def get_growth(self):
if os.path.isfile('growth.csv') and not self.is_update:
growth = data.read_data(['growth'], ['growth'])
self.bb_data.factor['growth'] = growth.ix['growth']
else:
self.get_g_egrlf()
self.get_g_egrsf()
self.get_g_egro()
self.get_g_sgro()
self.bb_data.discard_uninv_data()
# 计算四个成分因子的暴露
self.bb_data.raw_data[
'egrlf_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['egrlf'],
self.bb_data.stock_price.ix['FreeMarketValue'])
self.bb_data.raw_data[
'egrsf_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['egrsf'],
self.bb_data.stock_price.ix['FreeMarketValue'])
self.bb_data.raw_data[
'egro_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['egro'],
self.bb_data.stock_price.ix['FreeMarketValue'])
self.bb_data.raw_data[
'sgro_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['sgro'],
self.bb_data.stock_price.ix['FreeMarketValue'])
growth = 0.18*self.bb_data.raw_data.ix['egrlf_expo']+0.11*self.bb_data.raw_data.ix['egrsf_expo']+ \
0.24*self.bb_data.raw_data.ix['egro_expo']+0.47*self.bb_data.raw_data.ix['sgro_expo']
self.bb_data.factor['growth'] = growth
def get_earnings_yeild(self):
if os.path.isfile('ey.csv') and not self.is_update:
EarningsYield = data.read_data(['ey'], ['ey'])
self.bb_data.factor['ey'] = EarningsYield.ix['ey']
else:
self.get_ey_epfwd()
self.get_ey_cetop()
self.get_ey_etop()
self.bb_data.discard_uninv_data()
# 计算三个成分因子的暴露
self.bb_data.raw_data[
'epfwd_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['epfwd'],
self.bb_data.stock_price.ix['FreeMarketValue'])
self.bb_data.raw_data[
'cetop_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['cetop'],
self.bb_data.stock_price.ix['FreeMarketValue'])
self.bb_data.raw_data[
'etop_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['etop'],
self.bb_data.stock_price.ix['FreeMarketValue'])
EarningsYield = 0.68*self.bb_data.raw_data.ix['epfwd_expo']+0.21*self.bb_data.raw_data.ix['cetop_expo']+ \
0.11*self.bb_data.raw_data.ix['etop_expo']
self.bb_data.factor['ey'] = EarningsYield
def get_liquidity(self):
if os.path.isfile('liquidity.csv') and not self.is_update:
liquidity = data.read_data(['liquidity'], ['liquidity'])
self.bb_data.factor['liquidity'] = liquidity.ix['liquidity']
else:
self.get_liq_stom()
self.get_liq_stoq()
self.get_liq_stoa()
# 过滤数据
self.bb_data.discard_uninv_data()
# 计算三个成分因子的暴露
self.bb_data.raw_data[
'stom_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['stom'],
self.bb_data.stock_price.ix['FreeMarketValue'])
self.bb_data.raw_data[
'stoq_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['stoq'],
self.bb_data.stock_price.ix['FreeMarketValue'])
self.bb_data.raw_data[
'stoa_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['stoa'],
self.bb_data.stock_price.ix['FreeMarketValue'])
liquidity = 0.35*self.bb_data.raw_data.ix['stom_expo']+0.35*self.bb_data.raw_data.ix['stoq_expo']+ \
0.3*self.bb_data.raw_data.ix['stoa_expo']
# 计算liquidity的因子暴露，不再去极值
y = strategy_data.get_cap_wgt_exposure(
liquidity,
self.bb_data.stock_price.ix['FreeMarketValue'],
percentile=0)
# 计算市值因子的暴露
x = pd.Panel({
'lncap_expo':
strategy_data.get_cap_wgt_exposure(
self.bb_data.factor.ix['lncap'],
self.bb_data.stock_price.ix['FreeMarketValue'])
})
# 正交化
new_liq = strategy_data.simple_orth_gs(
y,
x,
weights=np.sqrt(
self.bb_data.stock_price.ix['FreeMarketValue']))[0]
self.bb_data.factor['liquidity'] = new_liq
def get_style_factor_exposure(self):
# 给因子暴露panel加上索引
self.bb_data.factor_expo = pd.Panel(
data=None,
major_axis=self.bb_data.factor.major_axis,
minor_axis=self.bb_data.factor.minor_axis)
# 循环计算暴露
for item, df in self.bb_data.factor.iteritems():
# 通过内部因子加总得到的因子，或已经计算过一次暴露的因子（如正交化过），不再需要去极值
if item in ['rv', 'nls', 'liquidity', 'ey', 'growth']:
self.bb_data.factor_expo[
item] = strategy_data.get_cap_wgt_exposure(
df,
self.bb_data.stock_price.ix['FreeMarketValue'],
percentile=0)
else:
self.bb_data.factor_expo[
item] = strategy_data.get_cap_wgt_exposure(
df, self.bb_data.stock_price.ix['FreeMarketValue'])
def select_stocks_pure_factor_bb(self,
*,
bb_expo,
cov_matrix='Empty',
reg_weight='Empty',
direction='+',
regulation_lambda=1):
# 计算因子值的暴露
factor_expo = strategy_data.get_cap_wgt_exposure(
self.strategy_data.factor.iloc[0],
self.strategy_data.stock_price.ix['FreeMarketValue'])
if direction == '-':
factor_expo = -factor_expo
self.strategy_data.factor_expo = pd.Panel(
{'factor_expo': factor_expo},
major_axis=self.strategy_data.factor.major_axis,
minor_axis=self.strategy_data.factor.minor_axis)
# 循环调仓日
for cursor, time in self.holding_days.iteritems():
# 当前的因子暴露向量，为n*1
x_alpha = self.strategy_data.factor_expo.ix['factor_expo',
time, :].fillna(0)
# 当前的其他因子暴露向量，为n*(k-1)，实际就是barra base因子的暴露
x_sigma = bb_expo.ix[:, time, :].fillna(0)
# 有协方差矩阵，优先用协方差矩阵
if type(cov_matrix) != str:
inv_v = np.linalg.pinv(cov_matrix.ix[time].fillna(0))
else:
assert type(reg_weight) != str, 'The construction of pure factor portfolio require one of following:\n' \
'Covariance matrix of factor returns (priority), OR \n' \
'Regression weight when getting factor return using linear regression.\n'
# 取当期的回归权重，每只股票的权重在对角线上
# inv_v = np.diag(reg_weight.ix[time].fillna(0))
curr_weight = reg_weight.ix[time]
curr_weight = (curr_weight / curr_weight.sum()).fillna(0)
inv_v = np.diag(curr_weight)
# 通过优化的解析解计算权重，解析解公式见barra, Efficient Replication of Factor Returns, equation (6)
temp_1 = np.linalg.pinv(np.dot(np.dot(x_sigma.T, inv_v), x_sigma))
temp_2 = np.dot(np.dot(x_sigma.T, inv_v), x_alpha)
temp_3 = x_alpha - np.dot(np.dot(x_sigma, temp_1), temp_2)
h_star = 1 / regulation_lambda * np.dot(inv_v, temp_3)
# 加权方式只能为这一种，只是需要归一化一下
self.position.holding_matrix.ix[time] = h_star
self.position.to_percentage()
pass
def get_nonlinear_size(self):
if os.path.isfile('nls.csv') and not self.is_update:
nls = data.read_data(['nls'], ['nls'])
self.bb_data.factor['nls'] = nls.ix['nls']
else:
size_cube = self.bb_data.factor.ix['lncap']**3
# 计算原始nls的暴露
y = strategy_data.get_cap_wgt_exposure(
size_cube, self.bb_data.stock_price.ix['FreeMarketValue'])
# 计算市值因子的暴露，注意解释变量需要为一个panel
x = pd.Panel({
'lncap_expo':
strategy_data.get_cap_wgt_exposure(
self.bb_data.factor.ix['lncap'],
self.bb_data.stock_price.ix['FreeMarketValue'])
})
# 对市值因子做正交化
new_nls = strategy_data.simple_orth_gs(
y,
x,
weights=np.sqrt(
self.bb_data.stock_price.ix['FreeMarketValue']))[0]
self.bb_data.factor['nls'] = new_nls
def construct_factor(self):
# 直接使用runner value中的rv8, rv36
self.strategy_data.raw_data = data.read_data(
['runner_value_8', 'runner_value_36'], item_name=['rv8', 'rv36'])
self.strategy_data.stock_price = data.read_data(['FreeMarketValue'])
# # 在投资域内进行标准化及回归
# self.strategy_data.discard_uninv_data()
# 由于rv没有进行标准化, 在这里进行标准化, 进行市值加权标准化
self.strategy_data.raw_data[
'rv8'] = strategy_data.get_cap_wgt_exposure(
self.strategy_data.raw_data['rv8'],
self.strategy_data.stock_price['FreeMarketValue'])
self.strategy_data.raw_data[
'rv36'] = strategy_data.get_cap_wgt_exposure(
self.strategy_data.raw_data['rv36'],
self.strategy_data.stock_price['FreeMarketValue'])
# 等权回归
reg_weight = None
# 以根号市值作为回归权重
# reg_weight = np.sqrt(self.strategy_data.stock_price['FreeMarketValue'])
outcome = strategy_data.simple_orth_gs(
self.strategy_data.raw_data['rv8'],
self.strategy_data.raw_data[['rv36']],
weights=reg_weight)
new_factor = outcome[0]
pvalues = outcome[1]
rsquared = outcome[2]
# 储存因子
self.strategy_data.factor = pd.Panel(
{'new_reversal': new_factor.shift(1)})
pass
def get_residual_volatility(self):
if os.path.isfile('rv.csv') and not self.is_update:
rv = data.read_data(['rv'], ['rv'])
self.bb_data.factor['rv'] = rv.ix['rv']
else:
self.get_rv_dastd()
self.get_rv_cmra()
self.get_rv_hsigma()
# 过滤数据，因为之前的因子数据之后要正交化，会影响计算
# 此处为barra base计算中第一次过滤掉uninv数据，此后的数据都不能再储存，因为依赖于stock pool
self.bb_data.discard_uninv_data()
# 计算三个成分因子的暴露
self.bb_data.raw_data[
'dastd_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['dastd'],
self.bb_data.stock_price.ix['FreeMarketValue'])
self.bb_data.raw_data[
'cmra_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['cmra'],
self.bb_data.stock_price.ix['FreeMarketValue'])
self.bb_data.raw_data[
'hsigma_expo'] = strategy_data.get_cap_wgt_exposure(
self.bb_data.raw_data.ix['hsigma'],
self.bb_data.stock_price.ix['FreeMarketValue'])
rv = 0.74*self.bb_data.raw_data.ix['dastd_expo']+0.16*self.bb_data.raw_data.ix['cmra_expo']+ \
0.1*self.bb_data.raw_data.ix['hsigma_expo']
# 计算rv的因子暴露，不再去极值
y = strategy_data.get_cap_wgt_exposure(
rv,
self.bb_data.stock_price.ix['FreeMarketValue'],
percentile=0)
# 计算市值因子与beta因子的暴露
x = pd.Panel({
'lncap_expo':
strategy_data.get_cap_wgt_exposure(
self.bb_data.factor.ix['lncap'],
self.bb_data.stock_price.ix['FreeMarketValue']),
'beta_expo':
strategy_data.get_cap_wgt_exposure(
self.bb_data.factor.ix['beta'],
self.bb_data.stock_price.ix['FreeMarketValue'])
})
# 正交化
new_rv = strategy_data.simple_orth_gs(
y,
x,
weights=np.sqrt(
self.bb_data.stock_price.ix['FreeMarketValue']))[0]
# 之后会再次的计算暴露，注意再次计算暴露后，new_rv依然保有对x的正交性
self.bb_data.factor['rv'] = new_rv
def get_nonlinear_size(self):
if os.path.isfile(os.path.abspath('.')+'/ResearchData/nls_total'+self.filename_appendix) \
and not self.is_update and self.try_to_read:
self.base_data.factor['nls'] = data.read_data(
'nls' + self.filename_appendix)
else:
# 计算市值因子的暴露，注意解释变量需要为一个panel
x = pd.Panel({
'lncap_expo':
strategy_data.get_cap_wgt_exposure(
self.base_data.factor.ix['lncap'],
self.base_data.stock_price.ix['FreeMarketValue'])
})
# 将市值因子暴露取3次方, 得到size cube
y = x['lncap_expo']**3
# 将size cube对市值因子做正交化
new_nls = strategy_data.simple_orth_gs(
y,
x,
weights=np.sqrt(
self.base_data.stock_price.ix['FreeMarketValue']))[0]
self.base_data.factor['nls'] = new_nls
def get_pure_factor_gs_orth(self,
base_expo,
*,
do_active_bb_pure_factor=False,
reg_weight=1,
add_constant=False,
use_factor_expo=True,
expo_weight=1):
# 计算当前因子的暴露，注意策略里的数据都已经lag过了
if use_factor_expo:
if expo_weight == 1:
factor_expo = strategy_data.get_cap_wgt_exposure(
self.strategy_data.factor.iloc[0],
self.strategy_data.stock_price.ix['FreeMarketValue'])
elif expo_weight == 0:
factor_expo = strategy_data.get_exposure(
self.strategy_data.factor.iloc[0])
# 如果计算的是相对基准的纯因子收益率
if do_active_bb_pure_factor:
if self.strategy_data.stock_pool == 'all':
benchmark_weight = data.read_data(['Weight_zz500'],
['Weight_zz500'],
shift=True)
else:
benchmark_weight = self.strategy_data.benchmark_price.ix[
'Weight_' + self.strategy_data.stock_pool]
# 计算bb base的调整后暴露，以及调整后benchmark在bb base上的暴露
adjusted_bb_expo = strategy_data.adjust_benchmark_related_expo(
base_expo, benchmark_weight,
self.strategy_data.if_tradable.ix['if_tradable'])
benchmark_bb_expo = np.einsum('ijk,jk->ji',
adjusted_bb_expo.fillna(0),
benchmark_weight.fillna(0))
benchmark_bb_expo = pd.DataFrame(benchmark_bb_expo,
index=base_expo.major_axis,
columns=base_expo.items)
# 计算当前因子的调整后暴露值，以及调整后benchmark在当前因子上的暴露
adjusted_factor_expo = strategy_data.adjust_benchmark_related_expo(
self.strategy_data.factor, benchmark_weight,
self.strategy_data.if_tradable.ix['if_tradable'])
adjusted_factor_expo = adjusted_factor_expo.iloc[0]
benchmark_factor_expo = (adjusted_factor_expo *
benchmark_weight).sum(1)
# 用暴露的绝对值减去基准的暴露值，得到相对基准的超额暴露值
base_expo = base_expo.sub(benchmark_bb_expo, axis=0)
factor_expo = factor_expo.sub(benchmark_factor_expo, axis=0)
# 在bb expo里去掉国家因子，去掉是为了保证有唯一解，而且去掉后残差值不变，不影响结果
# 因为国家因子向量已经能表示成行业暴露的线性组合了
if 'country_factor' in base_expo.items:
base_expo_no_cf = base_expo.drop('country_factor', axis=0)
else:
base_expo_no_cf = base_expo
# 利用多元线性回归进行提纯
if reg_weight == 1:
pure_factor_expo = strategy_data.simple_orth_gs(
factor_expo,
base_expo_no_cf,
weights=np.sqrt(
self.strategy_data.stock_price.ix['FreeMarketValue']),
add_constant=add_constant)[0]
elif reg_weight == 0:
pure_factor_expo = strategy_data.simple_orth_gs(
factor_expo, base_expo_no_cf, add_constant=add_constant)[0]
# 将得到的纯化因子放入因子值中储存
self.strategy_data.factor.iloc[0] = pure_factor_expo
def select_stocks(self,
*,
select_ratio=[0.8, 1],
direction='+',
weight=0,
use_factor_expo=True,
expo_weight=1):
# 对调仓期进行循环
for cursor, time in self.holding_days.iteritems():
curr_factor_data = self.strategy_data.factor.ix[0, time, :]
# 对因子值进行排序，注意这里的秩（rank），类似于得分
if direction is '+':
factor_score = curr_factor_data.rank(ascending=True)
elif direction is '-':
factor_score = curr_factor_data.rank(ascending=False)
else:
print('Please enter ' '+' ' or ' '-' ' for direction argument')
# 取有效的股票数
effective_num = curr_factor_data.dropna().size
# 无股票可选，进行下一次循环
if effective_num == 0:
continue
# 选取股票的得分范围
lower_bound = np.floor(effective_num * select_ratio[0])
upper_bound = np.floor(effective_num * select_ratio[1])
# 选取股票
selected_stocks = curr_factor_data.ix[np.logical_and(
factor_score >= lower_bound,
factor_score <= upper_bound)].index
# 被选取的股票都将持仓调为1
self.position.holding_matrix.ix[time, selected_stocks] = 1
# 循环结束
if self.strategy_data.stock_pool == 'all':
# 去除不可交易的股票
self.filter_untradable()
else:
# 有股票池的情况去除不可投资的股票
self.filter_uninv()
# 设置为等权重
self.position.to_percentage()
# 如果需要市值加权，则市值加权
if weight == 1:
self.position.weighted_holding(self.strategy_data.stock_price.ix[
'FreeMarketValue', self.position.holding_matrix.index, :])
# 如果是因子加权, 则进行因子值加权
elif weight == 2:
# 看是否需要计算因子暴露, 用因子暴露值进行加权
if use_factor_expo:
if expo_weight == 1:
factor_weight = strategy_data.get_cap_wgt_exposure(
self.strategy_data.factor.iloc[0],
self.strategy_data.stock_price.ix['FreeMarketValue'])
elif expo_weight == 0:
factor_weight = strategy_data.get_exposure(
self.strategy_data.factor.iloc[0])
else:
factor_weight = self.strategy_data.factor.iloc[0]
# 进行因子值加权的权重计算
self.position.weighted_holding(
factor_weight.ix[self.position.holding_matrix.index, :])
pass
def get_factor_return(self,
*,
holding_freq='m',
weights='default',
direction='+',
plot_cum=True,
start='default',
end='default'):
# 如果没有price的数据，读入price数据，注意要shift，
# 即本来的实现收益率应当是调仓日当天的开盘价，但这里计算调仓日前一个交易日的收盘价。
if 'ClosePrice_adj' not in self.strategy_data.stock_price.items:
temp_panel = data.read_data(['ClosePrice_adj'], ['ClosePrice_adj'],
shift=True)
self.strategy_data.stock_price['ClosePrice_adj'] = temp_panel.ix[
'ClosePrice_adj']
# 计算因子收益的频率
holding_days = strategy.resample_tradingdays(self.strategy_data.stock_price.\
ix['FreeMarketValue', :, 0], freq=holding_freq)
# 如果有指定，只取start和end之间的时间计算
if start != 'default':
holding_days = holding_days[start:]
if end != 'default':
holding_days = holding_days[:end]
# 计算股票对数收益以及因子暴露
holding_day_price = self.strategy_data.stock_price.ix['ClosePrice_adj',
holding_days, :]
holding_day_return = np.log(
holding_day_price.div(holding_day_price.shift(1)))
holding_day_factor = self.strategy_data.factor.ix[0, holding_days, :]
holding_day_factor_expo = strategy_data.get_cap_wgt_exposure(
holding_day_factor,
self.strategy_data.stock_price.ix['FreeMarketValue',
holding_days, :])
# 注意因子暴露要用前一期的数据
holding_day_factor_expo = holding_day_factor_expo.shift(1)
# 初始化因子收益序列以及估计量的t统计量序列
factor_return_series = np.empty(holding_days.size) * np.nan
t_stats_series = np.empty(holding_days.size) * np.nan
self.factor_return_series = pd.Series(factor_return_series,
index=holding_days)
self.t_stats_series = pd.Series(t_stats_series, index=holding_days)
# 进行回归，对调仓日进行循环
for cursor, time in holding_days.iteritems():
y = holding_day_return.ix[time, :]
x = holding_day_factor_expo.ix[time, :]
if y.isnull().all() or x.isnull().all():
continue
x = sm.add_constant(x)
if weights is 'default':
results = sm.WLS(y, x, missing='drop').fit()
else:
results = sm.WLS(y,
x,
weights=weights.ix[time],
missing='drop').fit()
self.factor_return_series.ix[time] = results.params[1]
self.t_stats_series.ix[time] = results.tvalues[1]
# 如果方向为负，则将因子收益和t统计量加个负号
if direction == '-':
self.factor_return_series = -self.factor_return_series
self.t_stats_series = -self.t_stats_series
# 输出的string
tstats_sig_ratio = self.t_stats_series[
np.abs(self.t_stats_series) >= 2].size / self.t_stats_series.size
target_str = 'The average return of this factor: {0:.4f}%\n' \
'Note that the return of factor is not annualized but corresponding to the holding days interval\n' \
'The average t-statistics value: {1:.4f}\n' \
'Ratio of t_stats whose absolute value >= 2: {2:.2f}%\n'.format(
self.factor_return_series.mean()*100, self.t_stats_series.mean(), tstats_sig_ratio*100
)
# 循环结束，输出结果
print(target_str)
with open(str(os.path.abspath('.')) + '/' +
self.strategy_data.stock_pool + '/performance.txt',
'a',
encoding='GB18030') as text_file:
text_file.write(target_str)
# 画图，默认画因子收益的累计收益图
fx = plt.figure()
ax = fx.add_subplot(1, 1, 1)
zero_series = pd.Series(np.zeros(self.factor_return_series.shape),
index=self.factor_return_series.index)
if plot_cum:
plt.plot(self.factor_return_series.cumsum() * 100, 'b-')
else:
plt.plot(self.factor_return_series * 100, 'b-')
plt.plot(zero_series, 'r-')
ax.set_xlabel('Time')
ax.set_ylabel('Return of The Factor (%)')
ax.set_title('The Return Series of The Factor')
plt.savefig(str(os.path.abspath('.')) + '/' +
self.strategy_data.stock_pool + '/' + 'FactorReturn.png',
dpi=1200)
if type(self.pdfs) != str:
plt.savefig(self.pdfs, format='pdf')
fx = plt.figure()
ax = fx.add_subplot(1, 1, 1)
plt.plot(self.t_stats_series, 'b-')
plt.plot(zero_series, 'r-')
ax.set_xlabel('Time')
ax.set_ylabel('T-Stats of The Factor Return')
ax.set_title('The T-Stats Series of The Factor Return')
plt.savefig(str(os.path.abspath('.')) + '/' +
self.strategy_data.stock_pool + '/' +
'FactorReturnTStats.png',
dpi=1200)
if type(self.pdfs) != str:
plt.savefig(self.pdfs, format='pdf')
def select_stocks_pure_factor(self,
*,
base_expo,
cov_matrix='Empty',
reg_weight='Empty',
direction='+',
benchmark_weight='Empty',
is_long_only=True):
# 计算因子值的暴露
factor_expo = strategy_data.get_cap_wgt_exposure(
self.strategy_data.factor.iloc[0],
self.strategy_data.stock_price.ix['FreeMarketValue'])
if direction == '-':
factor_expo = -factor_expo
self.strategy_data.factor_expo = pd.Panel(
{'factor_expo': factor_expo},
major_axis=self.strategy_data.factor.major_axis,
minor_axis=self.strategy_data.factor.minor_axis)
# 如果有benchmark，则计算benchmark的暴露
if type(benchmark_weight) != str:
benchmark_weight = (benchmark_weight.div(benchmark_weight.sum(1),
axis=0)).fillna(0)
adjusted_base_expo = strategy_data.adjust_benchmark_related_expo(
base_expo, benchmark_weight,
self.strategy_data.if_tradable.ix['if_tradable'])
benchmark_base_expo = np.einsum('ijk,jk->ji',
adjusted_base_expo.fillna(0),
benchmark_weight.fillna(0))
benchmark_base_expo = pd.DataFrame(benchmark_base_expo,
index=base_expo.major_axis,
columns=base_expo.items)
adjusted_factor_expo = strategy_data.adjust_benchmark_related_expo(
pd.Panel({'factor_expo': factor_expo}), benchmark_weight,
self.strategy_data.if_tradable.ix['if_tradable'])
adjusted_factor_expo = adjusted_factor_expo.ix['factor_expo']
benchmark_curr_factor_expo = (adjusted_factor_expo *
benchmark_weight).sum(1)
self.strategy_data.factor_expo.ix['factor_expo'] = factor_expo.sub(
benchmark_curr_factor_expo, axis=0)
# 循环调仓日
for cursor, time in self.holding_days.iteritems():
curr_factor_expo = self.strategy_data.factor_expo.ix['factor_expo',
time, :]
curr_base_expo = base_expo.ix[:, time, :]
# 有协方差矩阵，优先用协方差矩阵
if type(cov_matrix) != str:
curr_v = cov_matrix.ix[time]
curr_v_diag = curr_v.diagonal()
# 去除有nan的数据
all_data = pd.concat(
[curr_v_diag, curr_factor_expo, curr_base_expo], axis=1)
all_data = all_data.dropna()
# 如果有效数据小于等于1，当期不选股票
if all_data.shape[0] <= 1:
continue
# 指数中选股可能会出现一个行业暴露全是0的情况，所以关于这个行业的限制条件会冗余，于是要进行剔除
all_data = all_data.replace(0, np.nan).dropna(
axis=1, how='all').fillna(0.0)
curr_factor_expo = all_data.ix[:, 0]
curr_v_diag = all_data.ix[:, 1]
curr_base_expo = all_data.ix[:, 2:]
curr_v = curr_v.reindex(index=curr_v_diag.index,
columns=curr_v_diag.index)
else:
assert type(reg_weight) != str, 'The construction of pure factor portfolio require one of following:\n' \
'Covariance matrix of factor returns (priority), OR \n' \
'Regression weight when getting factor return using linear regression.\n'
# 取当期的回归权重，每只股票的权重在对角线上
curr_v_diag = reg_weight.ix[time]
# 去除有nan的数据
all_data = pd.concat(
[curr_v_diag, curr_factor_expo, curr_base_expo], axis=1)
all_data = all_data.dropna()
# 如果有效数据小于等于1，当期不选股票
if all_data.shape[0] <= 1:
continue
# 指数中选股可能会出现一个行业暴露全是0的情况，所以关于这个行业的限制条件会冗余，于是要进行剔除
all_data = all_data.replace(0, np.nan).dropna(
axis=1, how='all').fillna(0.0)
curr_v_diag = all_data.ix[:, 0]
curr_factor_expo = all_data.ix[:, 1]
curr_base_expo = all_data.ix[:, 2:]
# 将回归权重归一化
curr_v_diag = curr_v_diag / curr_v_diag.sum()
curr_v = np.linalg.pinv(np.diag(curr_v_diag))
curr_v = pd.DataFrame(curr_v,
index=curr_factor_expo.index,
columns=curr_factor_expo.index)
# 设置其他因子为0的限制条件，在有基准的时候，设置为基准的暴露
if type(benchmark_weight) != str:
expo_target = benchmark_base_expo.ix[time].reindex(
index=curr_base_expo.columns)
else:
expo_target = pd.Series(0.0, index=curr_base_expo.columns)
# 开始设置优化
# P = V
P = matrix(curr_v.as_matrix())
# q = - (factor_expo.T)
q = matrix(-curr_factor_expo.as_matrix().transpose())
# 其他因为暴露为0，或等于基准的限制条件
A = matrix(curr_base_expo.as_matrix().transpose())
b = matrix(expo_target.as_matrix())
solvers.options['show_progress'] = False
# 如果只能做多，则每只股票的比例都必须大于等于0
if is_long_only:
long_only_constraint = pd.DataFrame(
-1.0 * np.eye(curr_factor_expo.size),
index=curr_factor_expo.index,
columns=curr_factor_expo.index)
long_only_target = pd.Series(0.0, index=curr_factor_expo.index)
G = matrix(long_only_constraint.as_matrix())
h = matrix(long_only_target.as_matrix())
# 解优化问题
results = solvers.qp(P=P, q=q, A=A, b=b, G=G, h=h)
else:
results = solvers.qp(P=P, q=q, A=A, b=b)
results_np = np.array(results['x']).squeeze()
results_s = pd.Series(results_np, index=curr_factor_expo.index)
# 重索引为所有股票代码
results_s = results_s.reindex(
self.strategy_data.stock_price.minor_axis, fill_value=0)
# 股票持仓
self.position.holding_matrix.ix[time] = results_s
# 循环结束后，进行权重归一化
self.position.to_percentage()
pass
def get_table3(self,
*,
freq='w',
foldername=None,
startdate=None,
enddate=None):
# 首先需要按照频率生成holdingdays
self.generate_holding_days(holding_freq=freq,
loc=-1,
start_date=startdate,
end_date=enddate)
# 读取数据
self.strategy_data.raw_data['rv3'] = data.read_data('runner_value_3')
# 对rv3进行标准化
self.strategy_data.raw_data[
'rv3'] = strategy_data.get_cap_wgt_exposure(
self.strategy_data.raw_data['rv3'],
self.strategy_data.stock_price['FreeMarketValue'])
self.strategy_data.stock_price['ClosePrice_adj'] = data.read_data(
'ClosePrice_adj')
self.strategy_data.stock_price['daily_return'] = \
self.strategy_data.stock_price['ClosePrice_adj'].pct_change()
# 按照频率算收益率, 和holdingdays同步, 论文用月, 我们一般用w
r = self.strategy_data.stock_price['daily_return', startdate:enddate, :].\
resample(freq).sum()
# 注意, 回归的左边是未来一期的收益率, 因此要shift(-1), 即用到未来数据
r = r.shift(-1).dropna(how='all')
# 因为r的index为月末, 但是月末不一定是交易日, 因此将r的index重置为holding days
r = r.set_index(self.holding_days)
# 用于回归的右边
reg_panel = pd.Panel({
'lagged_ep':
self.strategy_data.raw_data['rv3'].shift(0),
'sue':
self.strategy_data.raw_data['rv36'],
'reversal':
self.strategy_data.raw_data['rv8']
})
# 储存table3的结果
table3 = pd.Panel(
items=['coef', 't_stats'],
major_axis=np.arange(5),
minor_axis=['intercept', 'lagged_ep', 'sue', 'reversal'])
# 使用holding days中的日期进行回归,
# 1. 用lagged lbm回归
results1 = strategy_data.fama_macbeth(
r, reg_panel.ix[['lagged_ep'], self.holding_days, :])
table3.ix['coef', 0, :] = results1[0]
table3.ix['t_stats', 0, :] = results1[1]
# 2. 使用bv return回归
results2 = strategy_data.fama_macbeth(
r, reg_panel.ix[['sue'], self.holding_days, :])
table3.ix['coef', 1, :] = results2[0]
table3.ix['t_stats', 1, :] = results2[1]
# 3. 使用lagged return回归
results3 = strategy_data.fama_macbeth(
r, reg_panel.ix[['reversal'], self.holding_days, :])
table3.ix['coef', 2, :] = results3[0]
table3.ix['t_stats', 2, :] = results3[1]
# 4. 使用lagged lbm与bv return回归
results4 = strategy_data.fama_macbeth(
r, reg_panel.ix[['lagged_ep', 'reversal'], self.holding_days, :])
table3.ix['coef', 3, :] = results4[0]
table3.ix['t_stats', 3, :] = results4[1]
# 5. 使用lagged lbm, bv return, lagged return一起回归
results5 = strategy_data.fama_macbeth(
r,
reg_panel.ix[['lagged_ep', 'sue', 'reversal'],
self.holding_days, :],
nw_lags=0)
table3.ix['coef', 4, :] = results5[0]
table3.ix['t_stats', 4, :] = results5[1]
# # 储存信息
# if foldername is None:
# table3.ix['coef'].to_csv(str(os.path.abspath('.')) + '/' + self.strategy_data.stock_pool +
# '/' + 'Table3_coef.csv', na_rep='N/A', encoding='GB18030')
# table3.ix['t_stats'].to_csv(str(os.path.abspath('.')) + '/' + self.strategy_data.stock_pool +
# '/' + 'Table3_t_stats.csv', na_rep='N/A', encoding='GB18030')
# else:
# table3.ix['coef'].to_csv(foldername +
# '/' + 'Table3_coef.csv', na_rep='N/A', encoding='GB18030')
# table3.ix['t_stats'].to_csv(foldername +
# '/' + 'Table3_t_stats.csv', na_rep='N/A', encoding='GB18030')
# pass
# 尝试使用linearmodels包
reg_panel['const'] = 1.0
fm_result1 = FamaMacBeth(
r, reg_panel.ix[['lagged_ep', 'const'],
self.holding_days, :]).fit()
fm_result2 = FamaMacBeth(
r, reg_panel.ix[['sue', 'const'], self.holding_days, :]).fit()
fm_result3 = FamaMacBeth(
r, reg_panel.ix[['reversal', 'const'],
self.holding_days, :]).fit()
fm_result4 = FamaMacBeth(
r, reg_panel.ix[['reversal', 'lagged_ep', 'const'],
self.holding_days, :]).fit()
fm_result5 = FamaMacBeth(
r, reg_panel.ix[['reversal', 'sue', 'const'],
self.holding_days, :]).fit()
fm_result6 = FamaMacBeth(r, reg_panel.ix[:,
self.holding_days, :]).fit()
pass
