '{}/data_exports/{}'.format(
self.sim.cwd, fitting_file)
data_used_in_environment = '{}/data_exports/{}'.format(
self.sim.cwd, testing_file)
fitting_data = self.sim.import_csv(filename=fitting_data_filepath)
fitting_data['coinbase_midpoint'] = np.log(
fitting_data['coinbase_midpoint'].values)
fitting_data['coinbase_midpoint'] = (
fitting_data['coinbase_midpoint'] -
fitting_data['coinbase_midpoint'].shift(1)).fillna(method='bfill')
self.sim.fit_scaler(fitting_data)
del fitting_data
return self.sim.import_csv(filename=data_used_in_environment)
class PriceJump(Env):
metadata = {'render.modes': ['human']}
id = 'long-short-v0'
action_repeats = 4
def __init__(self, training=True,
fitting_file='ETH-USD_2018-12-31.xz',
testing_file='ETH-USD_2019-01-01.xz',
step_size=1,
max_position=1,
window_size=50,
seed=1,
frame_stack=False):
# properties required for instantiation
self._random_state = np.random.RandomState(seed=seed)
self._seed = seed
self.training = training
self.step_size = step_size
self.fee = BROKER_FEE
self.max_position = max_position
self.window_size = window_size
self.frame_stack = frame_stack
self.frames_to_add = 3 if self.frame_stack else 0
self.inventory_features = ['long_inventory', 'short_inventory',
'long_unrealized_pnl', 'short_unrealized_pnl']
self._action = 0
# derive gym.env properties
self.actions = ((1, 0, 0), # 0. do nothing
(0, 1, 0), # 1. buy
(0, 0, 1) # 2. sell
)
self.sym = testing_file[:7] # slice the CCY from the filename
# properties that get reset()
self.reward = 0.0
self.done = False
self._local_step_number = 0
self.midpoint = 0.0
self.observation = None
# get historical data for simulations
self.broker = Broker(max_position=max_position)
self.sim = Sim(use_arctic=False)
# Turn to true if Bitifinex is in the dataset (e.g., include_bitfinex=True)
self.features = self.sim.get_feature_labels(include_system_time=False,
include_bitfinex=False)
fitting_data_filepath = '{}/data_exports/{}'.format(self.sim.cwd, fitting_file)
data_used_in_environment = '{}/data_exports/{}'.format(self.sim.cwd, testing_file)
print('Fitting data: {}\nTesting Data: {}'.format(fitting_data_filepath,
data_used_in_environment))
self.sim.fit_scaler(self.sim.import_csv(filename=fitting_data_filepath))
self.data = self.sim.import_csv(filename=data_used_in_environment)
self.prices = self.data['coinbase_midpoint'].values
self.data = self.data.apply(self.sim.z_score, axis=1)
self.data = self.data.values
self.data_buffer, self.frame_stacker = list(), list()
self.action_space = spaces.Discrete(len(self.actions))
variable_features_count = len(self.inventory_features) + len(self.actions) + 1
if self.frame_stack is False:
shape = (len(self.features) + variable_features_count, self.window_size)
else:
shape = (len(self.features) + variable_features_count, self.window_size, 4)
self.observation_space = spaces.Box(low=self.data.min(),
high=self.data.max(),
shape=shape,
dtype=np.int)
self.reset()
# print('PriceJump instantiated. ' +
# '\nself.observation_space.shape : {}'.format(
# self.observation_space.shape))
def __str__(self):
return '{} | {}-{}'.format(PriceJump.id, self.sym, self.seed)
@property
def step_number(self):
return self._local_step_number
def step(self, action):
for current_step in range(PriceJump.action_repeats):
if self.done:
self.reset()
return self.observation, self.reward, self.done
position_features = self._create_position_features()
action_features = self._create_action_features(action=action)
self.midpoint = self.prices[self._local_step_number]
self.broker.step(midpoint=self.midpoint)
if current_step == 0:
self.reward = 0.
self.reward += self._send_to_broker_and_get_reward(action=action)
_observation = np.concatenate((self.process_data(self.data[self._local_step_number]),
position_features,
action_features,
np.array([self.reward])),
axis=None)
self.data_buffer.append(_observation)
if len(self.data_buffer) >= self.window_size:
self.frame_stacker.append(np.array(self.data_buffer, dtype=np.float32))
del self.data_buffer[0]
if len(self.frame_stacker) > self.frames_to_add + 1:
del self.frame_stacker[0]
self._local_step_number += self.step_size
# output shape is [n_features, window_size, frames_to_add] eg [40, 100, 1]
self.observation = np.array(self.frame_stacker, dtype=np.float32).transpose()
# This removes a dimension to be compatible with the Keras-rl module
# because Keras-rl uses its own frame-stacker. There are future plans to integrate
# this repository with more reinforcement learning packages, such as baselines.
if self.frame_stack is False:
self.observation = self.observation.reshape(self.observation.shape[0], -1)
if self._local_step_number > self.data.shape[0] - 8:
self.done = True
order = Order(ccy=self.sym, side=None, price=self.midpoint,
step=self._local_step_number)
self.reward = self.broker.flatten_inventory(order=order)
return self.observation, self.reward, self.done, {}
def reset(self):
if self.training:
self._local_step_number = self._random_state.randint(low=1, high=5000)
else:
self._local_step_number = 0
logger.info(' %s-%i reset. Episode pnl: %.4f | First step: %i, max_pos: %i'
% (self.sym, self._seed,
self.broker.get_total_pnl(midpoint=self.midpoint),
self._local_step_number, self.max_position))
self.reward = 0.0
self.done = False
self.broker.reset()
self.data_buffer.clear()
self.frame_stacker.clear()
for step in range(self.window_size + self.frames_to_add):
position_features = self._create_position_features()
action_features = self._create_action_features(action=0)
_observation = np.concatenate((self.process_data(self.data[self._local_step_number]),
position_features,
action_features,
np.array([self.reward])),
axis=None)
self.data_buffer.append(_observation)
self._local_step_number += self.step_size
if step >= self.window_size - 1:
self.frame_stacker.append(np.array(self.data_buffer, dtype=np.float32))
del self.data_buffer[0]
if len(self.frame_stacker) > self.frames_to_add + 1:
del self.frame_stacker[0]
# output shape is [n_features, window_size, frames_to_add] eg [40, 100, 1]
self.observation = np.array(self.frame_stacker, dtype=np.float32).transpose()
# This removes a dimension to be compatible with the Keras-rl module
# because Keras-rl uses its own frame-stacker. There are future plans to integrate
# this repository with more reinforcement learning packages, such as baselines.
if self.frame_stack is False:
self.observation = self.observation.reshape(self.observation.shape[0], -1)
return self.observation
def render(self, mode='human'):
pass
def close(self):
logger.info('{}-{} is being closed.'.format(self.id, self.sym))
self.data = None
self.broker = None
self.sim = None
self.data_buffer = None
return
def seed(self, seed=1):
self._random_state = np.random.RandomState(seed=seed)
self._seed = seed
return [seed]
@staticmethod
def process_data(_next_state):
# return self.sim.scale_state(_next_state).values.reshape((1, -1))
return np.clip(_next_state.reshape((1, -1)), -10., 10.)
def _send_to_broker_and_get_reward(self, action):
reward = 0.0
if action == 0: # do nothing
pass
elif action == 1: # buy
price_fee_adjusted = self.midpoint + (self.fee * self.midpoint)
if self.broker.short_inventory_count > 0:
order = Order(ccy=self.sym, side='short',
price=price_fee_adjusted,
step=self._local_step_number)
self.broker.remove(order=order)
reward += self.broker.get_reward(side=order.side)
elif self.broker.long_inventory_count >= 0:
order = Order(ccy=self.sym, side='long',
price=price_fee_adjusted,
step=self._local_step_number)
if self.broker.add(order=order) is False:
reward -= 0.00000001
else:
logger.warning(('gym_trading.get_reward() ' +
'Error for action #{} - ' +
'unable to place an order with broker').format(action))
elif action == 2: # sell
price_fee_adjusted = self.midpoint - (self.fee * self.midpoint)
if self.broker.long_inventory_count > 0:
order = Order(ccy=self.sym, side='long',
price=price_fee_adjusted,
step=self._local_step_number)
self.broker.remove(order=order)
reward += self.broker.get_reward(side=order.side)
elif self.broker.short_inventory_count >= 0:
order = Order(ccy=self.sym, side='short',
price=price_fee_adjusted,
step=self._local_step_number)
if self.broker.add(order=order) is False:
reward -= 0.00000001
else:
logger.warning('gym_trading.get_reward() ' +
'Error for action #{} - ' +
'unable to place an order with broker'
.format(action))
else:
logger.warning(('Unknown action to take in get_reward(): ' +
'action={} | midpoint={}').format(action, self.midpoint))
return reward
def _create_position_features(self):
return np.array((self.broker.long_inventory.position_count / self.max_position,
self.broker.short_inventory.position_count / self.max_position,
self.broker.long_inventory.get_unrealized_pnl(self.midpoint),
self.broker.short_inventory.get_unrealized_pnl(self.midpoint)))
def _create_action_features(self, action):
return np.array(self.actions[action])
class MarketMaker(Env):
# gym.env required
metadata = {'render.modes': ['human']}
id = 'market-maker-v0'
# constants
inventory_features = [
'long_inventory', 'short_inventory',
'total_unrealized_and_realized_pnl', 'long_unrealized_pnl',
'short_unrealized_pnl', 'buy_distance_to_midpoint',
'short_distance_to_midpoint', 'buy_queue_vol', 'short_queue_vol'
]
# Turn to true if Bitifinex is in the dataset (e.g., include_bitfinex=True)
features = Sim.get_feature_labels(include_system_time=False,
include_bitfinex=False)
indicator_features = ['tns', 'rsi']
best_bid_index = features.index('coinbase-bid-distance-0')
best_ask_index = features.index('coinbase-ask-distance-0')
notional_bid_index = features.index('coinbase-bid-notional-0')
notional_ask_index = features.index('coinbase-ask-notional-0')
buy_trade_index = features.index('coinbase-buys')
sell_trade_index = features.index('coinbase-sells')
target_pnl = BROKER_FEE * 10 * 5 # e.g., 5 for max_positions
def __init__(self,
*,
training=True,
fitting_file='ETH-USD_2018-12-31.xz',
testing_file='ETH-USD_2019-01-01.xz',
step_size=1,
max_position=5,
window_size=10,
seed=1,
action_repeats=10,
format_3d=False):
# properties required for instantiation
self.action_repeats = action_repeats
self._seed = seed
self._random_state = np.random.RandomState(seed=self._seed)
self.training = training
self.step_size = step_size
self.max_position = max_position
self.window_size = window_size
self.format_3d = format_3d # e.g., [window, features, *NEW_AXIS*]
self.action = 0
# derive gym.env properties
self.actions = np.eye(17)
self.sym = testing_file[:7] # slice the CCY from the filename
# properties that get reset()
self.reward = 0.0
self.done = False
self.local_step_number = 0
self.midpoint = 0.0
self.observation = None
# get Broker class to keep track of PnL and orders
self.broker = Broker(max_position=max_position)
# get historical data for simulations
self.sim = Sim(use_arctic=False)
fitting_data_filepath = '{}/data_exports/{}'.format(
self.sim.cwd, fitting_file)
data_used_in_environment = '{}/data_exports/{}'.format(
self.sim.cwd, testing_file)
# print('Fitting data: {}\nTesting Data: {}'.format(fitting_data_filepath,
# data_used_in_environment))
fitting_data = self.sim.import_csv(filename=fitting_data_filepath)
fitting_data['coinbase_midpoint'] = np.log(
fitting_data['coinbase_midpoint'].values)
fitting_data['coinbase_midpoint'] = (
fitting_data['coinbase_midpoint'] -
fitting_data['coinbase_midpoint'].shift(1)).fillna(method='bfill')
self.sim.fit_scaler(fitting_data)
del fitting_data
self.data = self.sim.import_csv(filename=data_used_in_environment)
self.prices_ = self.data[
'coinbase_midpoint'].values # used to calculate PnL
self.normalized_data = self.data.copy()
self.data = self.data.values
self.max_steps = self.data.shape[0] - self.step_size * \
self.action_repeats - 1
self.normalized_data['coinbase_midpoint'] = \
np.log(self.normalized_data['coinbase_midpoint'].values)
self.normalized_data['coinbase_midpoint'] = (
self.normalized_data['coinbase_midpoint'] -
self.normalized_data['coinbase_midpoint'].shift(1)).fillna(
method='bfill')
self.tns = TnS()
self.rsi = RSI()
logger.info("Pre-scaling {}-{} data...".format(self.sym, self._seed))
self.normalized_data = self.normalized_data.apply(self.sim.z_score,
axis=1).values
logger.info("...{}-{} pre-scaling complete.".format(
self.sym, self._seed))
# rendering class
self._render = TradingGraph(sym=self.sym)
# graph midpoint prices
self._render.reset_render_data(
y_vec=self.prices_[:np.shape(self._render.x_vec)[0]])
self.data_buffer = list()
self.action_space = spaces.Discrete(len(self.actions))
variable_features_count = len(self.inventory_features) + len(self.actions) + 1 + \
len(MarketMaker.indicator_features)
if self.format_3d:
shape = (self.window_size,
len(MarketMaker.features) + variable_features_count, 1)
else:
shape = (self.window_size,
len(MarketMaker.features) + variable_features_count)
self.observation_space = spaces.Box(low=self.data.min(),
high=self.data.max(),
shape=shape,
dtype=np.int)
print(
'MarketMaker #{} instantiated.\nself.observation_space.shape : {}'.
format(self._seed, self.observation_space.shape))
def __str__(self):
return '{} | {}-{}'.format(MarketMaker.id, self.sym, self._seed)
def step(self, action):
for current_step in range(self.action_repeats):
if self.done:
self.reset()
return self.observation, self.reward, self.done
# reset the reward if there ARE action repeats
if current_step == 0:
self.reward = 0.
step_action = action
else:
step_action = 0
# Get current step's midpoint
self.midpoint = self.prices_[self.local_step_number]
# Pass current time step midpoint to broker to calculate PnL,
# or if any open orders are to be filled
step_best_bid, step_best_ask = self._get_nbbo()
buy_volume = self._get_book_data(MarketMaker.buy_trade_index)
sell_volume = self._get_book_data(MarketMaker.sell_trade_index)
self.tns.step(buys=buy_volume, sells=sell_volume)
self.rsi.step(price=self.midpoint)
step_reward = self.broker.step(bid_price=step_best_bid,
ask_price=step_best_ask,
buy_volume=buy_volume,
sell_volume=sell_volume,
step=self.local_step_number)
self.reward += self._send_to_broker_and_get_reward(step_action)
self.reward += step_reward
step_position_features = self._create_position_features()
step_action_features = self._create_action_features(
action=step_action)
step_indicator_features = self._create_indicator_features()
step_observation = np.concatenate(
(self.process_data(
self.normalized_data[self.local_step_number]),
step_indicator_features, step_position_features,
step_action_features, np.array([self.reward],
dtype=np.float32)),
axis=None)
self.data_buffer.append(step_observation)
if len(self.data_buffer) > self.window_size:
del self.data_buffer[0]
self.local_step_number += self.step_size
self.observation = np.array(self.data_buffer, dtype=np.float32)
# Expand the observation space from 2 to 3 dimensions.
# This is necessary for conv nets in Baselines.
if self.format_3d:
self.observation = np.expand_dims(self.observation, axis=-1)
if self.local_step_number > self.max_steps:
self.done = True
self.reward += self.broker.flatten_inventory(*self._get_nbbo())
return self.observation, self.reward, self.done, {}
def reset(self):
if self.training:
self.local_step_number = self._random_state.randint(
low=1, high=self.data.shape[0] // 4)
else:
self.local_step_number = 0
msg = ' {}-{} reset. Episode pnl: {:.4f} with {} trades | First step: {}'.format(
self.sym, self._seed,
self.broker.get_total_pnl(midpoint=self.midpoint),
self.broker.get_total_trade_count(), self.local_step_number)
logger.info(msg)
self.reward = 0.0
self.done = False
self.broker.reset()
self.data_buffer.clear()
self.rsi.reset()
self.tns.reset()
for step in range(self.window_size + self.tns.window):
self.midpoint = self.prices_[self.local_step_number]
step_buy_volume = self._get_book_data(MarketMaker.buy_trade_index)
step_sell_volume = self._get_book_data(
MarketMaker.sell_trade_index)
self.tns.step(buys=step_buy_volume, sells=step_sell_volume)
self.rsi.step(price=self.midpoint)
step_position_features = self._create_position_features()
step_action_features = self._create_action_features(action=0)
step_indicator_features = self._create_indicator_features()
step_observation = np.concatenate(
(self.process_data(
self.normalized_data[self.local_step_number]),
step_indicator_features, step_position_features,
step_action_features, np.array([self.reward])),
axis=None)
self.data_buffer.append(step_observation)
self.local_step_number += self.step_size
if len(self.data_buffer) > self.window_size:
del self.data_buffer[0]
self.observation = np.array(self.data_buffer, dtype=np.float32)
# Expand the observation space from 2 to 3 dimensions.
# This is necessary for conv nets in Baselines.
if self.format_3d:
self.observation = np.expand_dims(self.observation, axis=-1)
return self.observation
def render(self, mode='human'):
self._render.render(midpoint=self.midpoint, mode=mode)
def close(self):
logger.info('{}-{} is being closed.'.format(self.id, self.sym))
self.data = None
self.normalized_data = None
self.prices_ = None
self.broker = None
self.sim = None
self.data_buffer = None
self.tns = None
self.rsi = None
return
def seed(self, seed=1):
self._random_state = np.random.RandomState(seed=seed)
self._seed = seed
logger.info('Setting seed in MarketMaker.seed({})'.format(seed))
return [seed]
@staticmethod
def process_data(_next_state):
return np.clip(_next_state.reshape((1, -1)), -10., 10.)
# def process_data(self, _next_state):
# # return self.sim.scale_state(_next_state).values.reshape((1, -1))
# return np.reshape(_next_state, (1, -1))
def _send_to_broker_and_get_reward(self, action):
reward = 0.0
discouragement = 0.000000000001
if action == 0: # do nothing
reward += discouragement
elif action == 1:
reward += self._create_order_at_level(reward,
discouragement,
level=0,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=4,
side='short')
elif action == 2:
reward += self._create_order_at_level(reward,
discouragement,
level=0,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=9,
side='short')
elif action == 3:
reward += self._create_order_at_level(reward,
discouragement,
level=0,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=14,
side='short')
elif action == 4:
reward += self._create_order_at_level(reward,
discouragement,
level=4,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=0,
side='short')
elif action == 5:
reward += self._create_order_at_level(reward,
discouragement,
level=4,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=4,
side='short')
elif action == 6:
reward += self._create_order_at_level(reward,
discouragement,
level=4,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=9,
side='short')
elif action == 7:
reward += self._create_order_at_level(reward,
discouragement,
level=4,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=14,
side='short')
elif action == 8:
reward += self._create_order_at_level(reward,
discouragement,
level=9,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=0,
side='short')
elif action == 9:
reward += self._create_order_at_level(reward,
discouragement,
level=9,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=4,
side='short')
elif action == 10:
reward += self._create_order_at_level(reward,
discouragement,
level=9,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=9,
side='short')
elif action == 11:
reward += self._create_order_at_level(reward,
discouragement,
level=9,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=14,
side='short')
elif action == 12:
reward += self._create_order_at_level(reward,
discouragement,
level=14,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=0,
side='short')
elif action == 13:
reward += self._create_order_at_level(reward,
discouragement,
level=14,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=4,
side='short')
elif action == 14:
reward += self._create_order_at_level(reward,
discouragement,
level=14,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=9,
side='short')
elif action == 15:
reward += self._create_order_at_level(reward,
discouragement,
level=14,
side='long')
reward += self._create_order_at_level(reward,
discouragement,
level=14,
side='short')
elif action == 16:
reward += self.broker.flatten_inventory(*self._get_nbbo())
else:
logger.info("L'action n'exist pas ! Il faut faire attention !")
return reward
def _create_position_features(self):
return np.array(
(self.broker.long_inventory.position_count / self.max_position,
self.broker.short_inventory.position_count / self.max_position,
self.broker.get_total_pnl(midpoint=self.midpoint) /
MarketMaker.target_pnl,
self.broker.long_inventory.get_unrealized_pnl(self.midpoint) /
self.broker.reward_scale,
self.broker.short_inventory.get_unrealized_pnl(self.midpoint) /
self.broker.reward_scale,
self.broker.get_long_order_distance_to_midpoint(
midpoint=self.midpoint),
self.broker.get_short_order_distance_to_midpoint(
midpoint=self.midpoint),
*self.broker.get_queues_ahead_features()),
dtype=np.float32)
def _create_action_features(self, action):
return self.actions[action]
def _create_indicator_features(self):
return np.array((self.tns.get_value(), self.rsi.get_value()),
dtype=np.float32)
def _create_order_at_level(self,
reward,
discouragement,
level=0,
side='long'):
adjustment = 1 if level > 0 else 0
if side == 'long':
best_bid = self._get_book_data(MarketMaker.best_bid_index + level)
above_best_bid = round(
self._get_book_data(MarketMaker.best_bid_index + level -
adjustment), 2)
price_improvement_bid = round(best_bid + 0.01, 2)
if above_best_bid == price_improvement_bid:
bid_price = round(self.midpoint - best_bid, 2)
bid_queue_ahead = self._get_book_data(
MarketMaker.notional_bid_index)
else:
bid_price = round(self.midpoint - price_improvement_bid, 2)
bid_queue_ahead = 0.
bid_order = Order(ccy=self.sym,
side='long',
price=bid_price,
step=self.local_step_number,
queue_ahead=bid_queue_ahead)
if self.broker.add(order=bid_order) is False:
reward -= discouragement
else:
reward += discouragement
if side == 'short':
best_ask = self._get_book_data(MarketMaker.best_bid_index + level)
above_best_ask = round(
self._get_book_data(MarketMaker.best_ask_index + level -
adjustment), 2)
price_improvement_ask = round(best_ask - 0.01, 2)
if above_best_ask == price_improvement_ask:
ask_price = round(self.midpoint + best_ask, 2)
ask_queue_ahead = self._get_book_data(
MarketMaker.notional_ask_index)
else:
ask_price = round(self.midpoint + price_improvement_ask, 2)
ask_queue_ahead = 0.
ask_order = Order(ccy=self.sym,
side='short',
price=ask_price,
step=self.local_step_number,
queue_ahead=ask_queue_ahead)
if self.broker.add(order=ask_order) is False:
reward -= discouragement
else:
reward += discouragement
return reward
def _get_nbbo(self):
best_bid = round(
self.midpoint - self._get_book_data(MarketMaker.best_bid_index), 2)
best_ask = round(
self.midpoint + self._get_book_data(MarketMaker.best_ask_index), 2)
return best_bid, best_ask
def _get_book_data(self, index=0):
return self.data[self.local_step_number][index]
class MarketMaker(Env):
metadata = {'render.modes': ['human']}
id = 'market-maker-v0'
action_repeats = 4
bid_price_features = ['coinbase-bid-distance-0', 'coinbase-bid-distance-1',
'coinbase-bid-distance-2', 'coinbase-bid-distance-3',
'coinbase-bid-distance-4', 'coinbase-bid-distance-5',
'coinbase-bid-distance-6', 'coinbase-bid-distance-7',
'coinbase-bid-distance-8', 'coinbase-bid-distance-9']
ask_price_features = ['coinbase-ask-distance-0', 'coinbase-ask-distance-1',
'coinbase-ask-distance-2', 'coinbase-ask-distance-3',
'coinbase-ask-distance-4', 'coinbase-ask-distance-5',
'coinbase-ask-distance-6', 'coinbase-ask-distance-7',
'coinbase-ask-distance-8', 'coinbase-ask-distance-9']
bid_notional_features = ['coinbase-bid-notional-0', 'coinbase-bid-notional-1',
'coinbase-bid-notional-2', 'coinbase-bid-notional-3',
'coinbase-bid-notional-4', 'coinbase-bid-notional-5',
'coinbase-bid-notional-6', 'coinbase-bid-notional-7',
'coinbase-bid-notional-8', 'coinbase-bid-notional-9']
ask_notional_features = ['coinbase-ask-notional-0', 'coinbase-ask-notional-1',
'coinbase-ask-notional-2', 'coinbase-ask-notional-3',
'coinbase-ask-notional-4', 'coinbase-ask-notional-5',
'coinbase-ask-notional-6', 'coinbase-ask-notional-7',
'coinbase-ask-notional-8', 'coinbase-ask-notional-9']
def __init__(self, training=True,
fitting_file='ETH-USD_2018-12-31.xz',
testing_file='ETH-USD_2019-01-01.xz',
step_size=1,
max_position=5,
window_size=50,
seed=1,
frame_stack=False):
# properties required for instantiation
self._random_state = np.random.RandomState(seed=seed)
self._seed = seed
self.training = training
self.step_size = step_size
self.fee = BROKER_FEE
self.max_position = max_position
self.window_size = window_size
self.frame_stack = frame_stack
self.frames_to_add = 3 if self.frame_stack else 0
self.inventory_features = ['long_inventory', 'short_inventory',
'long_unrealized_pnl', 'short_unrealized_pnl',
'buy_distance_to_midpoint', 'short_distance_to_midpoint']
self._action = 0
# derive gym.env properties
self.actions = np.eye(24)
self.sym = testing_file[:7] # slice the CCY from the filename
# properties that get reset()
self.reward = 0.0
self.done = False
self._local_step_number = 0
self.midpoint = 0.0
self.observation = None
# get historical data for simulations
self.broker = Broker(max_position=max_position)
self.sim = Sim(use_arctic=False)
# Turn to true if Bitifinex is in the dataset (e.g., include_bitfinex=True)
self.features = self.sim.get_feature_labels(include_system_time=False,
include_bitfinex=False)
fitting_data_filepath = '{}/data_exports/{}'.format(self.sim.cwd, fitting_file)
data_used_in_environment = '{}/data_exports/{}'.format(self.sim.cwd, testing_file)
# print('Fitting data: {}\nTesting Data: {}'.format(fitting_data_filepath,
# data_used_in_environment))
self.sim.fit_scaler(self.sim.import_csv(filename=fitting_data_filepath))
self.data = self.sim.import_csv(filename=data_used_in_environment)
self.prices = self.data['coinbase_midpoint'].values # used to calculate PnL
self.bid_prices = self.data[MarketMaker.bid_price_features].values # used for LOB placement
self.ask_prices = self.data[MarketMaker.ask_price_features].values # used for LOB placement
self.bid_notionals = self.data[MarketMaker.bid_notional_features].values # used for LOB placement
self.ask_notionals = self.data[MarketMaker.ask_notional_features].values # used for LOB placement
# self.data = self.data.apply(self.sim.z_score, axis=1)
self.data_ = self.data.copy() # used for rendering data
self.data = self.data.values # used for the observation space
# self.data = None
self.data_buffer, self.frame_stacker = list(), list()
self.action_space = spaces.Discrete(len(self.actions))
variable_features_count = len(self.inventory_features) + len(self.actions) + 1
if self.frame_stack is False:
shape = (len(self.features) + variable_features_count, self.window_size)
else:
shape = (len(self.features) + variable_features_count, self.window_size, 4)
self.observation_space = spaces.Box(low=self.data.min(),
high=self.data.max(),
shape=shape,
dtype=np.int)
# attributes for rendering
self.line1 = []
self.screen_size = 200
self.y_vec = None
self.x_vec = None
self._reset_render_data()
self.reset()
# print('MarketMaker instantiated. ' +
# '\nself.observation_space.shape : {}'.format(
# self.observation_space.shape))
def __str__(self):
return '{} | {}-{}'.format(MarketMaker.id, self.sym, self.seed)
def _reset_render_data(self):
self.x_vec = np.linspace(0, self.screen_size * 10, self.screen_size + 1)[0:-1]
self.y_vec = np.array(self.prices[:np.shape(self.x_vec)[0]])
self.line1 = []
@property
def step_number(self):
return self._local_step_number
def step(self, action_):
for current_step in range(MarketMaker.action_repeats):
if self.done:
self.reset()
return self.observation, self.reward, self.done
# reset the reward if there are action repeats
if current_step == 0:
self.reward = 0.
action = action_
else:
action = 0
# Get current step's midpoint to calculate PnL, or if
# an open order got filled.
self.midpoint = self.prices[self._local_step_number]
_step_reward = self.broker.step(
bid_price=self.midpoint - self.bid_prices[self._local_step_number][0],
ask_price=self.midpoint + self.ask_prices[self._local_step_number][0],
buy_volume=self.data[self._local_step_number][-2],
sell_volume=self.data[self._local_step_number][-1],
step=self._local_step_number
)
self.reward += self._send_to_broker_and_get_reward(action) + _step_reward
position_features = self._create_position_features()
action_features = self._create_action_features(action=action)
_observation = np.concatenate((self.process_data(self.data[self._local_step_number]),
position_features,
action_features,
np.array([self.reward])),
axis=None)
self.data_buffer.append(_observation)
if len(self.data_buffer) >= self.window_size:
self.frame_stacker.append(np.array(self.data_buffer, dtype=np.float32))
del self.data_buffer[0]
if len(self.frame_stacker) > self.frames_to_add + 1:
del self.frame_stacker[0]
self._local_step_number += self.step_size
# output shape is [n_features, window_size, frames_to_add] e.g., [40, 100, 1]
self.observation = np.array(self.frame_stacker, dtype=np.float32).transpose()
# This removes a dimension to be compatible with the Keras-rl module
# because Keras-rl uses its own frame-stacker. There are future plans to integrate
# this repository with more reinforcement learning packages, such as baselines.
if self.frame_stack is False:
self.observation = self.observation.reshape(self.observation.shape[0], -1)
if self._local_step_number > self.data.shape[0] - 8:
self.done = True
best_bid = round(self.midpoint + self.bid_prices[self._local_step_number][0], 2)
best_ask = round(self.midpoint + self.ask_prices[self._local_step_number][0], 2)
self.reward += self.broker.flatten_inventory(bid_price=best_bid, ask_price=best_ask)
return self.observation, self.reward, self.done, {}
def reset(self):
if self.training:
self._local_step_number = self._random_state.randint(low=1, high=5000)
else:
self._local_step_number = 0
logger.info(' {}-{} reset. Episode pnl: {} | First step: {}, max_pos: {}'.format(
self.sym, self._seed,
self.broker.get_total_pnl(midpoint=self.midpoint),
self._local_step_number, self.max_position))
self.reward = 0.0
self.done = False
self.broker.reset()
self.data_buffer.clear()
self.frame_stacker.clear()
self._reset_render_data()
for step in range(self.window_size + self.frames_to_add):
position_features = self._create_position_features()
action_features = self._create_action_features(action=0)
_observation = np.concatenate((self.process_data(self.data[self._local_step_number]),
position_features,
action_features,
np.array([self.reward])),
axis=None)
self.data_buffer.append(_observation)
self._local_step_number += self.step_size
if step >= self.window_size - 1:
self.frame_stacker.append(np.array(self.data_buffer, dtype=np.float32))
del self.data_buffer[0]
if len(self.frame_stacker) > self.frames_to_add + 1:
del self.frame_stacker[0]
# output shape is [n_features, window_size, frames_to_add] eg [40, 100, 1]
self.observation = np.array(self.frame_stacker, dtype=np.float32).transpose()
# This removes a dimension to be compatible with the Keras-rl module
# because Keras-rl uses its own frame-stacker. There are future plans to integrate
# this repository with more reinforcement learning packages, such as baselines.
if self.frame_stack is False:
self.observation = self.observation.reshape(self.observation.shape[0], -1)
return self.observation
def render(self, mode='human'):
if mode == 'human':
self.line1 = _live_plotter(self.x_vec,
self.y_vec,
self.line1,
identifier=self.sym)
self.y_vec = np.append(self.y_vec[1:], self.midpoint)
def close(self):
logger.info('{}-{} is being closed.'.format(self.id, self.sym))
self.data = None
self.broker = None
self.sim = None
self.data_buffer = None
plt.close()
return
def seed(self, seed=1):
self._random_state = np.random.RandomState(seed=seed)
self._seed = seed
return [seed]
# @staticmethod
# def process_data(_next_state):
# return np.clip(_next_state.reshape((1, -1)), -10., 10.)
def process_data(self, _next_state):
# return self.sim.scale_state(_next_state).values.reshape((1, -1))
return np.reshape(_next_state, (1, -1))
def _send_to_broker_and_get_reward(self, action):
reward = 0.0
discouragement = 0.000000000001
if action == 0: # do nothing
reward += discouragement
elif action == 1: # set bid to inside spread or [ask_price - 0.01]
best_bid = self.bid_prices[self._local_step_number][0]
best_ask = self.ask_prices[self._local_step_number][0]
price = round(max(self.midpoint - best_bid, self.midpoint + best_ask - 0.01), 2)
order = Order(ccy=self.sym, side='long', price=price, step=self._local_step_number)
if self.broker.add(order=order) is False:
reward -= discouragement
elif action == 2: # set bid to best_bid - row 0
reward = self._create_bid_order_at_level(reward, discouragement, 0)
elif action == 3: # set bid to best_bid - row 1
reward = self._create_bid_order_at_level(reward, discouragement, 1)
elif action == 4: # set bid to best_bid - row 2
reward = self._create_bid_order_at_level(reward, discouragement, 2)
elif action == 5: # set bid to best_bid - row 3
reward = self._create_bid_order_at_level(reward, discouragement, 3)
if action == 6: # set bid to best_bid - row 4
reward = self._create_bid_order_at_level(reward, discouragement, 4)
elif action == 7: # set bid to best_bid - row 5
reward = self._create_bid_order_at_level(reward, discouragement, 5)
elif action == 8: # set bid to best_bid - row 6
reward = self._create_bid_order_at_level(reward, discouragement, 6)
elif action == 9: # set bid to best_bid - row 7
reward = self._create_bid_order_at_level(reward, discouragement, 7)
if action == 10: # set bid to best_bid - row 8
reward = self._create_bid_order_at_level(reward, discouragement, 8)
elif action == 11: # set bid to best_bid - row 9
reward = self._create_bid_order_at_level(reward, discouragement, 9)
elif action == 12: # set ask to inside spread or [bid_price + 0.01]
best_bid = self.bid_prices[self._local_step_number][0]
best_ask = self.ask_prices[self._local_step_number][0]
price = round(min(best_ask + self.midpoint, self.midpoint - best_bid + 0.01), 2)
order = Order(ccy=self.sym, side='long', price=price, step=self._local_step_number)
if self.broker.add(order=order) is False:
reward -= discouragement
if action == 13: # set ask to best_bid - row 0
reward = self._create_ask_order_at_level(reward, discouragement, 0)
elif action == 14: # set ask to best_bid - row 1
reward = self._create_ask_order_at_level(reward, discouragement, 1)
elif action == 15: # set ask to best_bid - row 2
reward = self._create_ask_order_at_level(reward, discouragement, 2)
if action == 16: # set ask to best_bid - row 3
reward = self._create_ask_order_at_level(reward, discouragement, 3)
elif action == 17: # set ask to best_bid - row 4
reward = self._create_ask_order_at_level(reward, discouragement, 4)
elif action == 18: # set ask to best_bid - row 5
reward = self._create_ask_order_at_level(reward, discouragement, 5)
if action == 19: # set ask to best_bid - row 6
reward = self._create_ask_order_at_level(reward, discouragement, 6)
elif action == 20: # set ask to best_bid - row 7
reward = self._create_ask_order_at_level(reward, discouragement, 7)
elif action == 21: # set ask to best_bid - row 8
reward = self._create_ask_order_at_level(reward, discouragement, 8)
elif action == 22: # set ask to best_bid - row 9
reward = self._create_ask_order_at_level(reward, discouragement, 9)
if action == 23: # flatten all positions
best_bid = round(self.midpoint + self.bid_prices[self._local_step_number][0], 2)
best_ask = round(self.midpoint + self.ask_prices[self._local_step_number][0], 2)
reward += self.broker.flatten_inventory(bid_price=best_bid, ask_price=best_ask)
elif action == 24: #
logger.info("L'action n.25 n'exist pas ! Il faut faire attention !")
pass
return reward
def _create_position_features(self):
return np.array((self.broker.long_inventory.position_count / self.max_position,
self.broker.short_inventory.position_count / self.max_position,
self.broker.long_inventory.get_unrealized_pnl(self.midpoint),
self.broker.short_inventory.get_unrealized_pnl(self.midpoint),
self.broker.get_long_order_distance_to_midpoint(midpoint=self.midpoint),
self.broker.get_short_order_distance_to_midpoint(midpoint=self.midpoint)))
def _create_action_features(self, action):
return self.actions[action]
def _create_bid_order_at_level(self, reward, discouragement, level=0):
if level > 0:
above_best_bid = self.bid_prices[self._local_step_number][level-1]
best_bid = self.bid_prices[self._local_step_number][level]
if round(above_best_bid, 2) == round(best_bid + 0.01, 2):
price = round(self.midpoint - best_bid, 2)
queue_ahead = self.bid_notionals[self._local_step_number][level]
else:
price = round(self.midpoint - best_bid + 0.01, 2)
queue_ahead = 0.
order = Order(ccy=self.sym, side='long', price=price,
step=self._local_step_number, queue_ahead=queue_ahead)
if self.broker.add(order=order) is False:
reward -= discouragement
else:
best_bid = self.bid_prices[self._local_step_number][level]
price = round(self.midpoint - best_bid, 2)
queue_ahead = self.bid_notionals[self._local_step_number][level]
order = Order(ccy=self.sym, side='long', price=price,
step=self._local_step_number, queue_ahead=queue_ahead)
if self.broker.add(order=order) is False:
reward -= discouragement
return reward
def _create_ask_order_at_level(self, reward, discouragement, level=0):
if level > 0:
above_best_ask = self.ask_prices[self._local_step_number][level - 1]
best_ask = self.ask_prices[self._local_step_number][level]
if round(above_best_ask, 2) == round(best_ask - 0.01, 2):
price = round(best_ask + self.midpoint, 2)
queue_ahead = self.ask_notionals[self._local_step_number][level]
else:
price = round(best_ask + 0.01 + self.midpoint, 2)
queue_ahead = 0.
order = Order(ccy=self.sym, side='short', price=price,
step=self._local_step_number, queue_ahead=queue_ahead)
if self.broker.add(order=order) is False:
reward -= discouragement
else:
best_ask = self.ask_prices[self._local_step_number][level]
price = round(best_ask + self.midpoint, 2)
queue_ahead = self.ask_notionals[self._local_step_number][level]
order = Order(ccy=self.sym, side='short', price=price,
step=self._local_step_number, queue_ahead=queue_ahead)
if self.broker.add(order=order) is False:
reward -= discouragement
return reward