def __init__(self,
alpha=None,
retain_alpha_sign=True,
universe='top2000',
delay=1,
capital=1e6):
if universe in stock_universes:
self.universe = universe
uni_file = database_dir + universe + '.adj_ammend.mat'
uni_var = scipy.io.loadmat(uni_file)[universe].astype(float)
uni_var[uni_var == 0.] = np.nan
self._uni_mask = uni_var
else:
raise KeyError('stock universe does not exist')
self.alpha = alpha
self.retain_alpha_sign = retain_alpha_sign
if not isinstance(delay, int) or delay < 0:
raise ValueError('delay must be int and >=0')
self.delay = delay
self.capital = capital
self._market_data = {}
dates = Simulator._read_var('dates')
dates = [time.strptime(str(idate), "[%Y%m%d]") for idate in dates]
dates = [datetime.datetime(*idate[0:6]) for idate in dates]
self.dates = dates
open_p = Simulator._read_var('open', self._uni_mask)
open_ret1 = open_p / trans.ts_delay(open_p, 1) - 1.0
self._market_data['open_ret1'] = open_ret1
close_p = Simulator._read_var('close', self._uni_mask)
close_ret1 = close_p / trans.ts_delay(close_p, 1) - 1.0
self._market_data['close_ret1'] = close_ret1
vwap = Simulator._read_var('vwap', self._uni_mask)
vwap_ret1 = vwap / trans.ts_delay(vwap, 1) - 1.0
Simulator._clean_data('vwap_ret1', vwap_ret1)
forward_vwap_ret1 = trans.ts_delay(vwap_ret1, -1 - self.delay)
self._market_data['vwap_ret1'] = vwap_ret1
self._market_data['forward_vwap_ret1'] = forward_vwap_ret1
return
def momen_smooth_inner(mysim):
momen_signal = np.sign(
trans.ts_delay(trans.ts_mean(mysim('ret1'), window_width),
window_width))
ret1_shock = np.zeros_like(mysim('ret1'))
ret1_shock[mysim('ret1') > threshold] = 1
ret1_shock[mysim('ret1') < -threshold] = -1
pshock, nshock = stats_util.time_to_lastshock(ret1_shock)
shock_index = np.logical_or(pshock < 60, nshock < 60)
momen_signal[np.logical_and(momen_signal > 0, ~shock_index)] = np.nan
momen_signal[np.logical_and(momen_signal < 0, shock_index)] = np.nan
return momen_signal
def momen_longterm_inner(mysim):
ave_past_ret = trans.ts_mean(
mysim('ret1'), lookback_period[1] - lookback_period[0] + 1)
ret1_shock = np.zeros_like(mysim('ret1'))
ret1_shock[mysim('ret1') > threshold] = 1
ret1_shock[mysim('ret1') < -threshold] = -1
pshock, nshock = stats_util.time_to_lastshock(ret1_shock)
shock_index = np.logical_or(pshock <= shock_effect_length,
nshock <= shock_effect_length)
ave_past_ret[shock_index] = np.nan
ave_past_ret = trans.ts_delay(ave_past_ret, -lookback_period[1])
alpha = trans.ts_make_squarewave(ave_past_ret, hold_length)
return alpha
def series_to_nfeatures(data2d, num_features, additional_feature=None):
"""from time series `data2d` select past `num_features` times to be used as
features. For example, time series [0, 1, 3, 5] and num_features = 2 becomes
[[nan, 0],
[0, 1],
[1, 3],
[3, 5]]
treat all the stocks as the same
"""
from trans import ts_delay
num_stocks = data2d.shape[0]
num_times = data2d.shape[1]
if additional_feature is None:
feature_vec = np.zeros((num_times*num_stocks, num_features))
else:
feature_vec = np.zeros((num_times*num_stocks, num_features+1))
for i in range(0, num_features):
feature_vec[:,i] = ts_delay(data2d, i).flatten()
if additional_feature is not None:
feature_vec[:,num_features] = np.tile(additional_feature, (1, num_stocks))
return feature_vec
def momen_HF_inner(mysim):
return trans.ts_delay(mysim(varname), delayed_days)
def momen_longterm_inner(mysim):
ave_past_ret = trans.ts_mean(
mysim('ret1'), lookback_period[1] - lookback_period[0] + 1)
ave_past_ret = trans.ts_delay(ave_past_ret, -lookback_period[1])
alpha = trans.ts_make_squarewave(ave_past_ret, hold_length)
return alpha