order = th.limit_sell(client, 'BTCUSDT', exit_price, quantity=exit_capital)
th.log(log_fname, order_type='NEW_SELL', quantity=order['origQty'],
price=exit_price, time=th.UTC('iso'))
# pause/sleep for the trading_interval
time.sleep(interval_s)
# get last time from data
t = data.index[-1]
# update data
tmp = h.get_klines([t.year, t.month, t.day, t.hour, t.minute, t.second], interval=trading_interval)
data.append(tmp)
forecast = model.predict(data['returns'][-win_sz:])
except ValueError:
pass # Handles lookback errors in beginning of dataset
"""## Live trading on Binance"""
# load pre-trainied model
#model_name = 'BidirLSTM_2layer_regress_epochs150.h5'
#model_dir = os.path.join('models', model_name)
#fname_model = 'BidirLSTM_2layer_regress_epochs150.h5' % (model_dir, model_name, epochs)
model = tf.keras.models.load_model('BidirLSTM_96Gold_epochs100.h5')
keys_fname = 'Group4_key.json' # set this file name to your file name
log_fname = 'tradingbot_group4_%s.log' % (th.UTC('iso')[:-10]) #
win_sz = 50 # example
trading_interval = '12h' # example
strategy(model, keys_fname, log_fname, win_sz=50, trading_interval=trading_interval)
def strategy(model, keys_fname, log_fname, win_sz, trading_interval='12h'):
"""
Arguments
---------
model : Some trained forecasting model
keys_fname : File name to the keys, see below
log_fname : You will need to log the trades you submit to Binance.
The file name should be on the format "tradingbot_group<number>_<date_time>.log".
E.g. "tradingbot_group1_2019-11-13_00:23:23.log"
win_sz : The window size that you specified when training the model,
i.e. how many time steps of historical data the model needs for a forecast.
trading_interval : How often the trading bot will query data from Binance,
make a new forecast and decide whether to trade.
E.g. '1min', '5min', '15min', '30min', '1h', '2h', '4h', '6h', '12h', '1d' or '1w'.
"""
# use the keys stored in KEYS_FNAME to initialize the Binance client
client = th.init_client(keys_fname)
# convert the trading_interval string to seconds (int)
interval_s = th.interval_string_to_seconds(trading_interval)
# initialize the log file
th.log(log_fname, new_log=True)
# initialize trading with historical data
t = datetime.utcnow()
# timedelta(weeks=2) -> 2 weeks back in time
t = th.UTC('time') - timedelta(weeks=7)
# get data to begin forecasting from
data = th.get_klines([t.year, t.month, t.day, t.hour, t.minute, t.second], interval=trading_interval)
# forecast
x = data['returns'][-win_sz:].to_numpy().reshape((win_sz, 1))
forecast = model.predict(x)
while True:
try:
if forecast[-1] > 0.0008: # forcasted price increase -> buy
# decide on how to set the entry price
entry_price = data['high'][-1]
risk = 0.7 # all in, max risk
# get the current balance
balance = client.get_asset_balance(asset='USDT')
entry_capital = np.float64(balance['free']) * risk # balance['free'] is a str -> convert to float
if entry_capital > 0:
order = th.limit_buy(client, 'BTCUSDT', entry_price, quantity=entry_capital)
th.log(log_fname, order_type='NEW_BUY', quantity=order['origQty'],
price=entry_price, time=th.UTC('iso'))
elif forecast[-1] < 0: # forcasted price decrease
exit_price = data['low'][-1]
balance = client.get_asset_balance(asset='BTC')
risk = 0.6 # all in, max risk
exit_capital = np.float64(balance['free']) * risk # balance['free'] is a str -> convert to float
if entry_capital > 0:
order = th.limit_sell(client, 'BTCUSDT', exit_price, quantity=exit_capital)
th.log(log_fname, order_type='NEW_SELL', quantity=order['origQty'],
price=exit_price, time=th.UTC('iso'))
# pause/sleep for the trading_interval
time.sleep(interval_s)
# get last time from data
t = data.index[-1]
# update data
tmp = h.get_klines([t.year, t.month, t.day, t.hour, t.minute, t.second], interval=trading_interval)
data.append(tmp)
forecast = model.predict(data['returns'][-win_sz:])
except ValueError:
pass # Handles lookback errors in beginning of dataset
def strategy(model, keys_fname, log_fname, win_sz, trading_interval='6h'):
"""
Arguments
---------
model : Some trained forecasting model
keys_fname : File name to the keys, see below
log_fname : You will need to log the trades you submit to Binance.
The file name should be on the format "tradingbot_group<number>_<date_time>.log".
E.g. "tradingbot_group1_2019-11-13_00:23:23.log"
win_sz : The window size that you specified when training the model,
i.e. how many time steps of historical data the model needs for a forecast.
trading_interval : How often the trading bot will query data from Binance,
make a new forecast and decide whether to trade.
E.g. '1min', '5min', '15min', '30min', '1h', '2h', '4h', '6h', '12h', '1d' or '1w'.
"""
# use the keys stored in KEYS_FNAME to initialize the Binance client
client = th.init_client(keys_fname)
# convert the trading_interval string to seconds (int)
interval_s = th.interval_string_to_seconds(trading_interval)
# initialize the log file
th.log(log_fname, new_log=True)
# initialize trading with historical data
t = datetime.utcnow()
# timedelta(weeks=2) -> 2 weeks back in time
t = th.UTC('time') - timedelta(weeks=2)
# get data to begin forecasting from
data = th.get_klines([t.year, t.month, t.day, t.hour, t.minute, t.second], interval=trading_interval)
# forecast
x = data['returns'][-win_sz:].to_numpy().reshape((win_sz, 1))
forecast = model.predict(x)
while True:
try:
if forecast[-1] >= 0.05: # forcasted big price increase, the average 6h window price increase from 2011 to 2019 is around 0.11%
# there 243 instances that the 6h window price increase is greater than 5%, so we think 5% is a good cutoff
# decide on how to set the entry price
entry_price = data['high'][-1] * 0.95 # since May 2019, the Bitcoin price has been dropping, if even the forecasted is positive
# we are not going to pay the full "high" price, but at a 5% discount price
risk = 0.4 # modified risk
# get the current balance
balance = client.get_asset_balance(asset='USDT')
entry_capital = np.float64(balance['free']) * risk # balance['free'] is a str -> convert to float
if entry_capital > 0:
order = th.limit_buy(client, 'BTCUSDT', entry_price, quantity=entry_capital)
th.log(log_fname, order_type='NEW_BUY', quantity=order['origQty'],
price=entry_price, time=th.UTC('iso'))
if 0.05 > forecast[-1] > 0.00075: # forcasted price increase to cover the commission
# decide on how to set the entry price
entry_price = data['high'][-1] * 0.9 # purchase at a 10% discount
risk = 0.3 # modified risk
# get the current balance
balance = client.get_asset_balance(asset='USDT')
entry_capital = np.float64(balance['free']) * risk # balance['free'] is a str -> convert to float
if entry_capital > 0:
order = th.limit_buy(client, 'BTCUSDT', entry_price, quantity=entry_capital)
th.log(log_fname, order_type='NEW_BUY', quantity=order['origQty'],
price=entry_price, time=th.UTC('iso'))
elif forecast[-1] < 0: # forcasted price decrease
exit_price = data['low'][-1] * 1.05 # we would like to sell at 5% premium
balance = client.get_asset_balance(asset='BTC')
risk = 0.5 # modified risk
exit_capital = np.float64(balance['free']) * risk # balance['free'] is a str -> convert to float
if entry_capital > 0:
order = th.limit_sell(client, 'BTCUSDT', exit_price, quantity=exit_capital)
th.log(log_fname, order_type='NEW_SELL', quantity=order['origQty'],
price=exit_price, time=th.UTC('iso'))
# pause/sleep for the trading_interval
time.sleep(interval_s)
# get last time from data
t = data.index[-1]
# update data
tmp = th.get_klines([t.year, t.month, t.day, t.hour, t.minute, t.second], interval=trading_interval)
data.append(tmp)
forecast = model.predict(data['returns'][-win_sz:])
except ValueError:
pass