def __init__(self, config):
self._config = config
self._eps_schedule = LinearSchedule(self._config.eps_begin,
self._config.eps_end,
self._config.nsteps)
self._lr_schedule = LinearSchedule(self._config.lr_begin,
self._config.lr_end,
self._config.lr_nsteps)
self._oq = Order_Queue(self._config.order_path)
self._mq = Message_Queue(self._config.message_path)
self._bf = ReplayBuffer(1000000, config)
self._action_fn = self.get_action_fn()
self.build()
def main():
config = Config()
config.mode = 'test'
config.dropout = 1.0
model = Neural_DQN(config)
#model = DQN(config)
model.initialize()
oq = Order_Queue(config.order_path)
mq = Message_Queue(config.message_path)
rewards = evaluate_policy(model, oq, mq)
print(np.mean(rewards))
class model(object):
def __init__(self, config):
self._config = config
self._eps_schedule = LinearSchedule(self._config.eps_begin,
self._config.eps_end,
self._config.nsteps)
self._lr_schedule = LinearSchedule(self._config.lr_begin,
self._config.lr_end,
self._config.lr_nsteps)
self._oq = Order_Queue(self._config.order_path)
self._mq = Message_Queue(self._config.message_path)
self._bf = ReplayBuffer(1000000, config)
self._action_fn = self.get_action_fn()
self.build()
def build(self):
pass
def initialize(self):
pass
def get_random_action(self, state):
pass
def get_best_action(self, state):
### return action, q value
pass
def get_action(self, state):
if np.random.random() < self._eps_schedule.get_epsilon():
return self.get_random_action(state)[0]
else:
return self.get_best_action(state)[0]
def get_random_action_fn(self):
def random_action_fn(t, amount, state, mid_price):
action = np.random.randint(
self._config.L) # action = L for market order
price = (action -
self._config.L // 2) * self._config.base_point + mid_price
return (price, action)
return random_action_fn
def get_action_fn(self):
def action_fn(t, amount, state, mid_price):
action = self.get_action(state)
price = (action -
self._config.L // 2) * self._config.base_point + mid_price
return (price, action)
return action_fn
def pad_state(self, states, state_history):
tmp_states, tmp_its = zip(*states)
tmp_state = np.concatenate(
[np.expand_dims(state, -1) for state in tmp_states], axis=-1)
tmp_state = np.pad(tmp_state,
((0, 0), (0, 0),
(state_history - tmp_state.shape[-1], 0)),
'constant',
constant_values=0)
tmp_it = tmp_its[-1]
return ([tmp_state], [tmp_it])
def simulate_an_episode(self, amount, T, H, start_time, order_direction,
action_fn, depth):
dH = H // T
self._mq.reset()
lob_data = self._oq.create_orderbook_time(start_time, self._mq)
lob = Limit_Order_book(**lob_data,
own_amount_to_trade=0,
own_init_price=-order_direction *
Limit_Order_book._DUMMY_VARIABLE,
own_trade_type=order_direction)
rewards = []
states = []
actions = []
done_mask = []
amount_remain = amount
cum_reward = 0
for t in range(start_time, start_time + H - dH, dH):
tmp1 = 1.0 * amount_remain / amount # amount remain
tmp2 = 1.0 * (start_time + H - t) / H # time remain
state = (lob.display_book(depth),
np.array([tmp1, tmp2], dtype=float))
state = self.process_state(state)
states.append(state)
mid_price = lob.get_mid_price()
state_input = self.pad_state(states[-self._config.state_history:],
self._config.state_history)
price, action = action_fn(start_time + H - t, amount_remain,
state_input, mid_price)
actions.append(action)
done_mask.append(False)
lob.update_own_order(price, amount_remain)
for idx, message in self._mq.pop_to_next_time(t + dH):
lob.process(**message)
if lob.own_amount_to_trade == 0:
done_mask.append(True)
state = (lob.display_book(depth),
np.array([
0, 1.0 * (start_time + H - self._mq._time) / H
],
dtype=float))
state = self.process_state(state)
states.append(state)
rewards.append(lob.own_reward - cum_reward)
break
if done_mask[-1]:
break
else:
# What is going on over here?
rewards.append(lob.own_reward - cum_reward)
cum_reward = lob.own_reward
amount_remain = lob.own_amount_to_trade
if not done_mask[-1]:
tmp1 = 1.0 * amount_remain / amount
tmp2 = 1.0 * (start_time + H - t - dH) / H
state = (lob.display_book(depth),
np.array([tmp1, tmp2], dtype=float))
state = self.process_state(state)
states.append(state)
done_mask.append(False)
lob.update_own_order(lob.own_trade_type *
Limit_Order_book._DUMMY_VARIABLE)
if lob.own_amount_to_trade == 0:
rewards.append(lob.own_reward - cum_reward)
else:
rewards.append(-Limit_Order_book._DUMMY_VARIABLE)
tmp1 = 1.0 * lob.own_amount_to_trade / amount
state = (lob.display_book(depth), np.array([tmp1, 0], dtype=float))
state = self.process_state(state)
states.append(state)
actions.append(self._config.L)
done_mask.append(True)
return (states, rewards, actions, done_mask[1:])
def sampling_buffer(self):
for start_time in range(self._config.train_start,
self._config.train_end, self._config.H):
states, rewards, actions, done_mask = self.simulate_an_episode(
self._config.I, self._config.T, self._config.H, start_time,
self._config.direction, self._action_fn, self._config.depth)
self._bf.store(states, actions, rewards, done_mask)
def process_state(self, state):
state_book, state_it = state
state_book = state_book.astype('float32')
state_book[:, 0] /= 1.e6
state_book[:, 1] /= 1.e2
state_book[:, 2] /= 1.e6
state_book[:, 3] /= 1.e2
return (state_book, state_it)
mq.reset() mq.jump_to_time(time) lob_copy = copy.deepcopy(lob) lob_copy.update_own_order(a_price, amount) for idx, message in mq.pop_to_next_time(next_time): lob_copy.process(**message) if lob_copy.own_amount_to_trade == 0: break return [lob_copy.own_amount_to_trade, lob_copy.own_reward] path_target = '../data/%s_Q_dp_%s.npy' % (args.tic,args.V) oq = Order_Queue(file_order) mq = Message_Queue(file_msg) if args.mode == 'train': np.save(path_target, Calculate_Q(args.V, args.H, args.T, args.I, args.L,oq,mq)) elif args.mode == 'test': Q = np.load(path_target) Optimal_Q = Optimal_strategy(Q) rewards = evaluate_policy(args.test_start, args.test_end, args.order_direction, args.V, args.H, args.T, oq, mq, Optimal_action) print(rewards) print(np.mean(rewards))
help='Buy 1, Sell -1',
type=int)
parser.add_argument('--start', default=34200, help='Start Time', type=float)
parser.add_argument('--end', default=34500, help='End Time', type=float)
parser.add_argument('--adj_freq',
default=100,
help='Adjustment Frequency',
type=float)
parser.add_argument('--tol',
default=1e-8,
help='Remaining Time To Submit Market Order',
type=float)
parser.add_argument('--base_point', default=100, help='Base Point', type=int)
args = parser.parse_args()
mq = Message_Queue(args.file_msg)
lob = Limit_Order_book(own_amount_to_trade=args.order_size,
own_init_price=-args.order_direction *
Limit_Order_book._DUMMY_VARIABLE,
own_trade_type=args.order_direction)
for idx, message in mq.pop_to_next_time(args.start):
lob.process(**message)
def optimal(time, lob, mq):
if time == (
args.end - args.tol
): # This code force that (args.end-args.start) is a multiple of args.tol
if lob.own_amount_to_trade == 0:
return lob.own_reward
else:
print('Add Buy Order %s' % status)
elif idx in [41]:
print('Add Sell Order %s' % status)
elif idx in [5]:
print('Execute Hidden Order %s' % status)
elif idx in [46]:
print('Delete Buy Order %s' % status)
elif idx in [47]:
print('Delete Sell Order %s' % status)
if status == '[FAIL]':
print('ERROR! idx %d msg %s' % (idx, str(msg)))
message_path = '../datasets/LOBSTER_SampleFile_GOOG_2012-06-21_10/GOOG_2012-06-21_34200000_57600000_message_10.csv'
mq = Message_Queue(message_path)
book_path = '../datasets/LOBSTER_SampleFile_GOOG_2012-06-21_10/GOOG_2012-06-21_34200000_57600000_orderbook_10.csv'
df_book = pd.read_csv(book_path, header=None)
level = 10
ask_book = df_book[np.arange(level) * 4].values
ask_size_book = df_book[1 + np.arange(level) * 4].values
bid_book = df_book[2 + np.arange(level) * 4].values
bid_size_book = df_book[3 + np.arange(level) * 4].values
book = np.concatenate([
tmp[:, :, np.newaxis]
for tmp in [bid_book, bid_size_book, ask_book, ask_size_book]
],
axis=2)
for idx, message in mq.iterate_queue():
default=1,
help='Buy 1, Sell -1',
type=int)
parser.add_argument('--start', default=34200, help='Start Time', type=float)
parser.add_argument('--end', default=34500, help='End Time', type=float)
parser.add_argument('--adj_freq',
default=100,
help='Adjustment Frequency',
type=float)
parser.add_argument('--tol',
default=1e-8,
help='Remaining Time To Submit Market Order',
type=float)
args = parser.parse_args()
mq = Message_Queue(args.file_msg)
lob = Limit_Order_book(own_amount_to_trade=args.order_size,
own_init_price=-args.order_direction *
Limit_Order_book._DUMMY_VARIABLE,
own_trade_type=args.order_direction)
for idx, message in mq.pop_to_next_time(args.start):
lob.process(**message)
lob.update_own_order(args.order_direction * Limit_Order_book._DUMMY_VARIABLE)
current_time = args.start
while lob.own_amount_to_trade > 0 and not mq.finished():
current_time += args.adj_freq
for idx, message in mq.pop_to_next_time(current_time):
lob.process(**message)
if lob.own_amount_to_trade == 0:
help='Adjustment Frequency',
type=float)
parser.add_argument('--tol',
default=100,
help='Remaining Time To Submit Market Order',
type=float)
parser.add_argument('--num',
default=10,
help='The number of base points to go',
type=int)
args = parser.parse_args()
# Use the train_start and train_end to find the best num. H: the total amount of time to execute the orders.
file_msg = '../datasets/%s_2012-06-21_34200000_57600000_message_10.csv' % (
args.tic)
mq = Message_Queue(file_msg)
lob = Limit_Order_book(own_amount_to_trade=args.order_size,
own_init_price=-args.order_direction *
Limit_Order_book._DUMMY_VARIABLE,
own_trade_type=args.order_direction)
for idx, message in mq.pop_to_next_time(args.train_start):
lob.process(**message)
current_mid_price = lob.bid[0] + (lob.ask[0] - lob.bid[0]) // 2
init_price = np.arange(current_mid_price - args.num * args.base_point,
current_mid_price + args.num * args.base_point,
args.base_point)
init_price = init_price[init_price > 0]
reward = np.zeros(init_price.shape)