def __init__(self,
balance,
tickers=None,
base_model=LinearRegression(),
n_prev=2,
wait=100,
steps_ahead=100,
k=5,
envelope='proportional',
log=False,
commission=.0002,
flat_rate=8):
if tickers is None:
tickers = s_and_p_names('2014-1-1', '2015-11-02')
super(TSEBuyAndHoldStrategy, self).__init__(balance,
tickers,
log=log,
commission=commission,
wait=wait,
flat_rate=flat_rate)
self.model = TimeSeriesRegressor(base_model, n_ahead=1, n_prev=n_prev)
self.steps_ahead = steps_ahead
self.k = k
self.envelope = envelope
def get_data_dependece(X,
data_sizes,
folds=20,
test_size=30,
n_prev=2,
log=True):
bests = np.empty((len(data_sizes), folds, X.shape[1]))
for i, data_size in enumerate(data_sizes):
pairs = cascade_cv(len(X),
folds,
data_size=data_size,
test_size=test_size,
number=True)
for j, pair in enumerate(pairs):
if log:
print('data size: {} trial {} '.format(data_size, j))
X_train, X_test = np.array(X.iloc[pair[0], :]), np.array(
X.iloc[pair[1], :])
tsr = TimeSeriesRegressor(LinearRegression(), n_prev=n_prev)
tsr.fit(X_train)
fc = tsr.forecast(X_train, len(X_test))
def changes(X, start=0, end=-1):
return np.array(
[X[end, i] - X[start, i] for i in range(X.shape[1])])
best_is = changes(fc).argsort()[::-1]
for k in range(X.shape[1]):
bests[i, j, k] = changes(X_test)[best_is[k]] - np.mean(
changes(X_test))
return bests
class TSEBuyAndHoldStrategy(InformedBuyAndHoldStrategy):
def __init__(self,
balance,
tickers=None,
base_model=LinearRegression(),
n_prev=2,
wait=100,
steps_ahead=100,
k=5,
envelope='proportional',
log=False,
commission=.0002,
flat_rate=8):
if tickers is None:
tickers = s_and_p_names('2014-1-1', '2015-11-02')
super(TSEBuyAndHoldStrategy, self).__init__(balance,
tickers,
log=log,
commission=commission,
wait=wait,
flat_rate=flat_rate)
self.model = TimeSeriesRegressor(base_model, n_ahead=1, n_prev=n_prev)
self.steps_ahead = steps_ahead
self.k = k
self.envelope = envelope
def choose_stocks(self):
self.model.fit(self.observed_data)
fc = self.model.forecast(self.observed_data, self.steps_ahead)
changes = np.array([fc[-1, i] - fc[0, i] for i in range(fc.shape[1])])
top_k = changes.argsort()[::-1][:self.k]
top_tickers = self.names[top_k]
if self.envelope == 'proportional':
top_weights = changes[top_k]
elif self.envelope == 'log_proportional':
top_weights = np.log(changes[top_k])
elif self.envelope == 'uniform':
top_weights = np.ones((self.k))
else:
raise ValueError("Chose a proper strategy name")
top_weights = np.array(map(lambda w: max(0, w), top_weights))
top_weights = top_weights / float(sum(top_weights))
return top_tickers, top_weights
def __str__(self):
return "TSE Buy and Hold Strategy"
def get_data_dependece(X, data_sizes, folds=20, test_size=30, n_prev=2, log=True):
bests = np.empty((len(data_sizes), folds, X.shape[1]))
for i, data_size in enumerate(data_sizes):
pairs = cascade_cv(len(X), folds, data_size=data_size, test_size=test_size, number=True)
for j, pair in enumerate(pairs):
if log:
print('data size: {} trial {} '.format(data_size, j))
X_train, X_test = np.array(X.iloc[pair[0], :]), np.array(X.iloc[pair[1], :])
tsr = TimeSeriesRegressor(LinearRegression(), n_prev=n_prev)
tsr.fit(X_train)
fc = tsr.forecast(X_train, len(X_test))
def changes(X, start=0, end=-1):
return np.array([X[end, i] - X[start, i] for i in range(X.shape[1])])
best_is = changes(fc).argsort()[::-1]
for k in range(X.shape[1]):
bests[i, j, k] = changes(X_test)[best_is[k]] - np.mean(changes(X_test))
return bests
def __init__(self, balance, tickers=None, base_model=LinearRegression(),
n_prev=2, wait=100, steps_ahead=100, k=5, envelope='proportional',
log=False, commission=.0002, flat_rate=8):
if tickers is None:
tickers = s_and_p_names('2014-1-1', '2015-11-02')
super(TSEBuyAndHoldStrategy, self).__init__(balance, tickers,
log=log, commission=commission, wait=wait, flat_rate=flat_rate)
self.model = TimeSeriesRegressor(base_model, n_ahead=1, n_prev=n_prev)
self.steps_ahead = steps_ahead
self.k = k
self.envelope = envelope
class TSEBuyAndHoldStrategy(InformedBuyAndHoldStrategy):
def __init__(self, balance, tickers=None, base_model=LinearRegression(),
n_prev=2, wait=100, steps_ahead=100, k=5, envelope='proportional',
log=False, commission=.0002, flat_rate=8):
if tickers is None:
tickers = s_and_p_names('2014-1-1', '2015-11-02')
super(TSEBuyAndHoldStrategy, self).__init__(balance, tickers,
log=log, commission=commission, wait=wait, flat_rate=flat_rate)
self.model = TimeSeriesRegressor(base_model, n_ahead=1, n_prev=n_prev)
self.steps_ahead = steps_ahead
self.k = k
self.envelope = envelope
def choose_stocks(self):
self.model.fit(self.observed_data)
fc = self.model.forecast(self.observed_data, self.steps_ahead)
changes = np.array([fc[-1, i] - fc[0, i] for i in range(fc.shape[1])])
top_k = changes.argsort()[::-1][:self.k]
top_tickers = self.names[top_k]
if self.envelope == 'proportional':
top_weights = changes[top_k]
elif self.envelope == 'log_proportional':
top_weights = np.log(changes[top_k])
elif self.envelope == 'uniform':
top_weights = np.ones((self.k))
else:
raise ValueError("Chose a proper strategy name")
top_weights = np.array(map(lambda w: max(0, w), top_weights))
top_weights = top_weights / float(sum(top_weights))
return top_tickers, top_weights
def __str__(self):
return "TSE Buy and Hold Strategy"
from TimeSeriesEstimator import TimeSeriesRegressor, time_series_split
def mse(X1, X2, multioutput='raw_values'):
if multioutput == 'raw_values':
return np.mean((X1 - X2)**2, axis=0)**.5
if multioutput == 'uniform_average':
return np.mean(np.mean((X1 - X2)**2, axis=0)**.5)
X = datasets('sp500')
names = list(X.columns.values)
X_train, X_test = time_series_split(X)
n_prev = 2
tsr = TimeSeriesRegressor(LinearRegression(), n_prev=n_prev)
tsr.fit(X_train)
fc = tsr.forecast(X_train, len(X_test), noise=.2, n_paths=200)
fc_mean = tsr.forecast(X_train,
len(X_test),
noise=.2,
n_paths=200,
combine='mean')
#or for speed
#fc_mean = np.mean(fc, axis=0)
plt.plot(np.transpose(fc[:, :, 1]), 'r', alpha=.05)
plt.plot(np.transpose(fc_mean[:, 1]), 'b', label='Mean Forecast')
plt.plot(X_test[:, 1], 'g', label='Actual Price')
plt.legend()
plt.xlabel('days')
plt.plot(pred_test[:, i], 'r', label="Predicted")
plt.plot(y_test[n_prev:, i], 'b--', label="Actual")
# nprev: because the first predicted point needed n_prev steps of data
# plt.title("Testing performance of " + titles[i])
# plt.legend(loc='lower left')
plt.gcf().set_size_inches(15, 6)
plt.show()
X = datasets('sp500')
X_train, X_test = time_series_split(X)
n_prev = 3
tsr = TimeSeriesRegressor(LinearRegression(), n_prev=n_prev)
tsr.fit(X_train)
fc = tsr.forecast(X_train, len(X_test))
def changes(X):
return np.array([X[-1, i] - X[0, i] for i in range(X.shape[1])])
#X_test_change = np.log(changes(X_test))
#fc_change = np.log(changes(fc))
#plt.plot(np.linspace(0,8),np.linspace(0,8),'b')
#plt.plot(X_test_change,fc_change, 'ro')
for i in range(min(16,X_train.shape[1])):
plt.subplot(4,4,i+1)
plt.plot(fc[:, i],'r')
from sklearn.cross_validation import KFold
def mse(X1, X2, multioutput="raw_values"):
if multioutput == "raw_values":
return np.mean((X1 - X2) ** 2, axis=0) ** 0.5
if multioutput == "uniform_average":
return np.mean(np.mean((X1 - X2) ** 2, axis=0) ** 0.5)
X = datasets("sp500")
names = list(X.columns.values)
X_train, X_test = time_series_split(X)
n_prev = 2
tsr = TimeSeriesRegressor(LinearRegression(), n_prev=n_prev)
tsr.fit(X_train)
fc = tsr.forecast(X_train, len(X_test), noise=0.2, n_paths=200)
fc_mean = tsr.forecast(X_train, len(X_test), noise=0.2, n_paths=200, combine="mean")
# or for speed
# fc_mean = np.mean(fc, axis=0)
plt.plot(np.transpose(fc[:, :, 1]), "r", alpha=0.05)
plt.plot(np.transpose(fc_mean[:, 1]), "b", label="Mean Forecast")
plt.plot(X_test[:, 1], "g", label="Actual Price")
plt.legend()
plt.xlabel("days")
plt.ylabel("Price")
plt.title("Forecasting Alcoa (AA)")
plt.show()
