class RLTrading:
def __init__(self):
self.agent = Agent(hist_ws=hist_ws, forward_ws=forward_ws)
self.trader = Trader(record=True)
self.fee_rate = 0.0001
def Train(self,
epochs=100,
train_files=['/running/2019-05-22/ni8888.csv'],
target_profit=2,
stoploss=2,
fee=0.5,
learn_step=2000):
for _ in range(epochs):
for f in train_files:
wallet = 0
count = 1
hist, forward = self._gen_sample(f)
state = np.hstack(
(hist[0], [[wallet]] * len(hist[0]))) # state defination
for i, hs in enumerate(hist):
h, f = hist[i], forward[i]
action = int(
self.agent.choose_action(np.expand_dims(state, 0)))
print('choose action %d' % (action))
reward, wallet = self._cal_reward(h, f, action, wallet)
s_ = np.hstack(
(h, [[wallet]] * len(h))) # state defination
self.agent.store_memory(state, action, reward, s_)
if count % learn_step == 0:
self.agent.learn()
state = s_
count += 1
def Plot(self, f):
df = pd.read_csv(f)
df.columns = shot.get_columns()
df['mid'] = (df['asks[0]'] + df['bids[0]']) / 2
plt.title(f)
plt.plot(df['mid'])
plt.show()
def Test(self,
test_file=[
'/running/2019-05-22/ni8888.csv',
'/running/2019-05-23/ni8888.csv',
'/running/2019-05-24/ni8888.csv',
'/running/2019-05-27/ni8888.csv',
]):
for f in test_file:
#self.Plot(f)
wallet = 0
count = 1
hist, forward = self._gen_sample(f)
for i, hs in enumerate(hist):
h, f = hist[i], forward[i]
state = np.hstack((h, [[wallet]] * len(h))) # state defination
shot, action = self.agent.real_act(np.expand_dims(state, 0))
shot = shot.flatten()
reward, wallet = self._cal_reward(h, f, action, wallet)
if reward > 0:
print(self.agent.print_qeval(np.expand_dims(state, 0)))
if action == 1: # buy
self.trader.RegisterOneTrade("ticker", 1,
shot[col_dict['asks[0]']])
elif action == 2: # sell
self.trader.RegisterOneTrade("ticker", 1,
shot[col_dict['bids[0]']])
print('in %d actionspace' % (len(hist)))
self.trader.Summary()
# gen_sample: generate sample, one by one, hist_sample->[i, i+hist_ws], forward_sample->[i, i+forward_ws]
def _gen_sample(self, file_name, hist_ws=hist_ws, forward_ws=forward_ws):
print('running read_csv')
df = pd.read_csv(file_name)
df.columns = shot.get_columns()
del df['ticker']
print('read over')
hist_sample, forward_sample = [], []
for i in range(hist_ws, len(df) - forward_ws):
hist_sample.append(df.iloc[i - hist_ws:i].values)
forward_sample.append(df.iloc[i:i + forward_ws].values)
print('gen_sample over')
return hist_sample, forward_sample
# cal_reward: reward based on action,forward and wallet, consider the best situation
def _cal_reward(self, hist, forward, action, wallet):
reward = 0.0
fee = hist[-1][col_dict['asks[0]']] * self.fee_rate
if action == 1: # buy action
buy_price = hist[-1][col_dict['asks[0]']] # ask price
if wallet == 0: # best sitution's pnl as reward
close_price = forward[:, col_dict['bids[0]']].max()
wallet = 1
return close_price - buy_price - fee, wallet
elif wallet == -1: # close sell, using buy
wallet = 0
avg_close_price_buy = forward[:, col_dict['asks[0]']].mean()
return avg_close_price_buy - buy_price, wallet
else:
return -fee / forward_ws, wallet
elif action == 2: # sell action
sell_price = hist[-1][col_dict['bids[0]']] # bid price
if wallet == 0: # best sitution's pnl as reward
close_price = forward[:, col_dict['asks[0]']].min()
wallet -= 1
return sell_price - close_price - fee, wallet
elif wallet == 1: # close buy, using sell, reward are price-avgprice
wallet = 0
avg_close_price_sell = forward[:, col_dict['bids[0]']].mean()
return sell_price - avg_close_price_sell, wallet
else:
return -fee / forward_ws, wallet
else: # sit
if wallet == 0:
return 0.0, wallet
else:
#return -fee/forward_ws, wallet
return 0.0, wallet
class BaseFactor:
def __init__(self):
self.f_value = []
self.tr = Trader()
def run(self, start_date, end_date, ticker):
self.m = self.LoadData(start_date, end_date, ticker)
self.CalFactor()
self.PlotFactor()
self.TestPnl()
self.PlotSignal()
def PlotSignal(self):
df_list = [i[1] for i in sorted(self.m.items(), key=lambda x: x[0])]
if len(df_list) < 1:
print('empty df')
return
df = df_list[0]
start = time.time()
for i in range(1, len(df_list)):
df = pd.merge(df, df_list[i], how='outer')
print('finished merge used %lfs' % (time.time() - start))
plt.plot(df['mid'], label='mid', alpha=0.3)
buy_x = df[df['money'] < 0].index.tolist()
buy = df[df['money'] < -0.1]
plt.scatter(x=buy.index.tolist(),
y=buy['mid'].tolist(),
marker='.',
s=[4] * len(buy),
c='red',
label='buy')
sell = df[df['money'] > 0.1]
plt.scatter(x=sell.index.tolist(),
y=sell['mid'].tolist(),
marker='.',
s=[4] * len(sell),
c='green',
label='sell')
plt.title('factor percentile signal')
plt.legend()
plt.grid()
plt.show()
def TestPnl(self):
for k in self.m:
df = self.m[k]
up = df['factor'].quantile(0.999)
down = df['factor'].quantile(0.001)
df['money'] = np.where(df['factor'] > up, df['mid'], 0.0)
df['money'] = np.where(df['factor'] < down, -df['mid'],
df['money'])
for i in df['money'].tolist():
if abs(i) > 1:
self.tr.RegisterOneTrade('ni8888', 1 if i > 0 else -1,
abs(i))
self.tr.Summary()
self.tr.PlotStratRawPnl(show=True)
self.tr.PlotStratPnl(show=True)
def PlotFactor(self):
plt.plot(self.f_value, label='factor value')
plt.title('factor value curve')
plt.legend()
plt.grid()
plt.show()
def CalFactor(self):
for k in sorted(self.m.keys()):
df = self.m[k]
df['factor'] = self.cal(df)
self.f_value += df['factor'].tolist()
@abstractmethod
def cal(self, df):
pass
def LoadData(self, start_date, end_date, ticker):
dl = dateRange(start_date, end_date)
m = {}
for date in dl:
path = '/root/' + date + '/' + ticker + '.csv'
if os.path.exists(path):
m[date] = self.ReadData(date, ticker)
else:
print("%s not existed!" % (path))
return m
def ReadData(self, date, ticker):
path = '/root/' + date + '/' + ticker + '.csv'
df = pd.read_csv(path, header=None)
df.columns = shot.get_columns()
df['mid'] = (df['asks[0]'] + df['bids[0]']) / 2
df['return1'] = df['mid'].diff(1).fillna(0.0) / df['mid']
return df
class BackTester:
def __init__(self):
self.reader = Reader()
self.order_file_size = 0
self.shot_file_size = 0
self.pos= {}
self.net_pnl = {}
self.gross_pnl = {}
self.ticker_strat_map = {}
self.avgcost = {}
self.gross_time_allpnl_map = {}
self.net_time_allpnl_map = {}
self.strat_data_map = {}
self.gross_strat_pnl_map = {}
self.net_strat_pnl_map = {}
self.pnl_contract = set([])
self.Caler = CALER(contract_config_path)
self.trader = Trader()
def GetStratPair(self, s):
exec('temp=' + s)
return temp
def GetStratPnlKey(self):
return self.net_strat_pnl_map.keys()
def load_binary_order(self, file_path):
self.reader.load_order_file(file_path)
self.order_file_size = self.reader.get_ordersize()
def load_binary_shot(self, file_path):
self.reader.load_shot_file(file_path)
self.shot_file_size = self.reader.get_shotsize()
def load_binary_strat(self, file_path):
start_sec = time.time()
self.reader.load_strat_file(file_path)
self.strat_file_size = self.reader.get_stratsize()
for i in range(self.strat_file_size):
strat = self.reader.read_bstrat(i)
self.pnl_contract.add(self.GetStratPair(strat.ticker)[-1])
for c in self.GetStratPair(strat.ticker):
self.ticker_strat_map[c] = strat.ticker
if not self.strat_data_map.has_key(strat.ticker):
self.strat_data_map[strat.ticker] = {}
self.strat_data_map[strat.ticker][strat.time] = strat.last_trade
print('load start cost ' + str(time.time()-start_sec))
#sys.exit(1)
def load_csv_data(self, file_path):
pass
def clear(self):
self.reader = Reader()
self.order_file_size = 0
self.shot_file_size = 0
self.pos= {}
self.net_pnl = {}
self.gross_pnl = {}
self.ticker_strat_map = {}
self.avgcost = {}
self.gross_time_allpnl_map = {}
self.net_time_allpnl_map = {}
self.strat_data_map = {}
self.gross_strat_pnl_map = {}
self.net_strat_pnl_map = {}
self.pnl_contract = set([])
self.Caler = CALER(contract_config_path)
def RunShot(self):
for i in range(self.shot_file_size):
shot = self.reader.read_bshot(i)
'''
def RunStrat(self):
for i in range(self.strat_file_size):
strat = self.reader.read_bstrat(i)
for c in GetStratPair(strat.contract):
self.ticker_strat_map[c] = strat.contract
self.strat_map[strat.contract] = strat.last_trade
'''
def RunOrder(self):
for i in range(self.order_file_size):
o = self.reader.read_border(i)
o.Show()
if o.action == 4:
print('passing order %s' %(o.order_ref))
continue
if o.shot_time < 156549195 or o.send_time < 1565491959:
print('passing order %s' %(o.order_ref))
continue
self.trader.RegisterOneTrade(o.contract, o.size, o.price)
continue
true_size = (o.size if o.side == 1 else -o.size)
if not self.pos.has_key(o.contract):
self.pos[o.contract] = 0
self.avgcost[o.contract] = 0.0
self.net_pnl[o.contract] = 0
self.gross_pnl[o.contract] = 0
pre_pos = self.pos[o.contract]
self.pos[o.contract] += true_size
is_close = (self.pos[o.contract]*true_size <= 0)
if is_close == True:
print('close order hit')
this_net_pnl = self.Caler.CalNetPnl(o.contract, self.avgcost[o.contract], abs(pre_pos), o.price, o.size, OrderSide.Buy if o.side==1 else OrderSide.Sell)
this_gross_pnl = self.Caler.CalPnl(o.contract, self.avgcost[o.contract], abs(pre_pos), o.price, o.size, OrderSide.Buy if o.side==1 else OrderSide.Sell)
#print("%s %f %i %f %i %i " %(o.contract, self.avgcost[o.contract], abs(pre_pos), o.price, o.size, o.side))
self.net_pnl[o.contract] += this_net_pnl
self.gross_pnl[o.contract] += this_gross_pnl
if self.pos[o.contract] == 0:
self.avgcost[o.contract] = 0.0
if o.contract in self.pnl_contract: # close contract
#self.time_allpnl_map[o.shot_time] = (np.sum(self.pnl.values()), o.contract)
strat_id = self.ticker_strat_map[o.contract]
strat_pair = self.GetStratPair(strat_id)
if not self.net_strat_pnl_map.has_key(strat_id):
self.net_strat_pnl_map[strat_id] = {}
self.net_strat_pnl_map[strat_id][o.shot_time] = np.sum([self.net_pnl[sp] for sp in strat_pair])
if not self.gross_strat_pnl_map.has_key(strat_id):
self.gross_strat_pnl_map[strat_id] = {}
self.gross_strat_pnl_map[strat_id][o.shot_time] = np.sum([self.gross_pnl[sp] for sp in strat_pair])
else:
self.avgcost[o.contract] = abs((self.avgcost[o.contract] * (self.pos[o.contract] - true_size) + o.price * true_size) / self.pos[o.contract])
def Plot(self):
keys = self.GetStratPnlKey()
print keys
ksize = len(keys)
ncol, nrow = 2, 3
width = int(math.sqrt(ksize)) + 1
height = int(math.sqrt(ksize)) +1
fig,ax = plt.subplots(nrows=nrow,ncols=ncol,figsize=(15,8))
count = 0
for i in range(ksize):
key = keys[i]
if count % (ncol*nrow) == 0 and count > 0:
fig.savefig('pnl@i' %(count))
fig.tight_layout()
plt.show()
fig,ax = plt.subplots(nrows=nrow,ncols=ncol,figsize=(15,8))
this_ax = ax[int(count/ncol)%nrow, count%ncol]
this_ax.set_title(key)
data_keys = sorted(self.strat_data_map[key].keys())
self.net_strat_pnl_map[key][np.min(data_keys)-10] = 0.0
self.gross_strat_pnl_map[key][np.min(data_keys)-10] = 0.0
net_pnl_keys = sorted(self.net_strat_pnl_map[key].keys())
gross_pnl_keys = sorted(self.gross_strat_pnl_map[key].keys())
print(net_pnl_keys)
this_ax.plot(net_pnl_keys, [self.net_strat_pnl_map[key][k] for k in net_pnl_keys], label='net_pnl', color='red')
this_ax.plot(gross_pnl_keys, [self.gross_strat_pnl_map[key][k] for k in gross_pnl_keys], label='gross_pnl', color='black')
twin_ax = this_ax.twinx()
#twin_ax.plot(data_keys, [self.strat_data_map[key][k] for k in data_keys], label='strat_data', color='blue', alpha=0.3)
this_ax.legend()
twin_ax.legend()
count += 1
fig.savefig('pnl@%i' %(count))
fig.tight_layout()
plt.show()
def Report(self):
self.trader.Summary()
def Run(self):
if self.order_file_size != 0:
self.RunOrder()
if self.shot_file_size != 0:
self.RunShot()