def get_reward(new_state, time_step, action, xdata, signal, terminal_state, eval=False, epoch=0):
reward = 0
signal.fillna(value=0, inplace=True)
if eval == False:
bt = twp.Backtest(pd.Series(data=[x for x in xdata[time_step - 2:time_step]],
index=signal[time_step - 2:time_step].index.values),
signal[time_step - 2:time_step],
signalType='shares')
reward = ((bt.data['price'].iloc[-1] - bt.data['price'].iloc[-2]) * bt.data['shares'].iloc[-1])
if terminal_state == 1 and eval == True:
# save a figure of the test set
bt = twp.Backtest(pd.Series(data=[x for x in xdata], index=signal.index.values), signal, signalType='shares')
reward = bt.pnl.iloc[-1]
plt.figure(figsize=(3, 4))
bt.plotTrades()
plt.axvline(x=400, color='black', linestyle='--')
plt.text(250, 400, 'training data')
plt.text(450, 400, 'test data')
plt.suptitle(str(epoch))
plt.savefig('plt/' + 'value_iter_' + str(epoch) + '.png', bbox_inches='tight', pad_inches=1, dpi=72)
plt.close('all')
print(time_step, terminal_state, eval, reward)
return reward
def get_reward(new_state, time_step, action, xdata, signal, terminal_state, eval=False, epoch=0):
reward = 0
signal.fillna(value=0, inplace=True)
if eval == False:
# bt = twp.Backtest(pd.Series(data=[x for x in xdata], index=signal.index.values), signal, signalType='shares')
# reward = bt.pnl.iloc[-1]
# print(reward)
bt = twp.Backtest(pd.Series(data=[x for x in xdata[time_step-2:time_step]], index=signal[time_step-2:time_step].index.values), signal[time_step-2:time_step], signalType='shares')
reward = ((bt.data['price'].iloc[-1] - bt.data['price'].iloc[-2])*bt.data['shares'].iloc[-1])
# print(reward)
# print(xdata[time_step-2:time_step])
# print(signal[time_step-2:time_step])
# print(bt.data)
# print('*******************************')
if terminal_state == 1 and eval == True:
#save a figure of the test set
# print(signal)
bt = twp.Backtest(pd.Series(data=[x for x in xdata], index=signal.index.values), signal, signalType='shares')
reward = bt.pnl.iloc[-1]
# print(reward)
# print(bt.data)
# plt.figure(figsize=(3,4))
# bt.plotTrades()
# plt.axvline(x=400, color='black', linestyle='--')
# plt.text(250, 400, 'training data')
# plt.text(450, 400, 'test data')
# plt.suptitle(str(epoch))
# plt.savefig('plt/'+str(epoch)+'.png')
#print(time_step, terminal_state, eval, reward)
return reward
def get_reward(new_state, time_step, action, xdata, signal, terminal_state, eval=False, epoch=0):
reward = 0
signal.fillna(value=0, inplace=True)
if eval == False:
bt = twp.Backtest(pd.Series(data=[x for x in xdata[time_step-2:time_step]], index=signal[time_step-2:time_step].index.values), signal[time_step-2:time_step], signalType='shares')
reward = ((bt.data['price'].iloc[-1] - bt.data['price'].iloc[-2])*bt.data['shares'].iloc[-1])
if terminal_state == 1 and eval == True:
#save a figure of the test set
bt = twp.Backtest(pd.Series(data=[x for x in xdata], index=signal.index.values), signal, signalType='shares')
reward = bt.pnl.iloc[-1]
#print(time_step, terminal_state, eval, reward)
return reward
def get_reward(new_state,
time_step,
action,
xdata,
signal,
terminal_state,
epoch=0):
reward = 0
signal.fillna(value=0, inplace=True)
if terminal_state == 0:
# get reward for the most current action
if signal[time_step] != signal[time_step - 1] and terminal_state == 0:
i = 1
while signal[time_step - i] == signal[time_step - 1 -
i] and time_step - 1 - i > 0:
i += 1
reward = (xdata[time_step - 1, 0] - xdata[time_step - i - 1, 0]
) * signal[time_step - 1] * -100 + i * np.abs(
signal[time_step - 1]) / 10.0
if signal[time_step] == 0 and signal[time_step - 1] == 0:
reward -= 10
# calculate the reward for all actions if the last iteration in set
if terminal_state == 1:
# run backtest, send list of trade signals and asset data to backtest function
bt = twp.Backtest(pd.Series(data=[x[0] for x in xdata]),
signal,
signalType='shares')
reward = bt.pnl.iloc[-1]
return reward
def get_reward(new_state,
time_step,
action,
xdata,
signal,
terminal_state,
epoch=0):
# counter = 0;
# counter +=1;
# # print(counter)
print('*******Get_reward********')
reward = 0
print('time_step', time_step)
signal.fillna(value=0, inplace=True)
if terminal_state == 0:
#get reward for the most current action
if signal[time_step] != signal[time_step - 1] and terminal_state == 0:
i = 1
print('signal[time_step]', signal[time_step])
print('SIGNALsignal[time_step-1]', signal[time_step - 1])
print('SIGNAL', signal)
while signal[time_step - i] == signal[time_step - 1 -
i] and time_step - 1 - i > 0:
print('time_step-1-i', time_step - 1 - i)
i += 1
print('*****i*****', i)
reward = (xdata[time_step - 1, 0] - xdata[time_step - i - 1, 0]
) * signal[time_step - 1] * -100 + i * np.abs(
signal[time_step - 1]) / 10.0
print('xdata', xdata)
print('xdata[time_step-1, 0]', xdata[time_step - 1, 0])
print('xdata[time_step - i-1, 0]', xdata[time_step - i - 1, 0])
print('(xdata[time_step-1, 0] - xdata[time_step - i-1, 0])',
(xdata[time_step - 1, 0] - xdata[time_step - i - 1, 0]))
print('signal[time_step - 1]', signal[time_step - 1])
print('signal[time_step - 1]*-100', signal[time_step - 1] * -100)
print('(-)*signal',
(xdata[time_step - 1, 0] - xdata[time_step - i - 1, 0]) *
signal[time_step - 1] * -100)
print('i*np.abs(signal[time_step - 1])',
i * np.abs(signal[time_step - 1]))
if signal[time_step] == 0 and signal[time_step - 1] == 0:
reward -= 10
# if you didn't trade for two days, you get a negative -10
#calculate the reward for all actions if the last iteration in set
if terminal_state == 1:
#run backtest, send list of trade signals and asset data to backtest function
bt = twp.Backtest(pd.Series(data=[x[0] for x in xdata]),
signal,
signalType='shares')
reward = bt.pnl.iloc[-1]
# print(bt.data)
# print(reward)
print('*******Get_reward********')
return reward
def get_reward(new_state,
time_step,
action,
xdata,
signal,
terminal_state,
reward_dict,
eval=False,
epoch=0):
reward = 0
signal.fillna(value=0, inplace=True)
if eval == False:
for j in range(xdata.shape[1]):
bt = twp.Backtest(pd.Series(
data=[x[j] for x in xdata[time_step - 2:time_step]],
index=signal[time_step - 2:time_step].index.values),
signal[time_step - 2:time_step],
signalType='shares')
reward = (bt.data['price'].iloc[-1] -
bt.data['price'].iloc[-2]) * bt.data['shares'].iloc[-1]
reward_dict[j].append(reward)
if terminal_state == 1 and eval == True:
for j in range(xdata.shape[1]):
bt = twp.Backtest(pd.Series(data=[x[j] for x in xdata],
index=signal.index.values),
signal,
signalType='shares')
reward = bt.pnl.iloc[-1]
reward_dict[j].append(reward)
plt.figure(figsize=(9, 16))
bt.plotTrades()
plt.axvline(x=xdata.shape[0] - 100, color='black', linestyle='--')
plt.text(250, 400, 'training data')
plt.text(450, 400, 'test data')
plt.suptitle(str(epoch))
plt.savefig('plt/' + 'knapsack_' + str(j) + ' ' + str(epoch) +
'.png')
plt.close('all')
print('@ eval=TRUE')
# print(time_step, terminal_state, eval, reward)
return reward_dict
def setUp(self):
self.data = data.Data(1, live=False)
print("Loading Data:", DATA_FILE_NAME)
data.loadData(DATA_FILE_NAME)
self.backtester = backtest.Backtest(cash=10000,
data=self.data,
minBrokerFee=1.00,
perShareFee=0.0075,
simple=True,
stateObj=True)
def get_reward(new_state, time_step, action, xdata, signal, terminal_state, eval=False, epoch=0):
reward = 0
signal.fillna(value=0, inplace=True)
if eval == False:
try:
bt = twp.Backtest(pd.Series(data=[x[0] for x in xdata[time_step - 2:time_step]], index=signal[time_step - 2:time_step].index.values),
signal[time_step - 2:time_step], signalType='shares')
reward = np.max((bt.data['price'].iloc[-1] - bt.data['price'].iloc[-2]) * bt.data['shares'].iloc[-1])
except:
pass
if terminal_state == 1 and eval == True:
bt = twp.Backtest(pd.Series(data=[x[0] for x in xdata], index=signal.index.values), signal, signalType='shares')
reward = bt.pnl.iloc[-1]
plt.figure(figsize=(9, 16))
bt.plotTrades()
plt.axvline(x=400, color='black', linestyle='--')
plt.text(250, 400, 'training data')
plt.text(450, 400, 'test data')
plt.suptitle(str(epoch))
plt.savefig('plt/' + 'knapsack_' + str(epoch) + '.png')
plt.close('all')
'''
# save a figure of the test set
plt.figure(figsize=(10, 25))
for i in range(xdata.T.shape[0]):
#frame = pd.concat(btFrame, axis=1)
bt = twp.Backtest(pd.Series(data=[x for x in xdata.T[i]], index=signal.index.values), signal, signalType='shares')
reward += np.max(bt.pnl.iloc[-1])
bt.plotTrades()
#plt.axvline(x=400, color='black', linestyle='--')
#plt.text(250, 400, 'training data')
#plt.text(450, 400, 'test data')
#plt.suptitle(str(epoch))
plt.savefig('plt/' + 'knapsack_' + str(epoch) + '.png', bbox_inches='tight', pad_inches=1, dpi=72)
plt.close('all')
'''
# print(time_step, terminal_state, eval, reward)
return reward
def test2(df):
ema2h = ind.iMA(df, 2, ma_method='EMA', applied_price='Hight')
ema2l = ind.iMA(df, 2, ma_method='EMA', applied_price='Low')
ema40 = ind.iMA(df, 40, ma_method='EMA')
# Buy : ema2lがema40よりも上側にあり、Close値がema2lにタッチした場合、エントリーし、
# BuyExit : High値がema2hにタッチしたら決済。
be = ((df.Close < enve['Lower']) & rsi10 < 30)
bExt = ema20 < df.Close
se = ((enve['Upper'] < df.Close) & 70 < rsi10)
sExt = df.Close < ema20
t, pl = bt.Backtest(df, be, se, bExt, sExt, TP=1000, SL=5000, Limit=20)
eq = bt.BacktestReport(t, pl)
def revTry(df):
rsi10 = ind.iRSI(df, 10, 'Close')
ema20 = ind.iMA(df, 20, ma_method='EMA')
enve = ind.iEnvelopes(df, 10, 0.05)
#ccc = pd.DataFrame({'Close' : df['Close'], 'EMA20': ema20, 'RSI' : rsi10, 'Upper' : enve['Upper'], 'Low' : enve['Lower']})
#print(ccc)
# Buy : エンベロープの下に突き抜ける + RSIが30以下で買い。
be = ((df.Close < enve['Lower']) & rsi10 < 30)
bExt = ema20 < df.Close
se = ((enve['Upper'] < df.Close) & 70 < rsi10)
sExt = df.Close < ema20
t, pl = bt.Backtest(df, be, se, bExt, sExt, TP=1000, SL=5000, Limit=20)
eq = bt.BacktestReport(t, pl)
def main():
adxcross = MyStrategy()
btest = bt.Backtest(500000.0, OTP_DATA, adxcross)
btest.run()
indicators = [
ind.cci(OTP_DATA, span=10),
ind.cci(OTP_DATA, span=20),
]
plotter = plt.plotter(OTP_DATA)
for i in indicators:
plotter.add_plot(i)
plotter.plot()
def pair_target(up_bound, low_bound):
up_bound = mu + up_bound * select_sigma
low_bound = mu - low_bound * select_sigma
Signal = arbitrage.Pairsignal()
pairsignal = Signal.generate(spread = spread, mu = mu, \
up = up_bound, low = low_bound)
data_backtest1 = copy.deepcopy(pairsignal)
data_backtest1['price'] = data[asset2] - beta * data[asset1]
data_backtest1['date'] = data_backtest.index
data_backtest1['volumn'] = 1
Pairbacktest = backtest.Backtest(stop_lose=-1, stop_profit=1)
backtest_result1 = Pairbacktest.run(data_backtest1, visual=False)
# annulized vol
vol = backtest_result1['daily_PLrate'].std() * np.sqrt(252)
# annualized return
return_ = (backtest_result1['new_value'].iloc[-1] -
1) * 252.0 / len(backtest_result1)
# shape ratio
sharpe = (return_ - rf) * 1.0 / (vol + 0.001)
return sharpe
def getReward(timeStep, signal, price, state, endState):
unit_reward = 1
if not endState:
# net earning from previous action
net = (price[timeStep] - price[timeStep - 1]) * signal[timeStep - 1]
else:
bt = twp.Backtest(price, signal, signalType='shares')
net = bt.pnl.iloc[-1]
rewards = 0
#intuition reward
if net > 0:
rewards += unit_reward * 2
elif net < 0:
rewards -= unit_reward / 2
else:
#rewards -= 1 #don't encourage hold
rewards += 0
'''# sma reward
if not endState:
#sma20 = state[0,0,2]
#sma80 = state[0,0,3]
sma20 = state[0, 2]
sma80 = state[0,3]
sma_net = sma20 - sma80
if sma_net < 0 : # short sma < long sma, down trend
if signal[timeStep - 1] < 0:
rewards += unit_reward
elif sma_net > 0: #short sma > long sma, up trend
if signal[timeStep - 1] > 0:
rewards += unit_reward'''
return rewards
def test_agent(pdata, model, data, episode_i):
signal = pd.Series(index=np.arange(len(data)))
signal.fillna(value=0, inplace=True)
signal.loc[0] = 1
state = initializeState(pdata)
endState = 0
timeStep = 1
realProfit = 0
realInventory = []
while not endState:
Q = model.predict(state, batch_size=1)
action = (np.argmax(Q))
nextState, timeStep, signal, endState, realProfit, reward = trade(
action, pdata, signal, timeStep, realInventory, data, realProfit)
state = nextState
while len(realInventory) > 0:
realProfit += (data.iloc[-1] - realInventory.pop(0)) * 10
long = 0
short = 0
hold = 0
newdici = changedecision(signal)
for j in range(signal.shape[0]):
if signal.iloc[j] < 0:
short += 1
elif signal.iloc[j] > 0:
long += 1
else:
hold += 1
newprofitwhold = profitcal(data, newdici, 5000)
print("Profit with hold: ", newprofitwhold, " Orignal Profit: ",
realProfit)
print('Episode #: ', episode_i, ' No Random, Buy: ', long, ', Sell: ',
short, ', Hold: ', hold)
bt = twp.Backtest(data, signal, signalType='shares')
plt.figure(figsize=(20, 20))
plt.subplot(2, 1, 1)
plt.title("trades " + str(episode_i))
plt.xlabel("timestamp")
bt.plotTrades()
plt.subplot(2, 1, 2)
plt.title("PnL " + str(episode_i))
plt.xlabel("timestamp")
bt.pnl.plot(style='-')
if episode_i % 10 == 0:
plt.savefig('plot/' + str(episode_i) + '.png')
plt.show()
plt.close()
bt2 = twp.Backtest(data, newdici, signalType='shares')
plt.figure(figsize=(20, 20))
plt.subplot(2, 1, 1)
plt.title("trades " + str(episode_i))
plt.xlabel("timestamp")
bt2.plotTrades()
plt.subplot(2, 1, 2)
plt.title("PnL " + str(episode_i))
plt.xlabel("timestamp")
bt2.pnl.plot(style='-')
if episode_i % 10 == 0:
plt.savefig('plot/' + str(episode_i) + '_hold.png')
plt.show()
plt.close()
endReward = bt.pnl.iloc[-1]
return endReward, realProfit
model.fit(state.reshape(1, 2), y, batch_size=1, nb_epoch=1, verbose=0)
state = new_state
if terminal_state == 1: # terminal state
status = 0
eval_reward = evaluate_Q(indata, model)
print("Epoch #: %s Reward: %f Epsilon: %f" % (i, eval_reward, epsilon))
learning_progress.append((eval_reward))
if epsilon > 0.1:
epsilon -= (1.0 / epochs)
elapsed = np.round(timeit.default_timer() - start_time, decimals=2)
print("Completed in %f" % (elapsed, ))
# plot results
bt = twp.Backtest(pd.Series(data=[x[0] for x in xdata]),
signal,
signalType='shares')
bt.data['delta'] = bt.data['shares'].diff().fillna(0)
print(bt.data)
plt.figure()
bt.plotTrades()
plt.suptitle('epoch' + str(i))
plt.savefig('plt/final_trades' + '.png',
bbox_inches='tight',
pad_inches=1,
dpi=72) # assumes there is a ./plt dir
plt.close('all')
plt.figure()
import pandas as pd
import backtest as twp
from functions import getData
import numpy as np
# get price & techinical indicator data as pandas dataframe
data, data_prev, sma20, sma80, slowD, slowK = getData()
signal = pd.Series(index=np.arange(len(data)))
for i in range(signal.shape[0]):
signal.loc[i] = 10
#signal.loc[1] = 0
signal.loc[0] = 1
print(signal)
bt = twp.Backtest(pd.Series(data=[x for x in data], index=signal.index.values),
signal,
signalType='shares')
endReward = bt.pnl.iloc[-1]
print(bt.data)
print(signal.shape[0])
print(endReward)
xform_data = self.transform_.transform(X, y)
except:
xform_data = self.transform_.transform(X)
return np.append(X, xform_data, axis=1)
class LogExpPipeline(Pipeline):
def fit(self, X, y):
super(LogExpPipeline, self).fit(X, np.log1p(y))
def predict(self, X):
return np.expm1(super(LogExpPipeline, self).predict(X))
import backtest
bt = backtest.Backtest()
bt.init_candle()
bt.resetbacktest()
bt.index = bt.size
bt.updateIndicators()
truth = bt.df
memory = joblib.Memory(cachedir=".")
n = truth.shape[1]
#
# XGBoost model
#
xgb_params = {}
xgb_params['objective'] = 'reg:linear'
xgb_params['learning_rate'] = 0.001
xgb_params['max_depth'] = int(6.0002117448743721)
up = up_bound, low = low_bound)
data_backtest = copy.deepcopy(pairsignal)
coin_beta = EGresult_df.loc['amid', 'gas'].beta[0][0]
data_backtest['price'] = data['gas'] - coin_beta * data['amid']
data_backtest['date'] = data_backtest.index
data_backtest['volumn'] = 1
# visualization
data_backtest['signal'].plot()
plt.title('Pair Trading Strategy Signal Result')
plt.xlabel('Time')
plt.ylabel('Spread')
plt.show()
# ---------------------------------------------------------------------------- #
# ------------------------------ Strategy Backtest --------------------------- #
# ---------------------------------------------------------------------------- #
Pairbacktest = backtest.Backtest(stop_lose=-1, stop_profit=1)
backtest_result = Pairbacktest.run(data_backtest, visual=True)
# return distribution analysis
#dist_return = backtest_result['daily_PLrate'][backtest_result['daily_PLrate'] != 0]
#Return_dist = evaluation.Distribution(dist_return)
#Return_dist.analysis()
# signal analysis
print '====================== Signal Analysis ======================'
Pair_signal_est = evaluation.Signal_est(data=pairsignal)
Pair_signal_est.analysis()
# risk analysis
print '======================= Risk Analysis ======================='
Pair_risk_est = evaluation.Risk_est(data=backtest_result)
Pair_risk_est.analysis()
# drawdown analysis
print '======================= P&L Analysis ========================'
if __name__ == "__main__":
# DATA_FILE_NAME = "resource\\OIH_adjusted.csv"
DATA_FILE_NAME = "resource\\SPY_EOD.csv"
trader = trader.Trader()
data = data.Data(1, live=False)
print("Loading Data:", DATA_FILE_NAME)
data.loadData(DATA_FILE_NAME)
print("Data Loaded")
featureB = featureBuild.FeatureBuilder(data=data, queueSize=QUEUE_SIZE)
print("Feature Builder Created")
backTester = backtest.Backtest(cash=10000,
data=data,
counter=QUEUE_SIZE,
minBrokerFee=1.00,
perShareFee=0.0075,
simple=True,
stateObj=True)
print("Backtester Created")
agent = ddpg.DeepDPG(state_shape=QUEUE_SIZE,
normal_state_shape=featureB.numFeatures,
criticParams=ddpg.CriticParams(),
policyParams=ddpg.PolicyParams(),
OUParams=ddpg.OUParams(),
episodes=500)
print("Agent Created")
manager = manager.Manager(data=data,
agent=agent,
featureB=featureB,
backtester=backTester,
# Update EMA calcs for imported data
d.update_ema(ma_1)
d.update_ema(ma_2)
# Loop for x minutes
# for mins in range(720):
# d.refresh_ohlc()
# print('Refreshing OHLC data. Last ID: ' + str(d.ohlc_last_id))
# time.sleep(60)
# Set up brain object with data and trade vol
b = brain.Brain(d)
# Backtest - pass data and brain object parameters
bt = backtest.Backtest(d, b)
# Monte carlo simulation from min to max MA
bt.run_ma_sim(type, 1, 50)
# Simulate single MA combo and export simulated trades
# bt.ma_sim(type, ma_1, ma_2)
# bt.export_trades()
# Save data
# d.export_ohlc()
# d.export_ohlc_csv()
# Testing
# d.get_cryptobal()
# d.get_fiatbal()
def bestow(self,
new_state,
step,
action,
prices,
trades,
term,
qs=None,
epoch=None,
plot=False,
evaluate=False):
'''
returns reward for agent given the action taken
'''
bestowal = [0] * len(self.symbols)
if evaluate == False:
for i in range(len(self.symbols)):
b_test = bt.Backtest(
pd.Series(data=prices[i, step - 2:step],
index=np.arange(step - 2, step)).astype(float),
pd.Series(data=trades[step - 2:step, i],
index=np.arange(step - 2, step)).astype(float),
roundShares=False,
signalType='shares')
bestowal[i] = ((b_test.data['price'].iloc[-1] -
b_test.data['price'].iloc[-2]) *
b_test.data['trades'].iloc[-1])
elif evaluate == True and term == True:
for i in range(len(self.symbols)):
close = self.data_test[:, 0, 0]
b_test = bt.Backtest(pd.Series(data=prices[i, :],
index=np.arange(len(trades))),
pd.Series(data=trades[:, i]),
roundShares=False,
initialCash=self.init_cash,
signalType='shares')
b_test2 = bt.Backtest(pd.Series(data=close,
index=np.arange(len(trades))),
pd.Series(data=trades[:, i]),
roundShares=False,
initialCash=self.init_cash,
signalType='shares')
bestowal[i] = b_test.pnl.iloc[-1]
if (plot == True) and ((epoch % 20 == 0) or epoch == 1):
matplotlib.rcParams['font.size'] = 24
f = plt.figure(figsize=(12, 12))
ax1 = f.add_subplot(2, 1, 1)
b_test2.plotTrades()
plt.axvline(x=round(self.data_size * 0.8),
color='black',
linestyle='--')
ax2 = f.add_subplot(2, 1, 2)
ax2.plot(b_test2.data.netProfit,
color='green',
label='Net Profit')
ax2.plot(b_test2.pnl, color='k', label='Adj. Profit')
plt.title('Profit and Loss')
plt.legend()
plt.savefig('./images/{}_{}sgd_summary.png'.format(
self.symbols[i], epoch),
bbox_inches='tight',
pad_inches=0.5,
dpi=60)
plt.close('all')
self.init_cash = 1000
self.init_cash = 1000
return bestowal