def forcastStkReturn(self):
'''
给出下一期股票收益的预测
:return:dataframe,纵轴:股票，横轴:e_ret,各因子暴露收益
'''
nextDate = fapi.TradingTimePoints(
asset_code=None,
starttime=self.timespan[1] + timedelta(days=1),
endtime=self.timespan[1] + timedelta(days=10),
freq=self.freq).index[0].to_pydatetime()
FactorMatDict = {
x.descriptorRefName(): get_feature(descriptor=x,
starttime=nextDate,
endtime=nextDate,
freq=self.freq,
check=False)
for x in self.factors
}
matchDs = xr.Dataset(FactorMatDict)
CalibratedFeats = {
fename: matchDs[fename].values.flatten()
for fename in matchDs
}
stocks = matchDs.indexes['StockCode']
df_factorLoading = pd.DataFrame(index=stocks, data=CalibratedFeats)
expFreturnLi = [
self.expFreturn[f.descriptorRefName()] for f in self.factors
]
df_ForcastReturns = df_factorLoading.mul(expFreturnLi, axis=1)
df_ForcastReturns['e_ret'] = df_ForcastReturns.sum(axis=1)
return df_ForcastReturns
def __init__(self,
asset_code: str,
factors: List[Union[Descriptor]],
starttime: datetime,
endtime: datetime,
freq: str = 'd',
forcast_period: int = 1):
'''
:param stockcode: 待预测标的
:param factors: 特征列表
:param starttime: 样本起始时间
:param endtime: 样本结束时间
:param freq: 学习频率 'm'(分)/'d'(日)/'w'(周)/'M'(月)
:param forcast_period: 预测期
'''
#todo 如何预测同标的的期货合约(要考虑合约换月)？
# (丢给feature算法/get_return特殊处理，此处仅考虑可以由stockcode顺利获取特征及收益)
ReturnSeries = fapi.getReturn(asset_code=asset_code,
starttime=starttime,
endtime=endtime,
forcast_period=forcast_period,
freq=freq)
for factor in factors:
factor.freq = freq
feSeries = get_feature(factor, starttime, endtime, asset_code,
True)
def forcastStkReturn(self, check=True) -> pd.DataFrame:
'''
给出下一期股票收益的预测
:param check: bool 是否检查更新特征数据
:return:dataframe,纵轴:股票，横轴:e_ret,各因子暴露收益
'''
nextDate = fapi.TradingTimePoints(
asset_code=None,
starttime=self.timespan[1] + timedelta(days=1),
endtime=self.timespan[1] + timedelta(days=10),
freq=self.freq).index[0].to_pydatetime()
FactorMatDict = {
x.descriptorRefName(): get_feature(descriptor=x,
starttime=nextDate,
endtime=nextDate,
freq=self.freq,
checkLevel=2 if check else 0)
for x in self.factors
}
matchDs = xr.Dataset(FactorMatDict)
CalibratedFeats = {
fename: matchDs[fename].values.flatten()
for fename in matchDs
}
stocks = matchDs.indexes['StockCode']
df_factorLoading = pd.DataFrame(index=stocks, data=CalibratedFeats)
expFreturnLi = [
self.expFreturn[f.descriptorRefName()] for f in self.factors
]
df_ForcastReturns: pd.DataFrame = df_factorLoading.mul(expFreturnLi,
axis=1)
df_ForcastReturns['e_ret'] = df_ForcastReturns.sum(axis=1)
cashdict = {i: 0 for i in list(df_ForcastReturns.columns)}
cash = pd.DataFrame(cashdict, index=['cash'])
aa = df_ForcastReturns.append(cash)
return aa
def __getfundamental(self, timespan):
'''
计算并返回指定时间段内的交易日、描述子数据、收益数据以及横截面回归结果.
:param timespan: 指定的时间段
:return: (out_dateIndex,out_endogs,out_exogs,out_dateIndex)
out_dateIndex:时间段内的交易时点序列
out_endogs:时间段内的股票收益数据
out_exogs：时间段内的描述子数据
out_rst:时间段内的横截面回归结果
'''
ReturnMat = fapi.getReturn(asset_code=None,
starttime=timespan[0],
endtime=timespan[1],
forcast_period=self.forcast_period,
freq=self.freq,
asset_type='stock')
#取因子矩阵
FactorMatDict = {
'feat_' + x.descriptorRefName(): get_feature(descriptor=x,
starttime=timespan[0],
endtime=timespan[1],
freq=self.freq,
check=False)
for x in self.factors
}
ControlMatDict = {
'ctrl_' + x.descriptorRefName(): get_feature(descriptor=x,
starttime=timespan[0],
endtime=timespan[1],
freq=self.freq,
check=False)
for x in self.controls
}
#FMB回归
matchDict = {'return': ReturnMat}
matchDict.update(FactorMatDict)
matchDict.update(ControlMatDict)
matchDs = xr.Dataset(matchDict)
def ProdTrue(*args):
for obj in args:
if (args is True):
continue
else:
return False
return True
CalibratedCtrls = [
matchDs[fe_name].values for fe_name in matchDs
if fe_name[0:4] == 'ctrl'
]
CalibratedFeats = [
matchDs[fe_name].values for fe_name in matchDs
if fe_name[0:4] == 'feat'
]
if (CalibratedCtrls.__len__() > 0):
ArrayProdTrue = np.frompyfunc(ProdTrue, CalibratedCtrls.__len__(),
1)
AssembledControlMat = ArrayProdTrue(
*CalibratedCtrls).astype('bool')
MaskedReturn = np.ma.MaskedArray(matchDs['return'].values,
mask=~AssembledControlMat,
fill_value=np.nan)
else:
MaskedReturn = matchDs['return'].values
l = MaskedReturn.shape[0]
out_rst = []
exogs = np.stack(CalibratedFeats)
noneCount = 0
for i in tqdm(range(l), desc='正在截面回归...'):
try:
out_rst.append(
sm.OLS(endog=MaskedReturn[i],
exog=exogs[:, i, :].T,
missing='drop').fit())
except:
out_rst.append(None)
noneCount += 1
if (noneCount == l):
raise Exception("所选时间区间内无有效回归样本，请检查基础数据！")
out_dateIndex = matchDs['return'].indexes['Time']
return (out_dateIndex, out_rst)
'''
择时模型:
根据选择的资产、特征，在样本时间区间内训练择时模型，并进行样本外回测.
输入:资产代码、特征对象列表、样本时间区间；
输出:一个可给出回测报告(输入时间区间即可)的对象。
对其它api的假定:
get_feature(descriptor:Descriptor,starttime:datetime, endtime:datetime,stock_name:list=None,check:bool=True)
descriptor的频率交由get_feature来指定，并在对象初始化中被剔除。
todo:现在回头补写一些基础数据和因子特征算法，用multifactor模型研究一下因子收益的持续性；
加入残差动量因子:将因子横街面回归残差
'''
from typing import List, Union
from datetime import datetime
from core.features.Descriptor import Descriptor
from core.fundamentals.getfundamentals import fundamentalApi as fapi
from core.features.getfeature import get_feature
class TSmodel:
def __init__(self,
asset_code: str,
factors: List[Union[Descriptor]],
starttime: datetime,
endtime: datetime,
freq: str = 'd',
forcast_period: int = 1):
'''
:param stockcode: 待预测标的