# """
# 因子描述性统计
factor_describe = {}
for fac in fac_data.keys():
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
# 保存因子的描述性统计
factor_describe[fac] = fac_data[fac].T.describe().T
# 打印因子的描述性统计均值
print(fac, fac_data[fac].T.describe().mean(axis=1))
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
# """
# 以下一日的开盘1小时tvwap到再下一日的开盘1小时tvwap收益率作为预测目标
begin = '2017-01-01'
end = '2021-03-02'
tvwap = fetch_data.fetch(begin, end, ['stock_twap_0930_1030']) # adjtvwap?
fac_data['next_re'] = uc.ts_delay(tvwap['stock_twap_0930_1030'],
-2) / uc.ts_delay(
tvwap['stock_twap_0930_1030'], -1) - 1
fac_data['next_re'] = fac_data['next_re'].dropna(how='all')
# 将每天的对应数据合并
new_f = {}
for k, v in fac_data.items():
new_v = pd.DataFrame(v.stack())
new_v.columns = [k]
new_f[k] = new_v
new_f = pd.concat(new_f.values(), axis=1)
f = open(data_pat + '/fac_reshape.pkl', 'wb') # 记得修改
pickle.dump(new_f, f, -1)
f.close()
# new_f.reset_index().to_csv(data_pat + '/fac_reshape.csv',index=False,encoding='gbk')
mine_summary = query_data.get_alphafactors_info(user='******')
# 调整正负
factor_value_adj = {}
for summa in mine_summary:
if summa['factor_name'] in list(factor_value.keys()):
if 'IC' in list(summa['perf']['1_d'].keys()):
factor_value_adj[summa['factor_name']] = factor_value[
summa['factor_name']] * uc.sign(summa['perf']['1_d']['IC'])
else:
factor_value_adj[summa['factor_name']] = factor_value[
summa['factor_name']] * uc.sign(
summa['perf']['1_d']['ic-mean'])
# 建立股票在未来n日的涨跌标签
oc_data = fetch_data.fetch(begin, end, ['stock_adjopen', 'stock_adjclose'])
ud_tag = uc.ts_delay(oc_data['stock_adjclose'], -pred_window) / uc.ts_delay(
oc_data['stock_adjopen'], -1) - 1 # 以第二日的开盘价买入
ud_tag = ud_tag.mask(ud_tag > 0, 1)
ud_tag = ud_tag.mask(ud_tag < 0, 0)
# 股票因子值的reshape
new_f = {}
for k, v in factor_value_adj.items():
new_v = pd.DataFrame(v.stack())
new_v.columns = [k]
new_f[k] = new_v
new_f = pd.concat(new_f.values(), axis=1)
# 滚动生成上涨概率预测
prediction = {}
bay = GaussianNB()
for date in trade_days:
sub_data = new_f.loc[date, ]
model = sm.OLS(sub_data.iloc[:, -1],
sm.add_constant(sub_data.iloc[:, 0:-1]),
missing='drop').fit()
coef[date] = model.params
R_sq[date] = model.rsquared_adj
print(date)
coef_param = pd.concat(coef.values(), axis=1, keys=coef.keys())
coef_param = pd.DataFrame(coef_param.values.T,
index=coef_param.columns,
columns=coef_param.index) # 转置
r2_param = pd.DataFrame(R_sq.values(),
index=R_sq.keys(),
columns=['R_square_adj'])
coef_param = uc.ts_delay(coef_param, 2) # 2天后才能用估计出的参数
r2_param = uc.ts_delay(r2_param, 2) # 2天后才能用估计出的参数
coef_param = coef_param.groupby(
coef_param.index.strftime('%Y-%m')).mean() # 每个月更新一次权重
r2_param = r2_param.groupby(
r2_param.index.strftime('%Y-%m')).mean() # 每个月更新一次权重
plt.figure()
plt.plot(r2_param.index, r2_param['R_square_adj'])
plt.show()
coef_param.to_csv(data_pat + '/linear_regress_m/coef_param.csv',
encoding='gbk')
r2_param.to_csv(data_pat + '/linear_regress_m/r2_param.csv', encoding='gbk')
# 画出因子暴露时间序列
le = np.size(coef_param, 0)
la = math.ceil(4 * (le / 100)**(2 / 9))
la = math.ceil(4*(le/100)**(2/9))
for coef_name in coef_param.columns:
plt.figure()
plt.plot(coef_param.index, coef_param[coef_name])
plt.plot(coef_param.index, coef_param[coef_name].rolling(20).mean())
plt.title(coef_name, fontproperties="SimSun")
plt.show()
model = sm.OLS(coef_param[coef_name], [1 for i in range(le)]).fit(cov_type='HAC', cov_kwds={'maxlags': la})
print(model.summary()) # 有些因子的系数显著为负?多因子回归的影响
# 求收益率预测值(只用最近一次截面回归得到的系数)
fac = {}
new_f['const'] = 1
new_f = new_f.drop(['next_re'], axis=1)
coef_param2 = pd.concat([new_f.reset_index(level=1).iloc[:, 0], uc.ts_delay(coef_param, 2)], axis=1) # 2天后才能用估计出的参数
coef_param2 = coef_param2.set_index([coef_param2.index, 'level_1'])
pred = (coef_param2 * new_f).sum(axis=1, min_count=2) # 至少包含一个变量和一个const
pred = pred.unstack()
pred = pred.dropna(how='all')
fac['nearest'] = pred
f = open(data_pat + '/linear_regress_7/nearest/fac.pkl', 'wb') # 记得修改
pickle.dump(fac, f, -1)
f.close()
# 求收益率预测值(用过去20日截面回归得到的系数的平均值)
fac = {}
coef_param3 = pd.concat([new_f.reset_index(level=1).iloc[:, 0], uc.ts_delay(coef_param.rolling(20).mean(), 2)], axis=1) # 2天后才能用估计出的参数
coef_param3 = coef_param3.set_index([coef_param3.index, 'level_1'])
pred2 = (coef_param3 * new_f).sum(axis=1, min_count=2) # 至少包含一个变量和一个const
plt.plot(coef_param.index, coef_param[coef_name])
plt.plot(coef_param.index, coef_param[coef_name].rolling(20).mean())
plt.title(coef_name, fontproperties="SimSun")
plt.show()
model = sm.OLS(coef_param[coef_name],
[1 for i in range(le)]).fit(cov_type='HAC',
cov_kwds={'maxlags': la})
print(model.summary()) # 有些因子的系数显著为负?多因子回归的影响
# 求收益率预测值(只用最近一次截面回归得到的系数)
fac = {}
new_f['const'] = 1
new_f = new_f.drop(['next_re'], axis=1)
coef_param2 = pd.concat(
[new_f.reset_index(level=1).iloc[:, 0],
uc.ts_delay(coef_param, 2)],
axis=1) # 2天后才能用估计出的参数
coef_param2 = coef_param2.set_index([coef_param2.index, 'level_1'])
pred = (coef_param2 * new_f).sum(axis=1, min_count=2) # 至少包含一个变量和一个const
pred = pred.unstack()
pred = pred.dropna(how='all')
fac['nearest'] = pred
f = open(data_pat + '/linear_regress_22/nearest/fac.pkl', 'wb') # 记得修改
pickle.dump(fac, f, -1)
f.close()
# 求收益率预测值(用过去20日截面回归得到的系数的平均值)
fac = {}
coef_param3 = pd.concat([
new_f.reset_index(level=1).iloc[:, 0],
uc.ts_delay(coef_param.rolling(20).mean(), 2)
from copy import deepcopy
import numpy as np
import time
import json
from collections import Counter
data_pat = 'E:/FT_Users/LihaiYang/Files/factor_comb_data/fac_meaning/5group' # 记得修改
# 计算未来1、3、5、10、20日收益率，以开盘1小时tvwap为标准
begin = '2015-01-01' # 记得修改
end = '2020-02-28'
end1 = '2019-12-31'
data = fetch_data.fetch(begin, end, ['stock_adjtwap_0930_1030'])
index_data = fetch_data.fetch(begin, end, ['index_close'], '000905')
stock_re = {}
stock_re['1_d'] = uc.ts_delay(data['stock_adjtwap_0930_1030'],
-2) / uc.ts_delay(
data['stock_adjtwap_0930_1030'], -1) - 1
stock_re['3_d'] = uc.ts_delay(data['stock_adjtwap_0930_1030'],
-4) / uc.ts_delay(
data['stock_adjtwap_0930_1030'], -1) - 1
stock_re['5_d'] = uc.ts_delay(data['stock_adjtwap_0930_1030'],
-6) / uc.ts_delay(
data['stock_adjtwap_0930_1030'], -1) - 1
stock_re['10_d'] = uc.ts_delay(data['stock_adjtwap_0930_1030'],
-11) / uc.ts_delay(
data['stock_adjtwap_0930_1030'], -1) - 1
stock_re['20_d'] = uc.ts_delay(data['stock_adjtwap_0930_1030'],
-21) / uc.ts_delay(
data['stock_adjtwap_0930_1030'], -1) - 1
trade_days = query_data.get_trade_days('d',
from_trade_day=begin,
plt.figure()
plt.plot(r2_param.index, r2_param['R_square_adj'])
plt.plot(r2_param.index, r2_param['R_square_adj'].rolling(20).mean())
plt.show()
coef_param.to_csv(data_pat + '/ols/coef_param.csv', encoding='gbk')
r2_param.to_csv(data_pat + '/ols/r2_param.csv', encoding='gbk')
# 求收益率预测值
new_f['const'] = 1
new_f = new_f.drop(['stock_rela'], axis=1)
fac = {}
# 只用最近一次截面回归得到的系数
coef_param2 = pd.concat(
[new_f.reset_index(level=1).iloc[:, 0],
uc.ts_delay(coef_param, 11)],
axis=1) # 11天后才能用估计出的参数，记得修改
coef_param2 = coef_param2.set_index([coef_param2.index, 'level_1'])
pred = (coef_param2 * new_f).sum(axis=1, min_count=1)
pred = pred.unstack()
pred = pred.dropna(how='all')
fac['nearest'] = pred
# 用过去20日截面回归得到的系数的平均值
coef_param3 = pd.concat([
new_f.reset_index(level=1).iloc[:, 0],
uc.ts_delay(coef_param.rolling(20).mean(), 11)
],
axis=1) # 11天后才能用估计出的参数，记得修改
coef_param3 = coef_param3.set_index([coef_param3.index, 'level_1'])
pred2 = (coef_param3 * new_f).sum(axis=1, min_count=1)