def testPolicy(symbol='JPM', sd=dt.datetime(2008,1,1), ed=dt.datetime(2009,12,31), sv=100000):
prices = get_data([symbol], pd.date_range(sd, ed))
prices = prices[symbol]
SMA = sma(prices)
BBP = bbp(prices, SMA)
SMA = prices / SMA
orders = prices.copy()
orders[:] = 0
orders[(SMA < 0.95) & (BBP < 0)] = 1
orders[(SMA > 1.05) & (BBP > 1)] = -1
holdings = 0.0
trades = []
for date in orders.index:
trade = 0
if orders.loc[date] == 1:
trade = 1000 - holdings
trades.append((date, trade))
elif orders.loc[date] == -1:
trade = -1000 - holdings
trades.append((date, trade))
elif orders.loc[date] == 0:
trade = 0
trades.append((date, trade))
holdings = holdings + trade
df_trades = pd.DataFrame(trades, columns=["Date", symbol])
df_trades.set_index("Date", inplace=True)
return df_trades
def testPolicy(self, symbol = "IBM", \
sd=dt.datetime(2009,1,1), \
ed=dt.datetime(2010,1,1), \
sv = 10000):
syms = [symbol]
dates = pd.date_range(sd, ed)
prices = ut.get_data(syms, dates)
prices = prices[syms]
SMA = sma(prices)
BBP = bbp(prices, SMA)
SMA = prices / SMA
Xtest = pd.concat([SMA, BBP], axis=1)
Xtest = Xtest.values
Y = self.learner.query(Xtest)
trades = pd.DataFrame(0.0, index=prices.index, columns=[symbol])
holdings = 0.0
for i in range(0, trades.shape[0] - 1):
if Y[0][i] >= 0.5:
trades[symbol].iloc[i] = 1000.0 - holdings
elif Y[0][i] <= -0.5:
trades[symbol].iloc[i] = -1000.0 - holdings
else:
trades[symbol].iloc[i] = 0.0
holdings += trades[symbol].iloc[i]
return trades
def testPolicy(symbol,
sd=dt.datetime(2008, 1, 1),
ed=dt.datetime(2009, 12, 31),
sv=100000):
# get volume and prices
pricesDF = get_data([symbol], pd.date_range(sd, ed)).drop(['SPY'], axis=1)
volumeDF = get_data([symbol], pd.date_range(sd, ed),
colname="Volume").drop(['SPY'], axis=1)
closeDF = get_data([symbol], pd.date_range(sd, ed),
colname="Close").drop(['SPY'], axis=1)
volumeDF = (volumeDF * closeDF / pricesDF)
# calc indicators
bbpDF = ind.bbp(pricesDF, 10)
rsi_obvDF = ind.rsi_obv(pricesDF, volumeDF, 14)
#trixDF = ind.trix(pricesDF, 14)
# check conditions:
# close when trix crosses the avarage
# buy - when bbf < 0 or rsi of obv < 30
# sell - when bbf > 1 or rsi of obv > 70
'''trix_cross = pd.DataFrame(0, index=trixDF.index, columns=["JPM"])
trix_cross[trixDF >= 1] = 1
trix_cross[1:] = trix_cross.diff()'''
orders = pd.DataFrame(np.nan,
index=pricesDF.index,
columns=pricesDF.columns)
orders[(bbpDF < 0) | (rsi_obvDF < 30)] = 1000
orders[(bbpDF > 1) | (rsi_obvDF > 70)] = -1000
#orders[(trix_cross != 0)] = 0
orders.ffill(inplace=True)
orders.fillna(0, inplace=True)
orders = orders.diff()
orders.iloc[0] = 0
return orders
def testPolicy(self, symbol="IBM", \
sd=dt.datetime(2009, 1, 1), \
ed=dt.datetime(2010, 1, 1), \
sv=10000):
## 14 day Lookup period
lookback_dates = pd.date_range(sd - dt.timedelta(days=30), ed)
## Get prices
prices = ut.get_data([symbol], lookback_dates)
prices = prices.ix[:, [symbol]]
prices = prices.fillna(method='ffill', inplace=False)
prices = prices.fillna(method='bfill', inplace=False)
## Get indicators
priceOverSMAValues = priceOverSMA(prices)
bbpValues = bbp(prices)
rsiValues = rsi(prices)
##Discretize
self.prices = prices.ix[sd:]
self.symbol = symbol
priceOverSMAValues = priceOverSMAValues.ix[sd:]
priceOverSMAValues[symbol] = pd.qcut(priceOverSMAValues[symbol].values,
10).codes
bbpValues = bbpValues.ix[sd:]
bbpValues[symbol] = pd.qcut(bbpValues[symbol].values, 10).codes
rsiValues = rsiValues.ix[sd:]
rsiValues[symbol] = pd.qcut(rsiValues[symbol].values, 10).codes
states = (priceOverSMAValues * 100) + (bbpValues * 10) + rsiValues * 1
df_trades = priceOverSMAValues.copy()
##Query
state = states.ix[0, symbol]
action = self.learner.querysetstate(state)
total_days = states.shape[0]
total_reward = 0
day = 0
net_position = self.NONE
df_trades.ix[day, symbol] = self.addNewAction(net_position, day,
action)
for day in range(1, total_days):
net_position = action
total_reward += self.getReward(net_position, day)
state = states.ix[day, symbol]
action = self.learner.querysetstate(state)
df_trades.ix[day,
symbol] = self.addNewAction(net_position, day, action)
df_trades = df_trades[(df_trades.T != 0).any()]
return df_trades
def addEvidence(self,
symbol="IBM",
sd=dt.datetime(2008, 1, 1),
ed=dt.datetime(2009, 1, 1),
sv=10000):
syms = [symbol]
dates = pd.date_range(sd, ed)
prices_all = ut.get_data(syms, dates)
pricesDF = prices_all[syms]
prices_SPY = prices_all['SPY']
volumeDF = ut.get_data(syms, dates, colname="Volume").drop(['SPY'],
axis=1)
closeDF = ut.get_data(syms, dates, colname="Close").drop(['SPY'],
axis=1)
volumeDF = (volumeDF * closeDF / pricesDF)
if self.verbose:
print pricesDF
# calc indicators
bbpDF = ind.bbp(pricesDF, 10)
rsi_obvDF = ind.rsi_obv(pricesDF, volumeDF, 14)
self.bbpBins = self.discretizing(bbpDF[symbol].values)
self.obvBins = self.discretizing(rsi_obvDF[symbol].values)
bbpDF = self.getDiscretValue(self.bbpBins, bbpDF[symbol].values)
rsi_obvDF = self.getDiscretValue(self.obvBins,
rsi_obvDF[symbol].values)
prices = pricesDF[symbol].values
orders = np.empty(len(prices))
min_iter = 0
while 1:
old_orders = np.copy(orders)
orders[0:15] = 0
s = self.getState([bbpDF[14], rsi_obvDF[14]])
self.learner.querysetstate(s)
hold_value = 0
impact_factor = 1
for i in range(15, len(bbpDF)):
r = ((prices[i] /
(prices[i - 1] * impact_factor)) - 1) * hold_value
a = self.learner.query(s, r)
hold_value = self.translateAction(a)
orders[i] = hold_value
impact_factor = self.calcImpact(hold_value)
s = self.getState([bbpDF[i], rsi_obvDF[i]])
if np.array_equal(
orders, old_orders
) and min_iter > self.min_iter: #converged and min iterations
break
min_iter += 1
return orders
def testPolicy(self,
symbol="IBM",
sd=dt.datetime(2009, 1, 1),
ed=dt.datetime(2010, 1, 1),
sv=10000):
syms = [symbol]
dates = pd.date_range(sd, ed)
prices_all = ut.get_data(syms, dates)
pricesDF = prices_all[syms]
prices_SPY = prices_all['SPY']
volumeDF = ut.get_data(syms, dates, colname="Volume").drop(['SPY'],
axis=1)
closeDF = ut.get_data(syms, dates, colname="Close").drop(['SPY'],
axis=1)
volumeDF = (volumeDF * closeDF / pricesDF)
if self.verbose:
print pricesDF
# calc indicators
bbpDF = ind.bbp(pricesDF, 10)
rsi_obvDF = ind.rsi_obv(pricesDF, volumeDF, 14)
bbpDF = self.getDiscretValue(self.bbpBins, bbpDF[symbol].values)
rsi_obvDF = self.getDiscretValue(self.obvBins,
rsi_obvDF[symbol].values)
prices = pricesDF[symbol].values
orders = np.empty(len(prices))
s = self.getState([bbpDF[14], rsi_obvDF[14]])
self.learner.querysetstate(s)
orders[0:15] = 0
for i in range(15, len(bbpDF)):
a = self.learner.querysetstate(s)
orders[i] = self.translateAction(a)
s = self.getState([bbpDF[i], rsi_obvDF[i]])
orders = pd.DataFrame(orders,
index=pricesDF.index,
columns=pricesDF.columns)
if self.verbose:
print type(orders) # it better be a DataFrame!
print orders
print prices_all
orders = orders.diff()
orders.iloc[0] = 0
return orders
def addEvidence(self, symbol = "IBM", \
sd=dt.datetime(2008,1,1), \
ed=dt.datetime(2009,1,1), \
sv = 10000):
leaf_size = 5
bags = 50
N = 10
YBUY = 0.01
YSELL = -0.01
syms = [symbol]
dates = pd.date_range(sd, ed)
prices = ut.get_data(syms, dates)
prices = prices[syms]
SMA = sma(prices)
BBP = bbp(prices, SMA)
SMA = prices / SMA
X = pd.concat([SMA, BBP], axis=1)
X = X[:-N]
X = X.values
ndayreturns = (prices.shift(-N) / prices) - 1.0
Y = ndayreturns.applymap(lambda x: 1.0 if x > (YBUY + self.impact) else
(-1.0 if x < (YSELL - self.impact) else 0.0))
Y = Y.dropna()
Y = Y.values
self.learner = bl.BagLearner(learner=rt.RTLearner,
kwargs={'leaf_size': leaf_size},
bags=bags,
boost=False,
verbose=False)
self.learner.addEvidence(X, Y)
def addEvidence(self, symbol = "IBM", \
sd=dt.datetime(2008,1,1), \
ed=dt.datetime(2009,1,1), \
sv = 10000):
## Initialize Q Learner
self.learner = ql.QLearner(num_states=3000, \
num_actions=3, \
alpha=0.2, \
gamma=0.9, \
rar=0.98, \
radr=0.999, \
dyna=0, \
verbose=False)
## 14 day lookback period
lookback_dates = pd.date_range(sd - dt.timedelta(days=30), ed)
## Get Prices
prices = ut.get_data([symbol], lookback_dates)
prices = prices.ix[:, [symbol]]
prices = prices.fillna(method='ffill', inplace=False)
prices = prices.fillna(method='bfill', inplace=False)
## Get indicators
priceOverSMAValues = priceOverSMA(prices)
bbpValues = bbp(prices)
rsiValues = rsi(prices)
self.prices = prices.ix[sd:]
self.symbol = symbol
## Discretize
priceOverSMAValues = priceOverSMAValues.ix[sd:]
priceOverSMAValues[symbol] = pd.qcut(priceOverSMAValues[symbol].values,
10).codes
bbpValues = bbpValues.ix[sd:]
bbpValues[symbol] = pd.qcut(bbpValues[symbol].values, 10).codes
rsiValues = rsiValues.ix[sd:]
rsiValues[symbol] = pd.qcut(rsiValues[symbol].values, 10).codes
## Set up state
states = (priceOverSMAValues * 100) + (bbpValues * 10) + rsiValues * 1
##Update Q Table until converged
converged = False
prev_total_reward = 0
reward_match_count = 0
while not converged:
state = states.ix[0, symbol]
action = self.learner.querysetstate(state)
net_position = self.NONE
total_days = states.shape[0]
total_reward = 0
for day in range(1, total_days):
net_position = action
reward = self.getReward(net_position, day)
total_reward += reward
state = states.ix[day, symbol]
action = self.learner.query(state, reward)
prev_total_reward = total_reward
if total_reward == prev_total_reward:
reward_match_count += 1
if reward_match_count == 5:
converged = True