class BaseEnvironment(Env, ABC):
metadata = {'render.modes': ['human']}
def __init__(self,
symbol: str,
fitting_file: str,
testing_file: str,
max_position: int = 10,
window_size: int = 100,
seed: int = 1,
action_repeats: int = 5,
training: bool = True,
format_3d: bool = False,
reward_type: str = 'default',
transaction_fee: bool = True,
ema_alpha: list or float or None = EMA_ALPHA):
"""
Base class for creating environments extending OpenAI's GYM framework.
:param symbol: currency pair to trade / experiment
:param fitting_file: prior trading day (e.g., T-1)
:param testing_file: current trading day (e.g., T)
:param max_position: maximum number of positions able to hold in inventory
:param window_size: number of lags to include in observation space
:param seed: random seed number
:param action_repeats: number of steps to take in environment after a given action
:param training: if TRUE, then randomize starting point in environment
:param format_3d: if TRUE, reshape observation space from matrix to tensor
:param reward_type: method for calculating the environment's reward:
1) 'default' --> inventory count * change in midpoint price returns
2) 'default_with_fills' --> inventory count * change in midpoint price returns
+ closed trade PnL
3) 'realized_pnl' --> change in realized pnl between time steps
4) 'differential_sharpe_ratio' -->
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.1.7210&rep=rep1&type=pdf
5) 'asymmetrical' --> extended version of *default* and enhanced with a
reward for being filled above or below midpoint, and returns only
negative rewards for Unrealized PnL to discourage long-term
speculation.
6) 'trade_completion' --> reward is generated per trade's round trip
:param ema_alpha: decay factor for EMA, usually between 0.9 and 0.9999; if NONE,
raw values are returned in place of smoothed values
"""
assert reward_type in VALID_REWARD_TYPES, \
'Error: {} is not a valid reward type. Value must be in:\n{}'.format(
reward_type, VALID_REWARD_TYPES)
self.viz = Visualize(
columns=['midpoint', 'buys', 'sells', 'inventory', 'realized_pnl'],
store_historical_observations=True)
# get Broker class to keep track of PnL and orders
self.broker = Broker(max_position=max_position,
transaction_fee=transaction_fee)
# properties required for instantiation
self.symbol = symbol
self.action_repeats = action_repeats
self._seed = seed
self._random_state = np.random.RandomState(seed=self._seed)
self.training = training
self.max_position = max_position
self.window_size = window_size
self.reward_type = reward_type
self.format_3d = format_3d # e.g., [window, features, *NEW_AXIS*]
self.testing_file = testing_file
# properties that get reset()
self.reward = np.array([0.0], dtype=np.float32)
self.step_reward = np.array([0.0], dtype=np.float32)
self.done = False
self.local_step_number = 0
self.midpoint = 0.0
self.observation = None
self.action = 0
self.last_pnl = 0.
self.last_midpoint = None
self.midpoint_change = None
self.A_t, self.B_t = 0., 0. # variables for Differential Sharpe Ratio
self.episode_stats = ExperimentStatistics()
self.best_bid = self.best_ask = None
# properties to override in sub-classes
self.actions = None
self.action_space = None
self.observation_space = None
# get historical data for simulations
self.data_pipeline = DataPipeline(alpha=ema_alpha)
# three different data sets, for different purposes:
# 1) midpoint_prices - midpoint prices that have not been transformed
# 2) raw_data - raw limit order book data, not including imbalances
# 3) normalized_data - z-scored limit order book and order flow imbalance
# data, also midpoint price feature is replace by midpoint log price change
self._midpoint_prices, self._raw_data, self._normalized_data = \
self.data_pipeline.load_environment_data(
fitting_file=fitting_file,
testing_file=testing_file,
include_imbalances=True,
as_pandas=True,
)
# derive best bid and offer
self._best_bids = self._raw_data['midpoint'] - (
self._raw_data['spread'] / 2)
self._best_asks = self._raw_data['midpoint'] + (
self._raw_data['spread'] / 2)
self.max_steps = self._raw_data.shape[0] - self.action_repeats - 1
# 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, alpha=ema_alpha)))
self.rsi.add(
('rsi_{}'.format(window), RSI(window=window, alpha=ema_alpha)))
# buffer for appending lags
self.data_buffer = deque(maxlen=self.window_size)
# Index of specific data points used to generate the observation space
features = self._raw_data.columns.tolist()
self.best_bid_index = features.index('bids_distance_0')
self.best_ask_index = features.index('asks_distance_0')
self.notional_bid_index = features.index('bids_notional_0')
self.notional_ask_index = features.index('asks_notional_0')
self.buy_trade_index = features.index('buys')
self.sell_trade_index = features.index('sells')
# typecast all data sets to numpy
self._raw_data = self._raw_data.to_numpy(dtype=np.float32)
self._normalized_data = self._normalized_data.to_numpy(
dtype=np.float32)
self._midpoint_prices = self._midpoint_prices.to_numpy(
dtype=np.float64)
self._best_bids = self._best_bids.to_numpy(dtype=np.float32)
self._best_asks = self._best_asks.to_numpy(dtype=np.float32)
# rendering class
self._render = TradingGraph(sym=self.symbol)
# graph midpoint prices
self._render.reset_render_data(
y_vec=self._midpoint_prices[:np.shape(self._render.x_vec)[0]])
@abstractmethod
def map_action_to_broker(self, action: int) -> (float, float):
"""
Translate agent's action into an order and submit order to broker.
:param action: (int) agent's action for current step
:return: (tuple) reward, pnl
"""
return 0., 0.
@abstractmethod
def _create_position_features(self) -> np.ndarray:
"""
Create agent space feature set reflecting the positions held in inventory.
:return: (np.array) position features
"""
return np.array([np.nan], dtype=np.float32)
def _get_step_reward(self, step_pnl: float, step_penalty: float,
long_filled: bool, short_filled: bool) -> float:
"""
Calculate current step reward using a reward function.
:param step_pnl: PnL realized from an open position that's been closed in the
current time step
:param step_penalty: Penalty signal for agent to discourage erroneous actions
:param long_filled: TRUE if open long limit order was filled in current time step
:param short_filled: TRUE if open short limit order was filled in current time
step
:return: reward for current time step
"""
reward = 0.
if self.reward_type == 'default':
reward += reward_types.default(
inventory_count=self.broker.net_inventory_count,
midpoint_change=self.midpoint_change) * 100. + step_penalty
elif self.reward_type == 'default_with_fills':
reward += reward_types.default_with_fills(
inventory_count=self.broker.net_inventory_count,
midpoint_change=self.midpoint_change,
step_pnl=step_pnl) * 100. + step_penalty
elif self.reward_type == 'asymmetrical':
reward += reward_types.asymmetrical(
inventory_count=self.broker.net_inventory_count,
midpoint_change=self.midpoint_change,
half_spread_pct=(self.midpoint / self.best_bid) - 1.,
long_filled=long_filled,
short_filled=short_filled,
step_pnl=step_pnl,
dampening=0.6) * 100. + step_penalty
elif self.reward_type == 'realized_pnl':
current_pnl = self.broker.realized_pnl
reward += reward_types.realized_pnl(
current_pnl=current_pnl,
last_pnl=self.last_pnl) * 100. + step_penalty
self.last_pnl = current_pnl
elif self.reward_type == 'differential_sharpe_ratio':
tmp_reward, self.A_t, self.B_t = reward_types.differential_sharpe_ratio(
R_t=self.midpoint_change * self.broker.net_inventory_count,
A_tm1=self.A_t,
B_tm1=self.B_t)
reward += tmp_reward + step_penalty
elif self.reward_type == 'trade_completion':
reward += reward_types.trade_completion(
step_pnl=step_pnl,
market_order_fee=MARKET_ORDER_FEE,
profit_ratio=2.) + step_penalty
else: # Default implementation
reward += reward_types.default(
inventory_count=self.broker.net_inventory_count,
midpoint_change=self.midpoint_change) * 100. + step_penalty
return reward
def step(self, action: int = 0) -> (np.ndarray, np.ndarray, bool, dict):
"""
Step through environment with action.
:param action: (int) action to take in environment
:return: (tuple) observation, reward, is_done, and empty `dict`
"""
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 and change in midpoint price percentage
self.midpoint = self._midpoint_prices[self.local_step_number]
self.midpoint_change = (self.midpoint / self.last_midpoint) - 1.
# Pass current time step bid/ask prices to broker to calculate PnL,
# or if any open orders are to be filled
self.best_bid, self.best_ask = self._get_nbbo()
# verify the data integrity
assert self.best_bid <= self.best_ask, (
"Error: best bid is more expensive than the best Ask:"
"\nBid = {}\nAsk = {}").format(self.best_bid, self.best_ask)
# get buy and sell trade volume to use by indicators and 'broker' to
# execute any open orders the agent has
buy_volume = self._get_book_data(index=self.buy_trade_index)
sell_volume = self._get_book_data(index=self.sell_trade_index)
# Update indicators
self.tns.step(buys=buy_volume, sells=sell_volume)
self.rsi.step(price=self.midpoint)
# Get PnL from any filled LIMIT orders, which is calculated by netting out
# whatever open position the agent already has in FIFO order
limit_pnl, long_filled, short_filled = self.broker.step_limit_order_pnl(
bid_price=self.best_bid,
ask_price=self.best_ask,
buy_volume=buy_volume,
sell_volume=sell_volume,
step=self.local_step_number)
# Get PnL from any filled MARKET orders AND action penalties for invalid
# actions made by the agent for future discouragement
action_penalty_reward, market_pnl = self.map_action_to_broker(
action=step_action)
step_pnl = limit_pnl + market_pnl
self.step_reward = self._get_step_reward(
step_pnl=step_pnl,
step_penalty=action_penalty_reward,
long_filled=long_filled,
short_filled=short_filled)
# Add current step's observation to the data buffer
step_observation = self._get_step_observation(
step_action=step_action)
self.data_buffer.append(step_observation)
# Store for visualization AFTER the episode
self.viz.add_observation(obs=step_observation)
self.viz.add(
self.midpoint, # arguments map to the column names in _init_
int(long_filled),
int(short_filled),
self.broker.net_inventory_count,
(self.broker.realized_pnl * 100) / self.max_position)
self.reward += self.step_reward
self.local_step_number += 1
self.last_midpoint = self.midpoint
self.observation = self._get_observation()
if self.local_step_number > self.max_steps:
self.done = True
had_long_positions = 1 if self.broker.long_inventory_count > 0 else 0
had_short_positions = 1 if self.broker.short_inventory_count > 0 else 0
flatten_pnl = self.broker.flatten_inventory(
bid_price=self.best_bid, ask_price=self.best_ask)
self.reward += self._get_step_reward(step_pnl=flatten_pnl,
step_penalty=0.,
long_filled=False,
short_filled=False)
# store for visualization after the episode
self.viz.add(
self.midpoint, # arguments map to the column names in _init_
had_long_positions,
had_short_positions,
self.broker.net_inventory_count,
(self.broker.realized_pnl * 100) / self.max_position)
# save rewards to derive cumulative reward
self.episode_stats.reward += self.reward
return self.observation, self.reward, self.done, {}
def reset(self) -> np.ndarray:
"""
Reset the environment.
:return: (np.array) Observation at first step
"""
if self.training:
self.local_step_number = self._random_state.randint(
low=0, high=self.max_steps // 5)
else:
self.local_step_number = 0
# print out episode statistics if there was any activity by the agent
if self.broker.total_trade_count > 0 or self.broker.realized_pnl != 0.:
self.episode_stats.number_of_episodes += 1
print(('-' * 25),
'{}-{} {} EPISODE RESET'.format(self.symbol, self._seed,
self.reward_type.upper()),
('-' * 25))
print('Episode Reward: {:.4f}'.format(self.episode_stats.reward))
print('Episode PnL: {:.2f}%'.format(
(self.broker.realized_pnl / self.max_position) * 100.))
print('Trade Count: {}'.format(self.broker.total_trade_count))
print('Average PnL per Trade: {:.4f}%'.format(
self.broker.average_trade_pnl * 100.))
print('Total # of episodes: {}'.format(
self.episode_stats.number_of_episodes))
print('\n'.join([
'{}\t=\t{}'.format(k, v)
for k, v in self.broker.get_statistics().items()
]))
print('First step:\t{}'.format(self.local_step_number))
print(('=' * 75))
else:
print('Resetting environment #{} on episode #{}.'.format(
self._seed, self.episode_stats.number_of_episodes))
self.A_t, self.B_t = 0., 0.
self.reward = 0.0
self.done = False
self.broker.reset()
self.data_buffer.clear()
self.episode_stats.reset()
self.rsi.reset()
self.tns.reset()
self.viz.reset()
for step in range(self.window_size + INDICATOR_WINDOW_MAX + 1):
self.midpoint = self._midpoint_prices[self.local_step_number]
if self.last_midpoint is None:
self.last_midpoint = self.midpoint
self.midpoint_change = (self.midpoint / self.last_midpoint) - 1.
self.best_bid, self.best_ask = self._get_nbbo()
step_buy_volume = self._get_book_data(index=self.buy_trade_index)
step_sell_volume = self._get_book_data(index=self.sell_trade_index)
self.tns.step(buys=step_buy_volume, sells=step_sell_volume)
self.rsi.step(price=self.midpoint)
# Add current step's observation to the data buffer
step_observation = self._get_step_observation(step_action=0)
self.data_buffer.append(step_observation)
self.local_step_number += 1
self.last_midpoint = self.midpoint
self.observation = self._get_observation()
return self.observation
def render(self, mode: str = 'human') -> None:
"""
Render midpoint prices.
:param mode: (str) flag for type of rendering. Only 'human' supported.
:return: (void)
"""
self._render.render(midpoint=self.midpoint, mode=mode)
def close(self) -> None:
"""
Free clear memory when closing environment.
:return: (void)
"""
self.broker.reset()
self.data_buffer.clear()
self.episode_stats = None
self._raw_data = None
self._normalized_data = None
self._midpoint_prices = None
self.tns = None
self.rsi = None
def seed(self, seed: int = 1) -> list:
"""
Set random seed in environment.
:param seed: (int) random seed number
:return: (list) seed number in a list
"""
self._random_state = np.random.RandomState(seed=seed)
self._seed = seed
return [seed]
def _get_nbbo(self) -> (float, float):
"""
Get best bid and offer.
:return: (tuple) best bid and offer
"""
best_bid = self._best_bids[self.local_step_number]
best_ask = self._best_asks[self.local_step_number]
return best_bid, best_ask
def _get_book_data(self, index: int = 0) -> np.ndarray or float:
"""
Return step 'n' of order book snapshot data.
:param index: (int) step 'n' to look up in order book snapshot history
:return: (np.array) order book snapshot vector
"""
return self._raw_data[self.local_step_number][index]
@staticmethod
def _process_data(observation: np.ndarray) -> np.ndarray:
"""
Reshape observation for function approximator.
:param observation: observation space
:return: (np.array) clipped observation space
"""
return np.clip(observation, -10., 10.)
def _create_action_features(self, action: int) -> np.ndarray:
"""
Create a features array for the current time step's action.
:param action: (int) action number
:return: (np.array) One-hot of current action
"""
return self.actions[action]
def _create_indicator_features(self) -> np.ndarray:
"""
Create features vector with environment indicators.
:return: (np.array) Indicator values for current time step
"""
return np.array((*self.tns.get_value(), *self.rsi.get_value()),
dtype=np.float32).reshape(1, -1)
def _get_step_observation(self, step_action: int = 0) -> np.ndarray:
"""
Current step observation, NOT including historical data.
:param step_action: (int) current step action
:return: (np.array) Current step observation
"""
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_environment_observation = self._normalized_data[
self.local_step_number]
observation = np.concatenate(
(step_environment_observation, step_indicator_features,
step_position_features, step_action_features, self.step_reward),
axis=None)
return self._process_data(observation)
def _get_observation(self) -> np.ndarray:
"""
Current step observation, including historical data.
If format_3d is TRUE: Expand the observation space from 2 to 3 dimensions.
(note: This is necessary for conv nets in Baselines.)
:return: (np.array) Observation state for current time step
"""
# Note: reversing the data to chronological order is actually faster when
# making an array in Python / Numpy, which is odd. #timeit
observation = np.asarray(self.data_buffer, dtype=np.float32)
if self.format_3d:
observation = np.expand_dims(observation, axis=-1)
return observation
def get_trade_history(self) -> pd.DataFrame:
"""
Get DataFrame with trades from most recent episode.
:return: midpoint prices, and buy & sell trades
"""
return self.viz.to_df()
def plot_trade_history(self, save_filename: str or None = None) -> None:
"""
Plot history from back-test with trade executions, total inventory, and PnL.
:param save_filename: filename for saving the image
"""
self.viz.plot(save_filename=save_filename)
def plot_observation_history(self) -> None:
"""
Plot observation space as an image.
"""
return self.viz.plot_obs()
def __init__(self,
symbol: str,
fitting_file: str,
testing_file: str,
max_position: int = 10,
window_size: int = 100,
seed: int = 1,
action_repeats: int = 5,
training: bool = True,
format_3d: bool = False,
reward_type: str = 'default',
transaction_fee: bool = True,
ema_alpha: list or float or None = EMA_ALPHA):
"""
Base class for creating environments extending OpenAI's GYM framework.
:param symbol: currency pair to trade / experiment
:param fitting_file: prior trading day (e.g., T-1)
:param testing_file: current trading day (e.g., T)
:param max_position: maximum number of positions able to hold in inventory
:param window_size: number of lags to include in observation space
:param seed: random seed number
:param action_repeats: number of steps to take in environment after a given action
:param training: if TRUE, then randomize starting point in environment
:param format_3d: if TRUE, reshape observation space from matrix to tensor
:param reward_type: method for calculating the environment's reward:
1) 'default' --> inventory count * change in midpoint price returns
2) 'default_with_fills' --> inventory count * change in midpoint price returns
+ closed trade PnL
3) 'realized_pnl' --> change in realized pnl between time steps
4) 'differential_sharpe_ratio' -->
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.1.7210&rep=rep1&type=pdf
5) 'asymmetrical' --> extended version of *default* and enhanced with a
reward for being filled above or below midpoint, and returns only
negative rewards for Unrealized PnL to discourage long-term
speculation.
6) 'trade_completion' --> reward is generated per trade's round trip
:param ema_alpha: decay factor for EMA, usually between 0.9 and 0.9999; if NONE,
raw values are returned in place of smoothed values
"""
assert reward_type in VALID_REWARD_TYPES, \
'Error: {} is not a valid reward type. Value must be in:\n{}'.format(
reward_type, VALID_REWARD_TYPES)
self.viz = Visualize(
columns=['midpoint', 'buys', 'sells', 'inventory', 'realized_pnl'],
store_historical_observations=True)
# get Broker class to keep track of PnL and orders
self.broker = Broker(max_position=max_position,
transaction_fee=transaction_fee)
# properties required for instantiation
self.symbol = symbol
self.action_repeats = action_repeats
self._seed = seed
self._random_state = np.random.RandomState(seed=self._seed)
self.training = training
self.max_position = max_position
self.window_size = window_size
self.reward_type = reward_type
self.format_3d = format_3d # e.g., [window, features, *NEW_AXIS*]
self.testing_file = testing_file
# properties that get reset()
self.reward = np.array([0.0], dtype=np.float32)
self.step_reward = np.array([0.0], dtype=np.float32)
self.done = False
self.local_step_number = 0
self.midpoint = 0.0
self.observation = None
self.action = 0
self.last_pnl = 0.
self.last_midpoint = None
self.midpoint_change = None
self.A_t, self.B_t = 0., 0. # variables for Differential Sharpe Ratio
self.episode_stats = ExperimentStatistics()
self.best_bid = self.best_ask = None
# properties to override in sub-classes
self.actions = None
self.action_space = None
self.observation_space = None
# get historical data for simulations
self.data_pipeline = DataPipeline(alpha=ema_alpha)
# three different data sets, for different purposes:
# 1) midpoint_prices - midpoint prices that have not been transformed
# 2) raw_data - raw limit order book data, not including imbalances
# 3) normalized_data - z-scored limit order book and order flow imbalance
# data, also midpoint price feature is replace by midpoint log price change
self._midpoint_prices, self._raw_data, self._normalized_data = \
self.data_pipeline.load_environment_data(
fitting_file=fitting_file,
testing_file=testing_file,
include_imbalances=True,
as_pandas=True,
)
# derive best bid and offer
self._best_bids = self._raw_data['midpoint'] - (
self._raw_data['spread'] / 2)
self._best_asks = self._raw_data['midpoint'] + (
self._raw_data['spread'] / 2)
self.max_steps = self._raw_data.shape[0] - self.action_repeats - 1
# 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, alpha=ema_alpha)))
self.rsi.add(
('rsi_{}'.format(window), RSI(window=window, alpha=ema_alpha)))
# buffer for appending lags
self.data_buffer = deque(maxlen=self.window_size)
# Index of specific data points used to generate the observation space
features = self._raw_data.columns.tolist()
self.best_bid_index = features.index('bids_distance_0')
self.best_ask_index = features.index('asks_distance_0')
self.notional_bid_index = features.index('bids_notional_0')
self.notional_ask_index = features.index('asks_notional_0')
self.buy_trade_index = features.index('buys')
self.sell_trade_index = features.index('sells')
# typecast all data sets to numpy
self._raw_data = self._raw_data.to_numpy(dtype=np.float32)
self._normalized_data = self._normalized_data.to_numpy(
dtype=np.float32)
self._midpoint_prices = self._midpoint_prices.to_numpy(
dtype=np.float64)
self._best_bids = self._best_bids.to_numpy(dtype=np.float32)
self._best_asks = self._best_asks.to_numpy(dtype=np.float32)
# rendering class
self._render = TradingGraph(sym=self.symbol)
# graph midpoint prices
self._render.reset_render_data(
y_vec=self._midpoint_prices[:np.shape(self._render.x_vec)[0]])
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)
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))
class BaseEnvironment(Env, ABC):
metadata = {'render.modes': ['human']}
# Index of specific data points used to generate the observation space
# 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,
include_imbalances=True,
include_ema=False,
include_spread=True)
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')
def __init__(self,
fitting_file='BTC-USD_2019-04-07.csv.xz',
testing_file='BTC-USD_2019-04-08.csv.xz',
step_size=1,
max_position=5,
window_size=10,
seed=1,
action_repeats=10,
training=True,
format_3d=True,
z_score=True,
reward_type='default',
scale_rewards=True,
ema_alpha=EMA_ALPHA):
"""
Base class for creating environments extending OpenAI's GYM framework.
:param fitting_file: historical data used to fit environment data (i.e.,
previous trading day)
:param testing_file: historical data used in environment
:param step_size: increment size for steps (NOTE: leave a 1, otherwise market
transaction data will be overlooked)
:param max_position: maximum number of positions able to hold in inventory
:param window_size: number of lags to include in observation space
:param seed: random seed number
:param action_repeats: number of steps to take in environment after a given action
:param training: if TRUE, then randomize starting point in environment
:param format_3d: if TRUE, reshape observation space from matrix to tensor
:param z_score: if TRUE, normalize data set with Z-Score, otherwise use Min-Max
(i.e., range of 0 to 1)
:param reward_type: method for calculating the environment's reward:
1) 'trade_completion' --> reward is generated per trade's round trip
2) 'continuous_total_pnl' --> change in realized & unrealized pnl between
time steps
3) 'continuous_realized_pnl' --> change in realized pnl between time steps
4) 'continuous_unrealized_pnl' --> change in unrealized pnl between time steps
5) 'normed' --> refer to https://arxiv.org/abs/1804.04216v1
6) 'div' --> reward is generated per trade's round trip divided by
inventory count (again, refer to https://arxiv.org/abs/1804.04216v1)
7) 'asymmetrical' --> extended version of *default* and enhanced
with a reward for being filled above/below midpoint,
and returns only negative rewards for Unrealized PnL to
discourage long-term speculation.
8) 'asymmetrical_adj' --> extended version of *default* and enhanced
with a reward for being filled above/below midpoint,
and weighted up/down unrealized returns.
9) 'default' --> Pct change in Unrealized PnL + Realized PnL of
respective time step.
:param ema_alpha: decay factor for EMA, usually between 0.9 and 0.9999; if NONE,
raw values are returned in place of smoothed values
"""
# 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.reward_type = reward_type
self.format_3d = format_3d # e.g., [window, features, *NEW_AXIS*]
self.sym = testing_file[:7] # slice the CCY from the filename
self.scale_rewards = scale_rewards
# properties that get reset()
self.reward = 0.0
self.done = False
self.local_step_number = 0
self.midpoint = 0.0
self.observation = None
self.action = 0
self.last_pnl = 0.
self.last_midpoint = None
self.midpoint_change = None
# properties to override in sub-classes
self.actions = None
self.broker = None
self.action_space = None
self.observation_space = None
# get historical data for simulations
self.sim = Sim(z_score=z_score, alpha=ema_alpha)
self.prices_, self.data, self.normalized_data = self.sim.load_environment_data(
fitting_file=fitting_file,
testing_file=testing_file,
include_imbalances=True,
as_pandas=False)
self.best_bid = self.best_ask = None
self.max_steps = self.data.shape[
0] - self.step_size * self.action_repeats - 1
# 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, alpha=ema_alpha)))
self.rsi.add(
('rsi_{}'.format(window), RSI(window=window, alpha=ema_alpha)))
# conditionally load PnlNorm, since it calculates in O(n) time complexity
self.pnl_norm = PnlNorm(
window=INDICATOR_WINDOW[0],
alpha=None) if self.reward_type == 'normed' else None
# 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()
@abstractmethod
def map_action_to_broker(self, action: int):
"""
Translate agent's action into an order and submit order to broker.
:param action: (int) agent's action for current step
:return: (tuple) reward, pnl
"""
return 0., 0.
@abstractmethod
def _create_position_features(self):
"""
Create agent space feature set reflecting the positions held in inventory.
:return: (np.array) position features
"""
return np.array([np.nan], dtype=np.float32)
@staticmethod
def _trade_completion_reward(step_pnl: float):
"""
Alternate approach for reward calculation which places greater importance on
trades that have returned at least a 1:1 profit-to-loss ratio after
transaction fees.
:param step_pnl: limit order pnl and any penalties for bad actions
:return: normalized reward (-0.1 to 0.1) range, which can be scaled to
(-1, 1) in self._get_step_reward() method
"""
reward = 0.0
if step_pnl > MARKET_ORDER_FEE * 2: # e.g., 2:1 profit to loss ratio
reward += 1.0
elif step_pnl > 0.0:
reward += step_pnl
elif step_pnl < -MARKET_ORDER_FEE: # skew penalty so
reward -= 1.0
else:
reward -= step_pnl
return reward
def _asymmetrical_reward(self,
long_filled: bool,
short_filled: bool,
step_pnl: float,
dampening=0.15):
"""
Asymmetrical reward type for environments, which is derived from percentage
changes and notional values.
The inputs are as follows:
(1) Change in exposure value between time steps, in percentage terms; and,
(2) Realized PnL from a open order being filled between time steps,
in dollar terms.
:param long_filled: TRUE if long order is filled within same time step
:param short_filled: TRUE if short order is filled within same time step
:param step_pnl: limit order pnl and any penalties for bad actions
:param dampening: discount factor towards pnl change between time steps
:return: (float)
"""
exposure_change = self.broker.total_inventory_count * self.midpoint_change
long_fill_reward = short_fill_reward = 0.
if long_filled:
long_fill_reward += ((self.midpoint / self.best_bid) - 1.)
print("long_fill_reward: {:.6f}".format(long_fill_reward))
if short_filled:
short_fill_reward += ((self.best_ask / self.midpoint) - 1.)
print("short_fill_reward: {:.6f}".format(short_fill_reward))
reward = (long_fill_reward + short_fill_reward) + \
min(0., exposure_change * dampening)
if long_filled:
reward += step_pnl
if short_filled:
reward += step_pnl
return reward
def _asymmetrical_reward_adj(self,
long_filled: bool,
short_filled: bool,
step_pnl: float,
dampening=0.25):
"""
Asymmetrical reward type for environments with balanced feedback, which is
derived from percentage
changes and notional values.
The inputs are as follows:
(1) Change in exposure value between time steps, in percentage terms; and,
(2) Realized PnL from a open order being filled between time steps,
in dollar terms.
:param long_filled: TRUE if long order is filled within same time step
:param short_filled: TRUE if short order is filled within same time step
:param step_pnl: limit order pnl and any penalties for bad actions
:param dampening: discount factor towards pnl change between time steps
:return: (float)
"""
exposure_change = self.broker.total_inventory_count * self.midpoint_change
long_fill_reward = short_fill_reward = 0.
if long_filled:
long_fill_reward += ((self.midpoint / self.best_bid) - 1.)
print("long_fill_reward: {:.6f}".format(long_fill_reward))
if short_filled:
short_fill_reward += ((self.best_ask / self.midpoint) - 1.)
print("short_fill_reward: {:.6f}".format(short_fill_reward))
reward = (long_fill_reward + short_fill_reward) + \
min(0., exposure_change * (1. - dampening)*0.1) + \
max(0., exposure_change * dampening*0.1)
if long_filled:
reward += step_pnl
if short_filled:
reward += step_pnl
return reward
def _default_reward(self, long_filled: bool, short_filled: bool,
step_pnl: float):
"""
Default reward type for environments, which is derived from PnL and order
quantity.
The inputs are as follows:
(1) Change in exposure value between time steps, in dollar terms; and,
(2) Realized PnL from a open order being filled between time steps,
in dollar terms.
:param long_filled: TRUE if long order is filled within same time step
:param short_filled: TRUE if short order is filled within same time step
:param step_pnl: limit order pnl and any penalties for bad actions
:return:
"""
reward = self.broker.total_inventory_count * self.midpoint_change
if long_filled:
reward += step_pnl
if short_filled:
reward += step_pnl
return reward
def _get_step_reward(self, step_pnl: float, long_filled: bool,
short_filled: bool):
"""
Get reward for current time step.
Note: 'reward_type' is set during environment instantiation.
:param step_pnl: (float) PnL accrued from order fills at current time step
:return: (float) reward
"""
reward = 0.0
if self.reward_type == 'default': # pnl in dollar terms
reward += self._default_reward(long_filled, short_filled, step_pnl)
elif self.reward_type == 'asymmetrical':
reward += self._asymmetrical_reward(long_filled=long_filled,
short_filled=short_filled,
step_pnl=step_pnl)
elif self.reward_type == 'asymmetrical_adj':
reward += self._asymmetrical_reward_adj(long_filled=long_filled,
short_filled=short_filled,
step_pnl=step_pnl)
elif self.reward_type == 'trade_completion': # reward is [-1,1]
reward += self._trade_completion_reward(step_pnl=step_pnl)
# Note: we do not need to update last_pnl for this reward approach
elif self.reward_type == 'continuous_total_pnl': # pnl in percentage
new_pnl = self.broker.get_total_pnl(self.best_bid, self.best_ask)
difference = new_pnl - self.last_pnl # Difference in PnL over time step
# include step_pnl to net out drops in unrealized PnL from position closing
reward += difference + step_pnl
self.last_pnl = new_pnl
elif self.reward_type == 'continuous_realized_pnl':
new_pnl = self.broker.realized_pnl
reward += new_pnl - self.last_pnl # Difference in PnL
self.last_pnl = new_pnl
elif self.reward_type == 'continuous_unrealized_pnl':
new_pnl = self.broker.get_unrealized_pnl(self.best_bid,
self.best_ask)
difference = new_pnl - self.last_pnl # Difference in PnL over time step
# include step_pnl to net out drops in unrealized PnL from position closing
reward += difference + step_pnl
self.last_pnl = new_pnl
elif self.reward_type == 'normed':
# refer to https://arxiv.org/abs/1804.04216v1
new_pnl = self.pnl_norm.raw_value
reward += new_pnl - self.last_pnl # Difference in PnL
self.last_pnl = new_pnl
elif self.reward_type == 'div':
reward += step_pnl / max(self.broker.total_inventory_count, 1)
else: # Default implementation
reward += self._default_reward(long_filled, short_filled, step_pnl)
if self.scale_rewards:
reward *= 100. # multiply to avoid division error
return reward
def step(self, action: int):
"""
Step through environment with action
:param action: (int) action to take in environment
:return: (tuple) observation, reward, is_done, and empty `dict`
"""
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]
self.midpoint_change = (self.midpoint / self.last_midpoint) - 1.
# Pass current time step bid/ask prices to broker to calculate PnL,
# or if any open orders are to be filled
self.best_bid, self.best_ask = self._get_nbbo()
buy_volume = self._get_book_data(BaseEnvironment.buy_trade_index)
sell_volume = self._get_book_data(BaseEnvironment.sell_trade_index)
# Update indicators
self.tns.step(buys=buy_volume, sells=sell_volume)
self.rsi.step(price=self.midpoint)
# Get PnL from any filled LIMIT orders
limit_pnl, long_filled, short_filled = self.broker.step_limit_order_pnl(
bid_price=self.best_bid,
ask_price=self.best_ask,
buy_volume=buy_volume,
sell_volume=sell_volume,
step=self.local_step_number)
# Get PnL from any filled MARKET orders AND action penalties for invalid
# actions made by the agent for future discouragement
step_reward, market_pnl = self.map_action_to_broker(
action=step_action)
step_pnl = limit_pnl + step_reward + market_pnl
# step thru pnl_norm if not None
if self.pnl_norm:
self.pnl_norm.step(pnl=self.broker.get_unrealized_pnl(
bid_price=self.best_bid, ask_price=self.best_ask))
self.reward += self._get_step_reward(step_pnl=step_pnl,
long_filled=long_filled,
short_filled=short_filled)
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.last_midpoint = self.midpoint
self.observation = self._get_observation()
if self.local_step_number > self.max_steps:
self.done = True
flatten_pnl = self.broker.flatten_inventory(
self.best_bid, self.best_ask)
self.reward += self._get_step_reward(step_pnl=flatten_pnl,
long_filled=False,
short_filled=False)
return self.observation, self.reward, self.done, {}
def reset(self):
"""
Reset the environment.
:return: (np.array) Observation at first step
"""
if self.training:
self.local_step_number = self._random_state.randint(
low=0, high=self.data.shape[0] // 5)
else:
self.local_step_number = 0
msg = (' {}-{} reset. Episode pnl: {:.4f} with {} trades. '
'Avg. Trade PnL: {:.4f}. First step: {}').format(
self.sym, self._seed, self.broker.realized_pnl,
self.broker.total_trade_count,
self.broker.average_trade_pnl, self.local_step_number)
print(msg)
self.reward = 0.0
self.done = False
self.broker.reset()
self.data_buffer.clear()
self.rsi.reset()
self.tns.reset()
if self.pnl_norm:
self.pnl_norm.reset()
for step in range(self.window_size + INDICATOR_WINDOW_MAX + 1):
self.midpoint = self.prices_[self.local_step_number]
self.best_bid, self.best_ask = self._get_nbbo()
step_buy_volume = self._get_book_data(
BaseEnvironment.buy_trade_index)
step_sell_volume = self._get_book_data(
BaseEnvironment.sell_trade_index)
self.tns.step(buys=step_buy_volume, sells=step_sell_volume)
self.rsi.step(price=self.midpoint)
# step thru pnl_norm if not None
if self.pnl_norm:
self.pnl_norm.step(pnl=self.broker.get_unrealized_pnl(
bid_price=self.best_bid, ask_price=self.best_ask))
step_observation = self._get_step_observation(action=0)
self.data_buffer.append(step_observation)
self.local_step_number += self.step_size
self.last_midpoint = self.midpoint
if len(self.data_buffer) > self.window_size:
del self.data_buffer[0]
self.midpoint_change = (self.midpoint / self.last_midpoint) - 1.
self.observation = self._get_observation()
return self.observation
def render(self, mode='human'):
"""
Render midpoint prices
:param mode: (str) flag for type of rendering. Only 'human' supported.
:return: (void)
"""
self._render.render(midpoint=self.midpoint, mode=mode)
def close(self):
"""
Free clear memory when closing environment
:return: (void)
"""
self.data = None
self.normalized_data = None
self.prices_ = None
self.broker.reset()
self.data_buffer.clear()
self.sim = None
self.tns = None
self.rsi = None
self.pnl_norm = None
def seed(self, seed=1):
"""
Set random seed in environment
:param seed: (int) random seed number
:return: (list) seed number in a list
"""
self._random_state = np.random.RandomState(seed=seed)
self._seed = seed
return [seed]
@staticmethod
def _process_data(_next_state):
"""
Reshape observation for function approximator
:param _next_state: observation space
:return: (np.array) clipped observation space
"""
return np.clip(_next_state.reshape((1, -1)), -10, 10)
def _create_action_features(self, action):
"""
Create a features array for the current time step's action.
:param action: (int) action number
:return: (np.array) One-hot of current action
"""
return self.actions[action]
def _create_indicator_features(self):
"""
Create features vector with environment indicators.
:return: (np.array) Indicator values for current time step
"""
return np.array((*self.tns.get_value(), *self.rsi.get_value()),
dtype=np.float32).reshape(1, -1)
def _get_nbbo(self):
"""
Get best bid and offer
:return: (tuple) best bid and offer
"""
best_bid = round(
self.midpoint -
self._get_book_data(BaseEnvironment.best_bid_index), 2)
best_ask = round(
self.midpoint +
self._get_book_data(BaseEnvironment.best_ask_index), 2)
return best_bid, best_ask
def _get_book_data(self, index=0):
"""
Return step 'n' of order book snapshot data
:param index: (int) step 'n' to look up in order book snapshot history
:return: (np.array) order book snapshot vector
"""
return self.data[self.local_step_number][index]
def _get_step_observation(self, action=0):
"""
Current step observation, NOT including historical data.
:param action: (int) current step action
:return: (np.array) Current step observation
"""
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], dtype=np.float32)),
axis=None)
def _get_observation(self):
"""
Current step observation, including historical data.
If format_3d is TRUE: Expand the observation space from 2 to 3 dimensions.
(note: This is necessary for conv nets in Baselines.)
:return: (np.array) Observation state for current time step
"""
# Note: reversing the data to chronological order is actually faster when
# making an array in Python / Numpy, which is odd. #timeit
observation = np.asarray(self.data_buffer, dtype=np.float32)
if self.format_3d:
observation = np.expand_dims(observation, axis=-1)
return observation
def __init__(self,
fitting_file='BTC-USD_2019-04-07.csv.xz',
testing_file='BTC-USD_2019-04-08.csv.xz',
step_size=1,
max_position=5,
window_size=10,
seed=1,
action_repeats=10,
training=True,
format_3d=True,
z_score=True,
reward_type='default',
scale_rewards=True,
ema_alpha=EMA_ALPHA):
"""
Base class for creating environments extending OpenAI's GYM framework.
:param fitting_file: historical data used to fit environment data (i.e.,
previous trading day)
:param testing_file: historical data used in environment
:param step_size: increment size for steps (NOTE: leave a 1, otherwise market
transaction data will be overlooked)
:param max_position: maximum number of positions able to hold in inventory
:param window_size: number of lags to include in observation space
:param seed: random seed number
:param action_repeats: number of steps to take in environment after a given action
:param training: if TRUE, then randomize starting point in environment
:param format_3d: if TRUE, reshape observation space from matrix to tensor
:param z_score: if TRUE, normalize data set with Z-Score, otherwise use Min-Max
(i.e., range of 0 to 1)
:param reward_type: method for calculating the environment's reward:
1) 'trade_completion' --> reward is generated per trade's round trip
2) 'continuous_total_pnl' --> change in realized & unrealized pnl between
time steps
3) 'continuous_realized_pnl' --> change in realized pnl between time steps
4) 'continuous_unrealized_pnl' --> change in unrealized pnl between time steps
5) 'normed' --> refer to https://arxiv.org/abs/1804.04216v1
6) 'div' --> reward is generated per trade's round trip divided by
inventory count (again, refer to https://arxiv.org/abs/1804.04216v1)
7) 'asymmetrical' --> extended version of *default* and enhanced
with a reward for being filled above/below midpoint,
and returns only negative rewards for Unrealized PnL to
discourage long-term speculation.
8) 'asymmetrical_adj' --> extended version of *default* and enhanced
with a reward for being filled above/below midpoint,
and weighted up/down unrealized returns.
9) 'default' --> Pct change in Unrealized PnL + Realized PnL of
respective time step.
:param ema_alpha: decay factor for EMA, usually between 0.9 and 0.9999; if NONE,
raw values are returned in place of smoothed values
"""
# 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.reward_type = reward_type
self.format_3d = format_3d # e.g., [window, features, *NEW_AXIS*]
self.sym = testing_file[:7] # slice the CCY from the filename
self.scale_rewards = scale_rewards
# properties that get reset()
self.reward = 0.0
self.done = False
self.local_step_number = 0
self.midpoint = 0.0
self.observation = None
self.action = 0
self.last_pnl = 0.
self.last_midpoint = None
self.midpoint_change = None
# properties to override in sub-classes
self.actions = None
self.broker = None
self.action_space = None
self.observation_space = None
# get historical data for simulations
self.sim = Sim(z_score=z_score, alpha=ema_alpha)
self.prices_, self.data, self.normalized_data = self.sim.load_environment_data(
fitting_file=fitting_file,
testing_file=testing_file,
include_imbalances=True,
as_pandas=False)
self.best_bid = self.best_ask = None
self.max_steps = self.data.shape[
0] - self.step_size * self.action_repeats - 1
# 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, alpha=ema_alpha)))
self.rsi.add(
('rsi_{}'.format(window), RSI(window=window, alpha=ema_alpha)))
# conditionally load PnlNorm, since it calculates in O(n) time complexity
self.pnl_norm = PnlNorm(
window=INDICATOR_WINDOW[0],
alpha=None) if self.reward_type == 'normed' else None
# 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()
def __init__(self,
*,
fitting_file='LTC-USD_2019-04-07.csv.xz',
testing_file='LTC-USD_2019-04-08.csv.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, dtype=np.float32)
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, z_score=z_score)
self.prices_, self.data, self.normalized_data = self.sim.load_environment_data(
fitting_file, testing_file)
self.max_steps = self.data.shape[0] - self.step_size * \
self.action_repeats - 1
# 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)))
# 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))
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 = 0.03 # 3.0% gain per episode (i.e., day)
fee = MARKET_ORDER_FEE
def __init__(self,
*,
fitting_file='LTC-USD_2019-04-07.csv.xz',
testing_file='LTC-USD_2019-04-08.csv.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, dtype=np.float32)
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, z_score=z_score)
self.prices_, self.data, self.normalized_data = self.sim.load_environment_data(
fitting_file, testing_file)
self.max_steps = self.data.shape[0] - self.step_size * \
self.action_repeats - 1
# 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)))
# 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):
"""
Reshape observation and clip outliers (values +/- 10)
:param _next_state: observation space
:return: (np.array) clipped observation space
"""
return np.clip(_next_state.reshape((1, -1)), -10., 10.)
def _send_to_broker_and_get_reward(self, action: int):
"""
Create or adjust orders per a specified action and adjust for penalties.
:param action: (int) current step's action
:return: (float) reward
"""
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) / \
self.broker.reward_scale # scale realized PnL
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) / \
self.broker.reward_scale # scale realized PnL
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):
"""
Create an array with features related to the agent's inventory
:return: (np.array) normalized position features
"""
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):
"""
Create a features array for the current time step's action.
:param action: (int) action number
:return: (np.array) One-hot of current action
"""
return self.actions[action]
def _create_indicator_features(self):
"""
Create features vector with environment indicators.
:return: (np.array) Indicator values for current time step
"""
return np.array((*self.tns.get_value(), *self.rsi.get_value()),
dtype=np.float32)
def _get_nbbo(self):
"""
Get best bid and offer
:return: (tuple) best bid and offer
"""
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 step 'n' of order book snapshot data
:param index: (int) step 'n' to look up in order book snapshot history
:return: (np.array) order book snapshot vector
"""
return self.data[self.local_step_number][index]
def _get_step_observation(self, action=0):
"""
Current step observation, NOT including historical data.
:param action: (int) current step action
:return: (np.array) Current step observation
"""
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):
"""
Current step observation, including historical data.
If format_3d is TRUE: Expand the observation space from 2 to 3 dimensions.
(note: This is necessary for conv nets in Baselines.)
:return: (np.array) Observation state for current time step
"""
observation = np.array(self.data_buffer, dtype=np.float32)
if self.format_3d:
observation = np.expand_dims(observation, axis=-1)
return observation
class MarketMaker(Env):
# gym.env required
metadata = {'render.modes': ['human']}
id = 'market-maker-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 = 0.03 # 3.0% gain per episode (i.e., day)
def __init__(self,
*,
fitting_file='LTC-USD_2019-04-07.csv.xz',
testing_file='LTC-USD_2019-04-08.csv.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(17, dtype=np.float32)
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, z_score=z_score)
self.prices_, self.data, self.normalized_data = self.sim.load_environment_data(
fitting_file, testing_file)
self.max_steps = self.data.shape[0] - self.step_size * \
self.action_repeats - 1
# 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)))
# 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(
'{} MarketMaker #{} instantiated\nself.observation_space.shape: {}'
.format(self.sym, self._seed, self.observation_space.shape))
def __str__(self):
return '{} | {}-{}'.format(MarketMaker.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
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(
action=step_action)
self.reward += step_reward
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
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 + INDICATOR_WINDOW_MAX):
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_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 MarketMaker.seed({})'.format(seed))
return [seed]
@staticmethod
def _process_data(_next_state):
"""
Reshape observation and clip outliers (values +/- 10)
:param _next_state: observation space
:return: (np.array) clipped observation space
"""
return np.clip(_next_state.reshape((1, -1)), -10., 10.)
def _send_to_broker_and_get_reward(self, action: int):
"""
Create or adjust orders per a specified action and adjust for penalties.
:param action: (int) current step's action
:return: (float) reward
"""
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):
"""
Create an array with features related to the agent's inventory
:return: (np.array) normalized position features
"""
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):
"""
Create a features array for the current time step's action.
:param action: (int) action number
:return: (np.array) One-hot of current action
"""
return self.actions[action]
def _create_indicator_features(self):
"""
Create features vector with environment indicators.
:return: (np.array) Indicator values for current time step
"""
return np.array((*self.tns.get_value(), *self.rsi.get_value()),
dtype=np.float32)
def _create_order_at_level(self,
reward: float,
discouragement: float,
level=0,
side='long'):
"""
Create a new order at a specified LOB level
:param reward: (float) current step reward
:param discouragement: (float) penalty deducted from reward for erroneous actions
:param level: (int) level in the limit order book
:param side: (str) direction of trade e.g., 'long' or 'short'
:return: (float) reward with penalties added
"""
adjustment = 1 if level > 0 else 0
if side == 'long':
best = self._get_book_data(MarketMaker.best_bid_index - level)
denormalized_best = round(self.midpoint * (best + 1), 2)
inside_best = self._get_book_data(MarketMaker.best_bid_index -
level + adjustment)
denormalized_inside_best = round(self.midpoint * (inside_best + 1),
2)
plus_one = denormalized_best + 0.01
if denormalized_inside_best == plus_one:
# stick to best bid
bid_price = denormalized_best
# since LOB is rendered as cumulative notional, deduct the prior price
# level to derive the notional value of orders ahead in the queue
bid_queue_ahead = self._get_book_data(
MarketMaker.notional_bid_index -
level) - self._get_book_data(
MarketMaker.notional_bid_index - level + adjustment)
else:
# insert a cent ahead to jump a queue
bid_price = plus_one
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 = self._get_book_data(MarketMaker.best_ask_index + level)
denormalized_best = round(self.midpoint * (best + 1), 2)
inside_best = self._get_book_data(MarketMaker.best_ask_index +
level - adjustment)
denormalized_inside_best = round(self.midpoint * (inside_best + 1),
2)
plus_one = denormalized_best + 0.01
if denormalized_inside_best == plus_one:
ask_price = denormalized_best
# since LOB is rendered as cumulative notional, deduct the prior price
# level to derive the notional value of orders ahead in the queue
ask_queue_ahead = self._get_book_data(
MarketMaker.notional_ask_index +
level) - self._get_book_data(
MarketMaker.notional_ask_index + level - adjustment)
else:
ask_price = plus_one
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):
"""
Get best bid and offer
:return: (tuple) best bid and offer
"""
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 step 'n' of order book snapshot data
:param index: (int) step 'n' to look up in order book snapshot history
:return: (np.array) order book snapshot vector
"""
return self.data[self.local_step_number][index]
def _get_step_observation(self, action=0):
"""
Current step observation, NOT including historical data.
:param action: (int) current step action
:return: (np.array) Current step observation
"""
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):
"""
Current step observation, including historical data.
If format_3d is TRUE: Expand the observation space from 2 to 3 dimensions.
(note: This is necessary for conv nets in Baselines.)
:return: (np.array) Observation state for current time step
"""
observation = np.array(self.data_buffer, dtype=np.float32)
if self.format_3d:
observation = np.expand_dims(observation, axis=-1)
return observation
def __init__(self,
fitting_file='LTC-USD_2019-04-07.csv.xz',
testing_file='LTC-USD_2019-04-08.csv.xz',
step_size=1,
max_position=5,
window_size=10,
seed=1,
action_repeats=10,
training=True,
format_3d=False,
z_score=True,
reward_type='trade_completion',
scale_rewards=True):
"""
Base class for creating environments extending OpenAI's GYM framework.
:param fitting_file: historical data used to fit environment data (i.e.,
previous trading day)
:param testing_file: historical data used in environment
:param step_size: increment size for steps (NOTE: leave a 1, otherwise market
transaction data will be overlooked)
:param max_position: maximum number of positions able to hold in inventory
:param window_size: number of lags to include in observation space
:param seed: random seed number
:param action_repeats: number of steps to take in environment after a given action
:param training: if TRUE, then randomize starting point in environment
:param format_3d: if TRUE, reshape observation space from matrix to tensor
:param z_score: if TRUE, normalize data set with Z-Score, otherwise use Min-Max
(i.e., range of 0 to 1)
:param reward_type: method for calculating the environment's reward:
1) 'trade_completion' --> reward is generated per trade's round trip
2) 'continuous_total_pnl' --> change in realized & unrealized pnl between
time steps
3) 'continuous_realized_pnl' --> change in realized pnl between time steps
4) 'continuous_unrealized_pnl' --> change in unrealized pnl between time steps
"""
# 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.reward_type = reward_type
self.format_3d = format_3d # e.g., [window, features, *NEW_AXIS*]
self.sym = testing_file[:7] # slice the CCY from the filename
self.scale_rewards = scale_rewards
# properties that get reset()
self.reward = 0.0
self.done = False
self.local_step_number = 0
self.midpoint = 0.0
self.observation = None
self.action = 0
self.last_pnl = 0.
# properties to override in sub-classes
self.actions = None
self.broker = None
self.action_space = None
self.observation_space = None
# get historical data for simulations
self.sim = Sim(use_arctic=False, z_score=z_score)
self.prices_, self.data, self.normalized_data = self.sim.load_environment_data(
fitting_file, testing_file)
self.best_bid = self.best_ask = None
self.max_steps = self.data.shape[
0] - self.step_size * self.action_repeats - 1
# 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)))
# 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()
class BaseEnvironment(Env, ABC):
metadata = {'render.modes': ['human']}
# Index of specific data points used to generate the observation space
# 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')
# Constants for scaling data
target_pnl = 0.03 # 3.0% gain per episode (i.e., day)
def __init__(self,
fitting_file='LTC-USD_2019-04-07.csv.xz',
testing_file='LTC-USD_2019-04-08.csv.xz',
step_size=1,
max_position=5,
window_size=10,
seed=1,
action_repeats=10,
training=True,
format_3d=False,
z_score=True,
reward_type='trade_completion',
scale_rewards=True):
"""
Base class for creating environments extending OpenAI's GYM framework.
:param fitting_file: historical data used to fit environment data (i.e.,
previous trading day)
:param testing_file: historical data used in environment
:param step_size: increment size for steps (NOTE: leave a 1, otherwise market
transaction data will be overlooked)
:param max_position: maximum number of positions able to hold in inventory
:param window_size: number of lags to include in observation space
:param seed: random seed number
:param action_repeats: number of steps to take in environment after a given action
:param training: if TRUE, then randomize starting point in environment
:param format_3d: if TRUE, reshape observation space from matrix to tensor
:param z_score: if TRUE, normalize data set with Z-Score, otherwise use Min-Max
(i.e., range of 0 to 1)
:param reward_type: method for calculating the environment's reward:
1) 'trade_completion' --> reward is generated per trade's round trip
2) 'continuous_total_pnl' --> change in realized & unrealized pnl between
time steps
3) 'continuous_realized_pnl' --> change in realized pnl between time steps
4) 'continuous_unrealized_pnl' --> change in unrealized pnl between time steps
"""
# 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.reward_type = reward_type
self.format_3d = format_3d # e.g., [window, features, *NEW_AXIS*]
self.sym = testing_file[:7] # slice the CCY from the filename
self.scale_rewards = scale_rewards
# properties that get reset()
self.reward = 0.0
self.done = False
self.local_step_number = 0
self.midpoint = 0.0
self.observation = None
self.action = 0
self.last_pnl = 0.
# properties to override in sub-classes
self.actions = None
self.broker = None
self.action_space = None
self.observation_space = None
# get historical data for simulations
self.sim = Sim(use_arctic=False, z_score=z_score)
self.prices_, self.data, self.normalized_data = self.sim.load_environment_data(
fitting_file, testing_file)
self.best_bid = self.best_ask = None
self.max_steps = self.data.shape[
0] - self.step_size * self.action_repeats - 1
# 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)))
# 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()
@abstractmethod
def map_action_to_broker(self, action: int):
"""
Translate agent's action into an order and submit order to broker.
:param action: (int) agent's action for current step
:return: (tuple) reward, pnl
"""
return 0., 0.
@abstractmethod
def _create_position_features(self):
"""
Create agent space feature set reflecting the positions held in inventory.
:return: (np.array) position features
"""
return np.array([np.nan], dtype=np.float32)
def _get_step_reward(self, step_pnl: float):
"""
Get reward for current time step.
Note: 'reward_type' is set during environment instantiation.
:param step_pnl: (float) PnL accrued from order fills at current time step
:return: (float) reward
"""
reward = 0.
if self.reward_type == 'trade_completion':
reward += step_pnl
# Note: we do not need to update last_pnl for this reward approach
elif self.reward_type == 'continuous_total_pnl':
new_pnl = self.broker.get_total_pnl(self.best_bid, self.best_ask)
reward += new_pnl - self.last_pnl # Difference in PnL
self.last_pnl = new_pnl
elif self.reward_type == 'continuous_realized_pnl':
new_pnl = self.broker.realized_pnl
reward += new_pnl - self.last_pnl # Difference in PnL
self.last_pnl = new_pnl
elif self.reward_type == 'continuous_unrealized_pnl':
new_pnl = self.broker.get_unrealized_pnl(self.best_bid,
self.best_ask)
reward += new_pnl - self.last_pnl # Difference in PnL
self.last_pnl = new_pnl
else:
print("_get_step_reward() Unknown reward_type: {}".format(
self.reward_type))
if self.scale_rewards:
reward /= self.broker.reward_scale
return reward
def step(self, action: int):
"""
Step through environment with action
:param action: (int) action to take in environment
:return: (tuple) observation, reward, is_done, and empty `dict`
"""
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 bid/ask prices to broker to calculate PnL,
# or if any open orders are to be filled
self.best_bid, self.best_ask = self._get_nbbo()
buy_volume = self._get_book_data(BaseEnvironment.buy_trade_index)
sell_volume = self._get_book_data(BaseEnvironment.sell_trade_index)
# Update indicators
self.tns.step(buys=buy_volume, sells=sell_volume)
self.rsi.step(price=self.midpoint)
# Get PnL from any filled LIMIT orders
limit_pnl = self.broker.step_limit_order_pnl(
bid_price=self.best_bid,
ask_price=self.best_ask,
buy_volume=buy_volume,
sell_volume=sell_volume,
step=self.local_step_number)
# Get PnL from any filled MARKET orders AND action penalties for invalid
# actions made by the agent for future discouragement
step_reward, market_pnl = self.map_action_to_broker(
action=step_action)
step_pnl = limit_pnl + step_reward + market_pnl
self.reward += self._get_step_reward(step_pnl=step_pnl)
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
flatten_pnl = self.broker.flatten_inventory(
self.best_bid, self.best_ask)
self.reward += self._get_step_reward(step_pnl=flatten_pnl)
return self.observation, self.reward, self.done, {}
def reset(self):
"""
Reset the environment.
:return: (np.array) Observation at first step
"""
if self.training:
self.local_step_number = self._random_state.randint(
low=0, 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(self.best_bid, self.best_ask),
self.broker.total_trade_count, self.local_step_number)
print(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]
self.best_bid, self.best_ask = self._get_nbbo()
step_buy_volume = self._get_book_data(
BaseEnvironment.buy_trade_index)
step_sell_volume = self._get_book_data(
BaseEnvironment.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'):
"""
Render midpoint prices
:param mode: (str) flag for type of rendering. Only 'human' supported.
:return: (void)
"""
self._render.render(midpoint=self.midpoint, mode=mode)
def close(self):
"""
Free clear memory when closing environment
:return: (void)
"""
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
def seed(self, seed=1):
"""
Set random seed in environment
:param seed: (int) random seed number
:return: (list) seed number in a list
"""
self._random_state = np.random.RandomState(seed=seed)
self._seed = seed
return [seed]
@staticmethod
def _process_data(_next_state):
"""
Reshape observation for function approximator
:param _next_state: observation space
:return: (np.array) clipped observation space
"""
return _next_state.reshape((1, -1))
def _create_action_features(self, action):
"""
Create a features array for the current time step's action.
:param action: (int) action number
:return: (np.array) One-hot of current action
"""
return self.actions[action]
def _create_indicator_features(self):
"""
Create features vector with environment indicators.
:return: (np.array) Indicator values for current time step
"""
return np.array((*self.tns.get_value(), *self.rsi.get_value()),
dtype=np.float32)
def _get_nbbo(self):
"""
Get best bid and offer
:return: (tuple) best bid and offer
"""
best_bid = round(
self.midpoint -
self._get_book_data(BaseEnvironment.best_bid_index), 2)
best_ask = round(
self.midpoint +
self._get_book_data(BaseEnvironment.best_ask_index), 2)
return best_bid, best_ask
def _get_book_data(self, index=0):
"""
Return step 'n' of order book snapshot data
:param index: (int) step 'n' to look up in order book snapshot history
:return: (np.array) order book snapshot vector
"""
return self.data[self.local_step_number][index]
def _get_step_observation(self, action=0):
"""
Current step observation, NOT including historical data.
:param action: (int) current step action
:return: (np.array) Current step observation
"""
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):
"""
Current step observation, including historical data.
If format_3d is TRUE: Expand the observation space from 2 to 3 dimensions.
(note: This is necessary for conv nets in Baselines.)
:return: (np.array) Observation state for current time step
"""
observation = np.array(self.data_buffer, dtype=np.float32)
if self.format_3d:
observation = np.expand_dims(observation, axis=-1)
return observation