def out_sample_perf(pnl_df_out, way=1, cut_point_list=None):
# 根据sharpe大小,统计样本外的表现
if cut_point_list is None:
cut_point_list = [0.30]
if way == 1:
rolling_sharpe, cut_sharpe = \
bt.AZ_Rolling_sharpe(pnl_df_out, roll_year=0.5, year_len=250, cut_point_list=cut_point_list, output=True)
else:
rolling_sharpe, cut_sharpe = \
bt.AZ_Rolling_sharpe(-pnl_df_out, roll_year=0.5, year_len=250, cut_point_list=cut_point_list, output=True)
sharpe_quantile = cut_sharpe.values[0]
out_condition = sharpe_quantile > 0.8
return out_condition, round(sharpe_quantile * way, 4)
def filter_all(cut_date,
pos_df_daily,
pct_n,
if_return_pnl=False,
if_only_long=False):
if if_only_long:
pnl_df = (pos_df_daily[pos_df_daily > 0] * pct_n).sum(axis=1)
pnl_df = pnl_df.replace(np.nan, 0)
else:
pnl_df = (pos_df_daily * pct_n).sum(axis=1)
pnl_df = pnl_df.replace(np.nan, 0)
# pnl_df = pd.Series(pnl_df)
# 样本内表现
return_in = pct_n[pct_n.index < cut_date]
pnl_df_in = pnl_df[pnl_df.index < cut_date]
asset_df_in = pnl_df_in.cumsum()
last_asset_in = asset_df_in.iloc[-1]
pos_df_daily_in = pos_df_daily[pos_df_daily.index < cut_date]
pot_in = AZ_Pot(pos_df_daily_in, last_asset_in)
leve_ratio = AZ_Leverage_ratio(asset_df_in)
if leve_ratio < 0:
leve_ratio = 100
sharpe_q_in_df = bt.AZ_Rolling_sharpe(pnl_df_in,
roll_year=1,
year_len=250,
min_periods=1,
cut_point_list=[0.3, 0.5, 0.7],
output=False)
sp_in = bt.AZ_Sharpe_y(pnl_df_in)
fit_ratio = bt.AZ_fit_ratio(pos_df_daily_in, return_in)
ic = round(
bt.AZ_Normal_IC(pos_df_daily_in, pct_n, min_valids=None, lag=0).mean(),
6)
sharpe_q_in_df_u, sharpe_q_in_df_m, sharpe_q_in_df_d = sharpe_q_in_df.values
in_condition_u = sharpe_q_in_df_u > 0.9 and leve_ratio > 1
in_condition_d = sharpe_q_in_df_d < -0.9 and leve_ratio > 1
# 分双边和只做多
if if_only_long:
in_condition = in_condition_u
else:
in_condition = in_condition_u | in_condition_d
if sharpe_q_in_df_m > 0:
way = 1
else:
way = -1
# 样本外表现
pnl_df_out = pnl_df[pnl_df.index >= cut_date]
out_condition, sharpe_q_out = out_sample_perf_c(pnl_df_out, way=way)
if if_return_pnl:
return in_condition, out_condition, ic, sharpe_q_in_df_u, sharpe_q_in_df_m, sharpe_q_in_df_d, pot_in, \
fit_ratio, leve_ratio, sp_in, sharpe_q_out, pnl_df
else:
return in_condition, out_condition, ic, sharpe_q_in_df_u, sharpe_q_in_df_m, sharpe_q_in_df_d, pot_in, \
fit_ratio, leve_ratio, sp_in, sharpe_q_out
def filter_pot_sharpe(cut_date,
signal,
pct_n,
index_df,
lag=1,
hold_time=5,
if_hedge=True,
hedge_ratio=1,
if_return_pnl=False):
# signal向下移动一天,避免未来函数
signal = signal.shift(lag)
# # 将所有的0替换为nan,使得计算ic时更加合理, 计算pnl时没影响
# signal = signal.replace(0, np.nan)
pos_df_daily = pos_daily_fun(signal, n=hold_time)
if if_hedge:
hedge_df = hedge_ratio * index_df.mul(pos_df_daily.sum(axis=1), axis=0)
# pnl_df = (signal * pct_n).sum(axis=1).sub(hedge_df, axis=0)
pnl_df = -hedge_df.sub((pos_df_daily * pct_n).sum(axis=1), axis=0)
pnl_df = pnl_df[pnl_df.columns[0]]
else:
pnl_df = (pos_df_daily * pct_n).sum(axis=1)
# pnl_df = pd.Series(pnl_df)
# 样本内表现
pnl_df_in = pnl_df[pnl_df.index < cut_date]
asset_df_in = pnl_df_in.cumsum()
last_asset_in = asset_df_in.iloc[-1]
pos_df_daily_in = pos_df_daily[pos_df_daily.index < cut_date]
pot = AZ_Pot(pos_df_daily_in, last_asset_in)
sharpe_q_in_df = bt.AZ_Rolling_sharpe(pnl_df_in,
roll_year=1,
year_len=250,
min_periods=1,
cut_point_list=[0.3, 0.5, 0.7],
output=False)
sharpe_q_in_df_u, sharpe_q_in_df_m, sharpe_q_in_df_d = sharpe_q_in_df.values
in_condition_u = sharpe_q_in_df_u > 0.9
in_condition_d = sharpe_q_in_df_d < -0.9
in_condition = in_condition_u | in_condition_d
if sharpe_q_in_df_m > 0:
way = 1
else:
way = -1
# 样本外表现
pnl_df_out = pnl_df[pnl_df.index >= cut_date]
out_condition, sharpe_q_out = out_sample_perf(pnl_df_out, way=way)
if if_return_pnl:
return in_condition, out_condition, sharpe_q_in_df_u, sharpe_q_in_df_m, sharpe_q_in_df_d, pot, sharpe_q_out, \
pnl_df
else:
return in_condition, out_condition, sharpe_q_in_df_u, sharpe_q_in_df_m, sharpe_q_in_df_d, pot, sharpe_q_out
def filter_time_para_fun(time_para_dict, pos_df_daily, adj_return,
if_return_pnl, if_only_long):
if if_only_long:
pnl_df = (pos_df_daily[pos_df_daily > 0] * adj_return).sum(axis=1)
pnl_df = pnl_df.replace(np.nan, 0)
else:
pnl_df = (pos_df_daily * adj_return).sum(axis=1)
pnl_df = pnl_df.replace(np.nan, 0)
result_dict = OrderedDict()
for time_key in time_para_dict.keys():
begin_para, cut_para, end_para_1, end_para_2, end_para_3, end_para_4 = time_para_dict[
time_key]
# 样本内索引
sample_in_index = (adj_return.index >= begin_para) & (adj_return.index
< cut_para)
# 样本外索引
sample_out_index_1 = (adj_return.index >= cut_para) & (adj_return.index
< end_para_1)
sample_out_index_2 = (adj_return.index >= cut_para) & (adj_return.index
< end_para_2)
sample_out_index_3 = (adj_return.index >= cut_para) & (adj_return.index
< end_para_3)
sample_out_index_4 = (adj_return.index >= cut_para) & (adj_return.index
< end_para_4)
# 样本内表现
pos_df_daily_in = pos_df_daily[sample_in_index]
if len(pos_df_daily_in.abs().sum(axis=1).replace(
0, np.nan).dropna()) / len(pos_df_daily_in) < 0.1:
continue
adj_return_in = adj_return[sample_in_index]
pnl_df_in = pnl_df[sample_in_index]
asset_df_in = pnl_df_in.cumsum()
last_asset_in = asset_df_in.iloc[-1]
pot_in = AZ_Pot(pos_df_daily_in, last_asset_in)
leve_ratio = AZ_Leverage_ratio(asset_df_in)
if leve_ratio < 0:
leve_ratio = 100
sharpe_q_in_df = bt.AZ_Rolling_sharpe(pnl_df_in,
roll_year=1,
year_len=250,
min_periods=1,
cut_point_list=[0.3, 0.5, 0.7],
output=False)
sharpe_q_in_df = round(sharpe_q_in_df, 4)
sp_in = bt.AZ_Sharpe_y(pnl_df_in)
fit_ratio = bt.AZ_fit_ratio(pos_df_daily_in, adj_return_in)
ic = round(
bt.AZ_Normal_IC(pos_df_daily_in,
adj_return_in,
min_valids=None,
lag=0).mean(), 6)
sp_in_u, sp_in_m, sp_in_d = sharpe_q_in_df.values
in_condition_u = sp_in_u > 0.9 and leve_ratio > 1
in_condition_d = sp_in_d < -0.9 and leve_ratio > 1
# 分双边和只做多
if if_only_long:
in_condition = in_condition_u
else:
in_condition = in_condition_u | in_condition_d
if sp_in_m > 0:
way = 1
else:
way = -1
# 样本外表现
pnl_df_out_1 = pnl_df[sample_out_index_1]
pnl_df_out_2 = pnl_df[sample_out_index_2]
pnl_df_out_3 = pnl_df[sample_out_index_3]
pnl_df_out_4 = pnl_df[sample_out_index_4]
out_condition_1, sp_out_1 = out_sample_perf_c(pnl_df_out_1, way=way)
out_condition_2, sp_out_2 = out_sample_perf_c(pnl_df_out_2, way=way)
out_condition_3, sp_out_3 = out_sample_perf_c(pnl_df_out_3, way=way)
out_condition_4, sp_out_4 = out_sample_perf_c(pnl_df_out_4, way=way)
if if_return_pnl:
result_dict[time_key] = [
in_condition, out_condition_1, out_condition_2,
out_condition_3, out_condition_4, ic, sp_in_u, sp_in_m,
sp_in_d, pot_in, fit_ratio, leve_ratio, sp_in, sp_out_1,
sp_out_2, sp_out_3, sp_out_4, pnl_df
]
else:
result_dict[time_key] = [
in_condition, out_condition_1, out_condition_2,
out_condition_3, out_condition_4, ic, sp_in_u, sp_in_m,
sp_in_d, pot_in, fit_ratio, leve_ratio, sp_in, sp_out_1,
sp_out_2, sp_out_3, sp_out_4
]
return result_dict
