之前写过一篇多因子选股代码框架,本文对部分模块进行了修改,并添加了一些新模块。
多因子策略主要包括以下几个部分:
1.数据获取及处理。借助机器学习及深度学习工具,获取的因子越多越好,一般数据越大越有优势。缺失数据处理本文采用行业均值填充。
2.因子选择。因子选择的评判标准通常有信息系数IC’包含稳定性的IR(IC除以IC的标准差)、回归系数,但这些方法都有一个问题,就是忽略了因子之间的相关性。本文添加了机器学习的特征选择方法,用于进行因子有效性评判,在FeatureSelection类中identify_collinear方法可以先去除相关性高的特征,identify_importance_lgbm方法使用lightgbm算法进行特征选择,此方法本身能有效避免因子间的影响,embedded_select方法可以使用岭回归或者lasso回归,去除因子间共线性。以上方法都可以作为新的因子选择的标准。
3.基于因子有效性进行选股或择时并回测。基于前面的因子,形成有效策略,本文代码中简单添加了基于机器学习进行探索的代码,有兴趣的小伙伴可以发挥想象力,探索有效策略,具体方法本文不做分享。
另外,文中获取数据量相对较大,部分算法需要时间较长,不建议在聚宽平台上跑,本文包括了在自己平台上跑的全部代码,建议自己复制到本地跑数据。
import pandas as pd
import numpy as np
import time
import datetime
import statsmodels.api as sm
import pickle
import warnings
from jqdata import *
warnings.filterwarnings('ignore')
#此代码在自己机器上运行时使用,在聚宽平台上跑数据不需要
#from jqdatasdk import *
#auth('用户名','密码')
start_date = '2013-01-01'
end_date = '2014-01-01'
all_trade_days = (get_trade_days(start_date=start_date,end_date=end_date)).tolist() #所有交易日
trade_days = all_trade_days[::20] #每隔20天取一次数据,基本面数据更新频率较慢,数据获取频率尽量与之对应
securities = get_all_securities()
start_data_dt = datetime.datetime.strptime(start_date,'%Y-%m-%d').date()
securities_after_start_date = securities[(securities['start_date']<start_data_dt)] #选择起始时间之前上市的股票
all_stocks = list(securities_after_start_date.index)
INDUSTRY_NAME = 'sw_l1'
ttm_factors = []
'''
基本面因子映射
'''
fac_dict = {
'MC':valuation.market_cap, # 总市值
'GP':indicator.gross_profit_margin * income.operating_revenue, # 毛利润
'OP':income.operating_profit,
'OR':income.operating_revenue, # 营业收入
'NP':income.net_profit, # 净利润
'EV':valuation.market_cap + balance.shortterm_loan+balance.non_current_liability_in_one_year+balance.longterm_loan+balance.bonds_payable+balance.longterm_account_payable - cash_flow.cash_and_equivalents_at_end,
'TOE':balance.total_owner_equities, # 股东权益合计(元)
'TOR':income.total_operating_revenue, # 营业总收入
'EBIT':income.net_profit+income.financial_expense+income.income_tax_expense,
'TOC':income.total_operating_cost,#营业总成本
'NOCF/MC':cash_flow.net_operate_cash_flow / valuation.market_cap, #经营活动产生的现金流量净额/总市值
'OTR':indicator.ocf_to_revenue, #经营活动产生的现金流量净额/营业收入(%)
'GPOA':indicator.gross_profit_margin * income.operating_revenue / balance.total_assets, #毛利润 / 总资产 = 毛利率*营业收入 / 总资产
'GPM':indicator.gross_profit_margin, # 毛利率
'OPM':income.operating_profit / income.operating_revenue, #营业利润率
'NPM':indicator.net_profit_margin, # 净利率
'ROA':indicator.roa, # ROA
'ROE':indicator.roe, # ROE
'INC':indicator.inc_return, # 净资产收益率(扣除非经常损益)(%)
'EPS':indicator.eps, # 净资产收益率(扣除非经常损益)(%)
'AP':indicator.adjusted_profit, # 扣除非经常损益后的净利润(元)
'OP':indicator.operating_profit, # 经营活动净收益(元)
'VCP':indicator.value_change_profit, # 价值变动净收益(元) = 公允价值变动净收益+投资净收益+汇兑净收益
'ETTR':indicator.expense_to_total_revenue, # 营业总成本/营业总收入(%)
'OPTTR':indicator.operation_profit_to_total_revenue, # 营业利润/营业总收入(%)
'NPTTR':indicator.net_profit_to_total_revenue, # 净利润/营业总收入(%)
'OETTR':indicator.operating_expense_to_total_revenue, # 营业费用/营业总收入
'GETTR':indicator.ga_expense_to_total_revenue, # 管理费用/营业总收入(%)
'FETTR':indicator.financing_expense_to_total_revenue, # 财务费用/营业总收入(%)
'OPTP':indicator.operating_profit_to_profit, # 经营活动净收益/利润总额(%)
'IPTP':indicator.invesment_profit_to_profit, # 价值变动净收益/利润总额(%)
'GSASTR':indicator.goods_sale_and_service_to_revenue, # 销售商品提供劳务收到的现金/营业收入(%)
'OTR':indicator.ocf_to_revenue, # 经营活动产生的现金流量净额/营业收入(%)
'OTOP':indicator.ocf_to_operating_profit, # 经营活动产生的现金流量净额/经营活动净收益(%)
'ITRYOY':indicator.inc_total_revenue_year_on_year, # 营业总收入同比增长率(%)
'ITRA':indicator.inc_total_revenue_annual, # 营业总收入环比增长率(%)
'IRYOY':indicator.inc_revenue_year_on_year, # 营业收入同比增长率(%)
'IRA':indicator.inc_revenue_annual, # 营业收入环比增长率(%)
'IOPYOY':indicator.inc_operation_profit_year_on_year, # 营业利润同比增长率(%)
'IOPA':indicator.inc_operation_profit_annual, # 营业利润环比增长率(%)
'INPYOY':indicator.inc_net_profit_year_on_year, # 净利润同比增长率(%)
'INPA':indicator.inc_net_profit_annual, # 净利润环比增长率(%)
'INPTSYOY':indicator.inc_net_profit_to_shareholders_year_on_year, # 归属母公司股东的净利润同比增长率(%)
'INPTSA':indicator.inc_net_profit_to_shareholders_annual, # 归属母公司股东的净利润环比增长率(%)
'INPTSA':indicator.inc_net_profit_to_shareholders_annual, # 归属母公司股东的净利润环比增长率(%)
'ROIC':(income.net_profit+income.financial_expense+income.income_tax_expense)/(balance.total_owner_equities+balance.shortterm_loan+balance.non_current_liability_in_one_year+balance.longterm_loan+balance.bonds_payable+balance.longterm_account_payable),
'OPTT':income.operating_profit / income.total_profit, # 营业利润占比
'TP/TOR':income.total_profit / income.total_operating_revenue, #利润总额/营业总收入
'OP/TOR':income.operating_profit / income.total_operating_revenue,
'NP/TOR':income.net_profit / income.total_operating_revenue,
'NP':income.net_profit, # 净利润
'TA':balance.total_assets, # 总资产
'DER':balance.total_liability / balance.equities_parent_company_owners, # 产权比率 = 负债合计/归属母公司所有者权益合计
'FCFF/TNCL':(cash_flow.net_operate_cash_flow - cash_flow.net_invest_cash_flow) / balance.total_non_current_liability, #自由现金流比非流动负债
'NOCF/TL': cash_flow.net_operate_cash_flow / balance.total_liability, # 经营活动产生的现金流量净额/负债合计
'TCA/TCL':balance.total_current_assets / balance.total_current_liability, # 流动比率
'PE':valuation.pe_ratio, # PE 市盈率
'PB':valuation.pb_ratio, # PB 市净率
'PR':valuation.pcf_ratio, # PR 市现率
'PS':valuation.ps_ratio, # PS 市销率
'TOR/TA':income.total_operating_revenue / balance.total_assets, #总资产周转率
'TOR/FA':income.total_operating_revenue / balance.fixed_assets, #固定资产周转率
'TOR/TCA':income.total_operating_revenue / balance.total_current_assets, #流动资产周转率
'LTL/OC':balance.longterm_loan / income.operating_cost, #长期借款/营业成本
'TL/TA':balance.total_liability / balance.total_assets, #总资产/总负债
'TL/TOE':balance.total_liability / balance.total_owner_equities,#负债权益比
}
adjust_factors = {
'TOR/TA':income.total_operating_revenue / balance.total_assets, #总资产周转率
'TOR/FA':income.total_operating_revenue / balance.fixed_assets, #固定资产周转率
'TOR/TCA':income.total_operating_revenue / balance.total_current_assets, #流动资产周转率
'LTL/OC':balance.longterm_loan / income.operating_cost, #长期借款/营业成本
'TL/TA':balance.total_liability / balance.total_assets, #总资产/总负债
'TL/TOE':balance.total_liability / balance.total_owner_equities,#负债权益比
'DER':balance.total_liability / balance.equities_parent_company_owners, # 产权比率 = 负债合计/归属母公司所有者权益合计
'FCFF/TNCL':(cash_flow.net_operate_cash_flow - cash_flow.net_invest_cash_flow) / balance.total_non_current_liability, #自由现金流比非流动负债
'NOCF/TL': cash_flow.net_operate_cash_flow / balance.total_liability, # 经营活动产生的现金流量净额/负债合计
'TCA/TCL':balance.total_current_assets / balance.total_current_liability, # 流动比率
'ROIC':(income.net_profit+income.financial_expense+income.income_tax_expense)/(balance.total_owner_equities+balance.shortterm_loan+balance.non_current_liability_in_one_year+balance.longterm_loan+balance.bonds_payable+balance.longterm_account_payable),
'OPTT':income.operating_profit / income.total_profit, # 营业利润占比
'TP/TOR':income.total_profit / income.total_operating_revenue, #利润总额/营业总收入
'OP/TOR':income.operating_profit / income.total_operating_revenue,
'NP/TOR':income.net_profit / income.total_operating_revenue,
'NOCF/MC':cash_flow.net_operate_cash_flow / valuation.market_cap, #经营活动产生的现金流量净额/总市值
'GPOA':indicator.gross_profit_margin * income.operating_revenue / balance.total_assets, #毛利润 / 总资产 = 毛利率*营业收入 / 总资产
'OPM':income.operating_profit / income.operating_revenue, #营业利润率
'EBIT':income.net_profit+income.financial_expense+income.income_tax_expense,
}
#获取所有因子列表
factor_list = list(fac_dict.keys())
def get_fundamental_data(securities,factor_list,ttm_factors, date):
'''
获取基本面数据,横截面数据,时间、股票、因子三个参数确定
获取的数据中含有Nan值,一般用行业均值填充
输入:
factor_list:list, 普通因子
ttm_factors:list, ttm因子,获取过去四个季度财报数据的和
date:str 或者 datetime.data, 获取数据的时间
securities:list,查询的股票
输出:
DataFrame,普通因子和ttm因子的合并,index为股票代码,values为因子值
'''
if len(factor_list) == 0:
return 'factors list is empty, please input data'
#获取查询的factor list
q = query(valuation.code)
for fac in factor_list:
q = q.add_column(fac_dict[fac])
q = q.filter(valuation.code.in_(securities))
fundamental_df = get_fundamentals(q,date)
fundamental_df.index = fundamental_df['code']
fundamental_df.columns = ['code'] + factor_list
if type(date) == str:
year = int(date[:4])
month_day = date[5:]
elif type(date) == datetime.date:
date = date.strftime('%Y-%m-%d')
year = int(date[:4])
month_day = date[5:]
else:
return 'input date error'
if month_day < '05-01':
statdate_list = [str(year-2)+'q4', str(year-1)+'q1', str(year-1)+'q2', str(year-1)+'q3']
elif month_day >= '05-01' and month_day < '09-01':
statdate_list = [str(year-1)+'q1', str(year-1)+'q2', str(year-1)+'q3',str(year)+'q1']
elif month_day >= '09-01' and month_day < '11-01':
statdate_list = [str(year-1)+'q2', str(year-1)+'q3', str(year)+'q1', str(year)+'q2']
elif month_day >= '11-01':
statdate_list = [str(year-1)+'q4', str(year)+'q1', str(year)+'q2', str(year)+'q3']
ttm_fundamental_data = ''
ttm_q = query(valuation.code)
for fac in ttm_factors:
ttm_q = ttm_q.add_column(fac_dict[fac])
ttm_q = ttm_q.filter(valuation.code.in_(securities))
for statdate in statdate_list:
if type(ttm_fundamental_data) == str:
fundamental_data = get_fundamentals(ttm_q, statDate=statdate)
fundamental_data.index = fundamental_data['code']
ttm_fundamental_data = fundamental_data
else:
fundamental_data = get_fundamentals(ttm_q, statDate=statdate)
fundamental_data.index = fundamental_data['code']
ttm_fundamental_data.iloc[:,1:] += fundamental_data.iloc[:,1:]
ttm_fundamental_data.columns = ['code'] + ttm_factors
results = pd.merge(fundamental_df,ttm_fundamental_data,on=['code'],how='inner')
results = results.sort_values(by='code')
results.index = results['code']
results = results.drop(['code'],axis=1)
#删除非数值列
columns = list(results.columns)
for column in columns:
if not(isinstance(results[column][0],int) or isinstance(results[column][0],float)):
results = results.drop([column],axis=1)
return results
def get_all_fundamentals(securities, date):
'''
获取所有基本面因子
输入:
securies:list,查询的股票代码
date:str or datetime,查询的时间
输出:
fundamentals:dataframe,index为股票代码,values为因子值
'''
q = query(valuation,balance,cash_flow,income,indicator).filter(valuation.code.in_(securities))
fundamentals = get_fundamentals(q,date)
fundamentals.index = fundamentals['code']
#删除非数值列
columns = list(fundamentals.columns)
for column in columns:
if not(isinstance(fundamentals[column][0],int) or isinstance(fundamentals[column][0],float)):
fundamentals = fundamentals.drop([column],axis=1)
fundamentals = fundamentals.sort_index()
return fundamentals
all_fundamentals = get_all_fundamentals(all_stocks,start_date)
def get_stock_industry(industry_name,date,output_csv = False):
'''
获取股票对应的行业
input:
industry_name: str,
"sw_l1": 申万一级行业
"sw_l2": 申万二级行业
"sw_l3": 申万三级行业
"jq_l1": 聚宽一级行业
"jq_l2": 聚宽二级行业
"zjw": 证监会行业
date:时间
output: DataFrame,index 为股票代码,columns 为所属行业代码
'''
industries = list(get_industries(industry_name).index)
all_securities = get_all_securities(date=date) #获取当天所有股票代码
all_securities['industry_code'] = 1
for ind in industries:
industry_stocks = get_industry_stocks(ind,date)
#有的行业股票不在all_stocks列表之中
industry_stocks = set(all_securities) & set(industry_stocks)
all_securities['industry_code'][industry_stocks] = ind
stock_industry = all_securities['industry_code'].to_frame()
if output_csv == True:
stock_industry.to_csv('stock_industry.csv') #输出csv文件,股票对应行业
return stock_industry
def fillna_with_industry(data,date,industry_name='sw_l1'):
'''
使用行业均值填充nan值
input:
data:DataFrame,输入数据,index为股票代码
date:string,时间必须和data数值对应时间一致
output:
DataFrame,缺失值用行业中值填充,无行业数据的用列均值填充
'''
stocks = list(data.index)
stocks_industry = get_stock_industry(industry_name,date)
stocks_industry_merge = data.merge(stocks_industry, left_index=True,right_index=True,how='left')
stocks_dropna = stocks_industry_merge.dropna()
columns = list(data.columns)
select_data = []
group_data = stocks_industry_merge.groupby('industry_code')
group_data_mean = group_data.mean()
group_data = stocks_industry_merge.merge(group_data_mean,left_on='industry_code',right_index=True,how='left')
for column in columns:
if type(data[column][0]) != str:
group_data[column+'_x'][pd.isnull(group_data[column+'_x'])] = group_data[column+'_y'][pd.isnull(group_data[column+'_x'])]
group_data[column] = group_data[column+'_x']
#print(group_data.head())
select_data.append(group_data[column])
result = pd.concat(select_data,axis=1)
#行业均值为Nan,用总体均值填充
mean = result.mean()
for i in result.columns:
result[i].fillna(mean[i],inplace=True)
return result
#获取日期列表
def get_tradeday_list(start,end,frequency=None,count=None):
'''
input:
start:str or datetime,起始时间,与count二选一
end:str or datetime,终止时间
frequency:
str: day,month,quarter,halfyear,默认为day
int:间隔天数
count:int,与start二选一,默认使用start
'''
if isinstance(frequency,int):
all_trade_days = get_trade_days(start,end)
trade_days = all_trade_days[::frequency]
days = [datetime.datetime.strftime(i,'%Y-%m-%d') for i in trade_days]
return days
if count != None:
df = get_price('000001.XSHG',end_date=end,count=count)
else:
df = get_price('000001.XSHG',start_date=start,end_date=end)
if frequency == None or frequency =='day':
days = df.index
else:
df['year-month'] = [str(i)[0:7] for i in df.index]
if frequency == 'month':
days = df.drop_duplicates('year-month').index
elif frequency == 'quarter':
df['month'] = [str(i)[5:7] for i in df.index]
df = df[(df['month']=='01') | (df['month']=='04') | (df['month']=='07') | (df['month']=='10') ]
days = df.drop_duplicates('year-month').index
elif frequency =='halfyear':
df['month'] = [str(i)[5:7] for i in df.index]
df = df[(df['month']=='01') | (df['month']=='06')]
days = df.drop_duplicates('year-month').index
trade_days = [datetime.datetime.strftime(i,'%Y-%m-%d') for i in days]
return trade_days
tl = get_tradeday_list(start_date,end_date,frequency='month')
def get_date_list(begin_date, end_date):
'''
得到datetime类型时间序列
'''
dates = []
dt = datetime.datetime.strptime(begin_date,"%Y-%m-%d")
date = begin_date[:]
while date <= end_date:
dates.append(date)
dt += datetime.timedelta(days=1)
date = dt.strftime("%Y-%m-%d")
return dates
#去极值函数
#mad中位数去极值法
def filter_extreme_MAD(series,n): #MAD: 中位数去极值
median = series.quantile(0.5)
new_median = ((series - median).abs()).quantile(0.50)
max_range = median + n*new_median
min_range = median - n*new_median
return np.clip(series,min_range,max_range)
#进行标准化处理
def winsorize(factor, std=3, have_negative = True):
'''
去极值函数
factor:以股票code为index,因子值为value的Series
std为几倍的标准差,have_negative 为布尔值,是否包括负值
输出Series
'''
r=factor.dropna().copy()
if have_negative == False:
r = r[r>=0]
else:
pass
#取极值
edge_up = r.mean()+std*r.std()
edge_low = r.mean()-std*r.std()
r[r>edge_up] = edge_up
r[r<edge_low] = edge_low
return r
#标准化函数:
def standardize(s,ty=2):
'''
s为Series数据
ty为标准化类型:1 MinMax,2 Standard,3 maxabs
'''
data=s.dropna().copy()
if int(ty)==1:
re = (data - data.min())/(data.max() - data.min())
elif ty==2:
re = (data - data.mean())/data.std()
elif ty==3:
re = data/10**np.ceil(np.log10(data.abs().max()))
return re
#数据去极值及标准化
def winsorize_and_standarlize(data,qrange=[0.05,0.95],axis=0):
'''
input:
data:Dataframe or series,输入数据
qrange:list,list[0]下分位数,list[1],上分位数,极值用分位数代替
'''
if isinstance(data,pd.DataFrame):
if axis == 0:
q_down = data.quantile(qrange[0])
q_up = data.quantile(qrange[1])
index = data.index
col = data.columns
for n in col:
data[n][data[n] > q_up[n]] = q_up[n]
data[n][data[n] < q_down[n]] = q_down[n]
data = (data - data.mean())/data.std()
data = data.fillna(0)
else:
data = data.stack()
data = data.unstack(0)
q = data.quantile(qrange)
index = data.index
col = data.columns
for n in col:
data[n][data[n] > q[n]] = q[n]
data = (data - data.mean())/data.std()
data = data.stack().unstack(0)
data = data.fillna(0)
elif isinstance(data,pd.Series):
name = data.name
q = data.quantile(qrange)
data = np.clip(data,q.values[0],q.values[1])
data = (data - data.mean())/data.std()
return data
def neutralize(data,date,market_cap,industry_name='sw_l1'):
'''
中性化,使用行业和市值因子中性化
input:
data:DataFrame,index为股票代码,columns为因子,values为因子值
name:str,行业代码
"sw_l1": 申万一级行业
"sw_l2": 申万二级行业
"sw_l3": 申万三级行业
"jq_l1": 聚宽一级行业
"jq_l2": 聚宽二级行业
"zjw": 证监会行业
date:获取行业数据的时间
maket_cap:市值因子
'''
industry_se = get_stock_industry(industry_name,date)
columns = list(data.columns)
if isinstance(industry_se,pd.Series):
industry_se = industry_se.to_frame()
if isinstance(market_cap,pd.Series):
market_cap = market_cap.to_frame()
index = list(data.index)
industry_se = np.array(industry_se.ix[index,0].tolist())
industry_dummy = sm.categorical(industry_se,drop=True)
industry_dummy = pd.DataFrame(industry_dummy,index=index)
market_cap = np.log(market_cap.loc[index])
x = pd.concat([industry_dummy,market_cap],axis=1)
model = sm.OLS(data,x)
result = model.fit()
y_fitted = result.fittedvalues
neu_result = data - y_fitted
return neu_result
def get_month_profit(stocks,start_date,end_date,month_num=1,cal_num=3):
'''
获取月收益率数据,数据为本月相对于上月的增长率
input:
stocks:list 股票代码
start_date:str, 初始日期
end_date:str,终止日期
month_num:计算几个月的收益率,默认为1,即一个月的收益率
cal_num:int,计算每月最后n天的收盘价均值,默认为3
'''
start_year = int(start_date[:4])
end_year = int(end_date[:4])
start_month = int(start_date[5:7])
end_month = int(end_date[5:7])
len_month = (end_year - start_year)*12 + (end_month - start_month)
price_list = []
#获取初始时间之前一个月的价格数据
if start_month == 1:
last_date = str(start_year-1)+'-'+'12'+'-'+'01'
else:
last_date = str(start_year-1)+'-'+str(start_month-1)+'-'+'01'
last_price = get_price(stocks,fields=['close'],count=cal_num,end_date=last_date)['close']
last_price = last_price.mean().to_frame()
last_price.columns = [last_date]
price_list.append(last_price)
#计算给定时间段内的月度价格数据
for i in range(len_month):
date = str(start_year+i//12)+'-'+str(start_month+i%12)+'-'+'01'
price = get_price(stocks,fields=['close'],count=cal_num,end_date=date)['close']
price_mean = price.mean().to_frame()
price_mean.columns = [date]
price_list.append(price_mean)
month_profit = pd.concat(price_list,axis=1)
#计算月度收益率
month_profit_pct = month_profit.pct_change(month_num,axis=1).dropna(axis=1,how='all')
return month_profit_pct
def get_profit_depend_timelist(stocks,timelist,month_num=1,cal_num=3):
'''
input:
stocks:list 股票代码
timelist: 时间序列
month_num:计算几个月的收益率,默认为1,即一个月的收益率
cal_num:int,计算每月最后n天的收盘价均值,默认为3
'''
price_list = []
for date in timelist:
price = get_price(stocks,fields=['close'],count=cal_num,end_date=date)['close']
price_mean = price.mean().to_frame()
price_mean.columns = [date]
price_list.append(price_mean)
profit = pd.concat(price_list,axis=1)
profit_pct = profit.pct_change(month_num,axis=1).dropna(axis=1,how='all')
return profit_pct
def get_day_profit_forward(stocks,end_date,start_date=None,count=-1,pre_num=20):
'''
获取收益率,pre_num为计算时间差,在时间轴上的当期值是未来计算周期内的收益率,
例如:pre_num=3,2013-01-01对应的收益率是2013-01-04的收益率与01-01日收益率之差
input:
stocks:list or Series,股票代码
start_date:开始时间
end_date:结束时间
count:与start_date二选一,向前取值个数
pre_num:int,向后计算的天数
output:
profit:dataframe,index为日期,columns为股票代码,values为收益率
'''
if count == -1:
price = get_price(stocks,start_date,end_date,fields=['close'])['close']
date_list = get_trade_days(start_date=start_date,end_date=end_date)
price.index = date_list
else:
price = get_price(stocks,end_date=end_date,count=count,fields=['close'])['close']
date_list = get_trade_days(end_date=end_date,count=count)
price.index = date_list
profit = price.pct_change(periods=pre_num).shift(-pre_num).dropna()
return profit
def get_one_day_data(stocks,factor_list,ttm_factors,date,neu=False):
'''
获取一天的基本面数据
input:
stocks:list,股票列表
factor_list:list,普通因子列表
ttm_factors:list,ttm因子列表
date:str or datetime, 获取数据时间
neu:bool,是否进行中性化处理,使用市值和行业进行中性化,默认不进行中性化
'''
fund_data = get_fundamental_data(stocks,factor_list,ttm_factors,date)
fillna_data = fillna_with_industry(fund_data,date)
if neu == False:
results = winsorize_and_standarlize(fillna_data)
elif 'MC' in fillna_data.columns:
neu_data = neutralize(fillna_data,date,fillna_data['MC'])
results = winsorize_and_standarlize(neu_data)
elif 'market_cap' in fillna_data.columns:
neu_data = neutralize(fillna_data,date,fillna_data['market_cap'])
results = winsorize_and_standarlize(neu_data)
else:
print("error: please input 'market_cap' for neutralize")
return None
return results
def get_timelist_data(stocks,factor_list,ttm_factors,timelist,neu=False):
dic = {}
for date in timelist:
fund_date = get_one_day_data(stocks,factor_list,ttm_factors,date,neu=neu)
dic[date] = fund_date
return dic
fund_data = get_timelist_data(all_stocks,factor_list,ttm_factors,tl)
fund_data_neu = get_timelist_data(all_stocks,factor_list,ttm_factors,tl,neu=True)
profit = get_profit_depend_timelist(all_stocks,tl,month_num=2,cal_num=3)
res = []
res.append(fund_data)
res.append(fund_data_neu)
res.append(profit)
#将数据输出到pickle文件
with open('fundamental_data.pkl','wb') as pk_file:
pickle.dump(res,pk_file)
import numpy as np
import pandas as pd
import pickle
import datetime
import statsmodels.api as sm
import warnings
from sklearn.feature_selection import RFE,SelectKBest,SelectPercentile,SelectFromModel,f_classif
import lightgbm as lgb
from sklearn.tree import DecisionTreeClassifier
from sklearn.svm import SVC,LinearSVC
from sklearn.ensemble import RandomForestClassifier,AdaBoostClassifier,GradientBoostingClassifier
from sklearn.model_selection import train_test_split,GridSearchCV
import gc
from sklearn.metrics import accuracy_score,recall_score
warnings.filterwarnings('ignore')
with open('fundamental_data.pkl','rb') as pk_file:
data_pk = pickle.load(pk_file)
fund_data = data_pk[0]
fund_data_neu = data_pk[1]
profit = data_pk[2]
keys = list(fund_data.keys())
#截面数据,将profit与基本面数据对齐,基本面数据对应下一月的profit
def get_fund_profit_data(fund_data,profit):
'''
input:
fund_data:dic key为日期,values为dataframe,基本面数据,index为股票代码,columns为因子
profit:dataframe,index为股票代码,columns为时间
注意:此函数针对于fund_data keys日期与profit日期在位置上已经对应
'''
keys = list(fund_data.keys())
columns = profit.columns.tolist()
l = min(len(keys),len(columns))
fund_profit = {}
for i in range(l):
fd = fund_data[keys[i]].copy() #复制新的dataframe,否则fund_profit为引用,在fund_profit上修改值会直接影响到fund_data
p = profit[columns[i]].to_frame()
p.columns = ['profit']
fd_merge = pd.merge(fd,p,left_index=True,right_index=True,how='inner')
fund_profit[keys[i]] = fd_merge
return fund_profit
def get_fund_profit_class_data(fund_data,profit):
'''
profit不再是数值,而是类别,大于0标记为1,小于0标记为0
input:
fund_data:dic key为日期,values为dataframe,基本面数据,index为股票代码,columns为因子
profit:dataframe,index为股票代码,columns为时间
output:
fund_profit:dic, 在fund_data每个dataframe后面加了profit列
注意:此函数针对于fund_data keys日期与profit日期在位置上已经对应
'''
pf = profit.copy(deep=True)
pf[pf>0] = 1
pf[pf<0] = 0
keys = list(fund_data.keys())
columns = pf.columns.tolist()
l = min(len(keys),len(columns))
fund_profit = {}
for i in range(l):
fd = fund_data[keys[i]].copy() #复制新的dataframe,否则fund_profit为引用,在fund_profit上修改值会直接影响到fund_data
p = pf[columns[i]].to_frame()
p.columns = ['profit']
fd_merge = pd.merge(fd,p,left_index=True,right_index=True,how='inner')
fund_profit[keys[i]] = fd_merge
return fund_profit
fund_profit_data = get_fund_profit_data(fund_data,profit)
fund_profit_data_neu = get_fund_profit_data(fund_data_neu,profit)
fund_profit_class_data = get_fund_profit_class_data(fund_data,profit)
fund_profit_class_data_neu = get_fund_profit_class_data(fund_data_neu,profit)
#使用稳健回归(sm.RLM)robust linear model
'''
用回归系数作为因子有效性的指标,如果因子与收益之间非线性,则此指标不准确,此指标作为参考之一
'''
def get_RLM_res(fund_profit_data):
'''
input:
fund_profit_data:dic, keys为日期,values为dataframe,基本面数据,index为股票代码,columns为因子,columns最后一列为profit
output:
f: dataframe, index为因子,columns为时间,values为稳健回归系数
t: dataframe,index为因子,columns为时间,values为稳健回归系数的t值
'''
keys = fund_profit_data.keys()
f_dic = {}
t_dic = {}
f = []
t = []
for k in keys:
col = fund_profit_data[k].columns
f_list = []
t_list = []
for c in col[:-1]:
df = fund_profit_data[k]
rlm_model = sm.RLM(df[col[-1]],df[c],M=sm.robust.norms.HuberT()).fit()
f_list.append(rlm_model.params)
t_list.append(rlm_model.tvalues)
f_list = [f_list[i].values[0] for i in range(len(f_list))]
t_list = [t_list[i].values[0] for i in range(len(t_list))]
f_df_k = pd.DataFrame(f_list,index=list(col[:-1]),columns=[k])
t_df_k = pd.DataFrame(t_list,index=col[:-1],columns=[k])
f.append(f_df_k)
t.append(t_df_k)
f = pd.concat(f,axis=1)
t = pd.concat(t,axis=1)
return f,t
rlm_res = get_RLM_res(fund_profit_data)
rlm_neu_res = get_RLM_res(fund_profit_data_neu)
'''
信息系数IC值,可以有效的观察到某个因子收益率预测的稳定性和动量特征,以便在组合优化时用作筛选的指标。常见的IC值计算方法有两种:
相关系数(Pearson Correlation)和秩相关系数(Spearman Rank Correlation),此例中IC值统计用到的是秩相关系数,
与IC相关的用来判断因子的有效性和预测能力指标如下:
IC值的均值
IC值的标准差
IC值大于0的比例
IC绝对值大于0.02的比例
IR (IC均值与IC标准差的比值)
参考:https://www.joinquant.com/post/16105?tag=algorithm
'''
def get_IC(fund_profit_data):
'''
input:
fund_profit_data:dic, keys为日期,values为dataframe,基本面数据,index为股票代码,columns为因子,columns最后一列为profit
output:
p: dataframe, index为因子,columns为时间,values为pearson相关系数
s: dataframe,index为因子,columns为时间,values为spearman相关系数
'''
keys = fund_profit_data.keys()
p_dic = {}
s_dic = {}
p = []
s = []
for k in keys:
columns = fund_profit_data[k].columns
p_list = []
s_list = []
for c in columns[:-1]:
df = fund_profit_data[k]
p_value = df[c].corr(df[columns[-1]],method='pearson')
s_value = df[c].corr(df[columns[-1]],method='spearman')
p_list.append(p_value)
s_list.append(s_value)
p_df_k = pd.DataFrame(p_list,index=list(columns[:-1]),columns=[k])
s_df_k = pd.DataFrame(s_list,index=columns[:-1],columns=[k])
p.append(p_df_k)
s.append(s_df_k)
p = pd.concat(p,axis=1)
s = pd.concat(s,axis=1)
return p,s
ic_res = get_IC(fund_profit_data)
ic_res_neu = get_IC(fund_profit_data_neu)
#好不容易跑完的数据,赶紧保存一下子
res = []
res.append(rlm_res)
res.append(rlm_neu_res)
res.append(ic_res)
res.append(ic_res_neu)
with open('judge_data.pkl','wb') as pf:
pickle.dump(res,pf)
#计算IC的各个指标
def cal_IC_indicator(data):
'''
input:
data:dataframe,index为因子,columns为日期
output:
res:dataframe,index为因子,输出计算好的各个评判标准
'''
data = data.stack().unstack(0)
data_mean = data.mean()
data_std = data.std()
data_ir = data_mean / data_std
data_mean_df = data_mean.to_frame()
data_mean_df.columns = ['mean']
data_std_df = data_std.to_frame()
data_std_df.columns = ['std']
data_ir_df = data_ir.to_frame()
data_ir_df.columns = ['IR']
data_nagative = (data[data > 0]).count() / len(data)
data_nagative_df = data_nagative.to_frame()
data_nagative_df.columns = ['IC正值比例']
data_abs_dayu = (data[data.abs() > 0.02]).count() / len(data)
data_abs_dayu_df = data_abs_dayu.to_frame()
data_abs_dayu_df.columns = ['IC绝对值>0.02']
res = pd.concat([data_mean_df,data_std_df,data_nagative_df,data_abs_dayu_df,data_ir_df],axis=1)
return res
ic_indicator_pearson = cal_IC_indicator(ic_res[0])
class FeatureSelection():
'''
特征选择:
identify_collinear:基于相关系数,删除小于correlation_threshold的特征
identify_importance_lgbm:基于LightGBM算法,得到feature_importance,选择和大于p_importance的特征
filter_select:单变量选择,指定k,selectKBest基于method提供的算法选择前k个特征,selectPercentile选择前p百分百的特征
wrapper_select:RFE,基于estimator递归特征消除,保留n_feature_to_select个特征
embedded_select: 基于estimator,
'''
def __init__(self):
self.supports = None #bool型,特征是否被选中
self.columns = None #选择的特征
self.record_collinear = None #自相关矩阵大于门限值
self.feature_importances = None #lgbm算法保存特征重要性值
def identify_collinear(self, data, correlation_threshold):
"""
Finds collinear features based on the correlation coefficient between features.
For each pair of features with a correlation coefficient greather than `correlation_threshold`,
only one of the pair is identified for removal.
Using code adapted from: https://gist.github.com/Swarchal/e29a3a1113403710b6850590641f046c
Parameters
--------
data : dataframe
Data observations in the rows and features in the columns
correlation_threshold : float between 0 and 1
Value of the Pearson correlation cofficient for identifying correlation features
"""
columns = data.columns
self.correlation_threshold = correlation_threshold
# Calculate the correlations between every column
corr_matrix = data.corr()
self.corr_matrix = corr_matrix
# Extract the upper triangle of the correlation matrix
upper = corr_matrix.where(np.triu(np.ones(corr_matrix.shape), k = 1).astype(np.bool))
# Select the features with correlations above the threshold
# Need to use the absolute value
to_drop = [column for column in upper.columns if any(upper[column].abs() > correlation_threshold)]
obtain_columns = [column for column in columns if column not in to_drop]
self.columns = obtain_columns
# Dataframe to hold correlated pairs
record_collinear = pd.DataFrame(columns = ['drop_feature', 'corr_feature', 'corr_value'])
# Iterate through the columns to drop
for column in to_drop:
# Find the correlated features
corr_features = list(upper.index[upper[column].abs() > correlation_threshold])
# Find the correlated values
corr_values = list(upper[column][upper[column].abs() > correlation_threshold])
drop_features = [column for _ in range(len(corr_features))]
# Record the information (need a temp df for now)
temp_df = pd.DataFrame.from_dict({'drop_feature': drop_features,
'corr_feature': corr_features,
'corr_value': corr_values})
# Add to dataframe
record_collinear = record_collinear.append(temp_df, ignore_index = True)
self.record_collinear = record_collinear
return data[obtain_columns]
def identify_importance_lgbm(self, features, labels,p_importance=0.8, eval_metric='auc', task='classification',
n_iterations=10, early_stopping = True):
# One hot encoding
data = features
features = pd.get_dummies(features)
# Extract feature names
feature_names = list(features.columns)
# Convert to np array
features = np.array(features)
labels = np.array(labels).reshape((-1, ))
# Empty array for feature importances
feature_importance_values = np.zeros(len(feature_names))
print('Training Gradient Boosting Model\n')
# Iterate through each fold
for _ in range(n_iterations):
if task == 'classification':
model = lgb.LGBMClassifier(n_estimators=100, learning_rate = 0.05, verbose = -1)
elif task == 'regression':
model = lgb.LGBMRegressor(n_estimators=100, learning_rate = 0.05, verbose = -1)
else:
raise ValueError('Task must be either "classification" or "regression"')
# If training using early stopping need a validation set
if early_stopping:
train_features, valid_features, train_labels, valid_labels = train_test_split(features, labels, test_size = 0.15)
# Train the model with early stopping
model.fit(train_features, train_labels, eval_metric = eval_metric,
eval_set = [(valid_features, valid_labels)],
verbose = -1)
# Clean up memory
gc.enable()
del train_features, train_labels, valid_features, valid_labels
gc.collect()
else:
model.fit(features, labels)
# Record the feature importances
feature_importance_values += model.feature_importances_ / n_iterations
feature_importances = pd.DataFrame({'feature': feature_names, 'importance': feature_importance_values})
# Sort features according to importance
feature_importances = feature_importances.sort_values('importance', ascending = False).reset_index(drop = True)
# Normalize the feature importances to add up to one
feature_importances['normalized_importance'] = feature_importances['importance'] / feature_importances['importance'].sum()
feature_importances['cumulative_importance'] = np.cumsum(feature_importances['normalized_importance'])
#obtain feature importance
self.feature_importances = feature_importances
select_df = feature_importances[feature_importances['cumulative_importance']<=p_importance]
select_columns = select_df['feature']
self.columns = list(select_columns.values)
res = data[self.columns]
return res
def filter_select(self, data_x, data_y, k=None, p=50,method=f_classif):
columns = data_x.columns
if k != None:
model = SelectKBest(method,k)
res = model.fit_transform(data_x,data_y)
supports = model.get_support()
else:
model = SelectPercentile(method,p)
res = model.fit_transform(data_x,data_y)
supports = model.get_support()
self.support_ = supports
self.columns = columns[supports]
return res
def wrapper_select(self,data_x,data_y,n,estimator):
columns = data_x.columns
model = RFE(estimator=estimator,n_features_to_select=n)
res = model.fit_transform(data_x,data_y)
supports = model.get_support() #标识被选择的特征在原数据中的位置
self.supports = supports
self.columns = columns[supports]
return res
def embedded_select(self,data_x,data_y,estimator,threshold=None):
columns = data_x.columns
model = SelectFromModel(estimator=estimator,prefit=False,threshold=threshold)
res = model.fit_transform(data_x,data_y)
supports = model.get_support()
self.supports = supports
self.columns = columns[supports]
return res
#使用特征选择方法选择因子值
test_fund = fund_profit_class_data_neu[keys[0]] #取一期数据测试
test_fund1 = fund_profit_class_data_neu[keys[1]]
#test_fund = pd.concat([test_fund0,test_fund1])
columns = test_fund.columns
fs = FeatureSelection()
x = test_fund[columns[:-1]]
y = test_fund[columns[-1]]
lgbm = fs.identify_importance_lgbm(x,y) #使用特征选择方法选择有效因子
fs.feature_importances #各个因子重要性
'''
以下代码是示例代码,简单走一遍机器学习探索策略及调参,具体有效的策略请大家自己探索,不做分享
'''
fund_data_train = fund_profit_class_data_neu[keys[1]]
columns_s = lgbm.columns
col_s = columns_s[:-1]
fund_data_train_y = fund_data_train[fund_data_train.columns[-1]]
lgbm_x_train,lgbm_x_test,lgbm_y_train,lgbm_y_test = train_test_split(fund_data_train[col_s],fund_data_train_y,test_size=0.3)
lgbm_svm = SVC(max_iter=1000)
param_grid = {'C':[0.1,1,3],'kernel':['rbf','sigmoid','linear','poly'],'gamma':np.arange(0.3,0.8,0.1)}
lgbm_model = GridSearchCV(estimator=lgbm_svm,param_grid=param_grid,scoring='accuracy')
lgbm_model.fit(lgbm_x_train,lgbm_y_train)
lgbm_test_res = lgbm_model.predict(lgbm_x_test)
accuracy = accuracy_score(lgbm_y_test,lgbm_test_res)
print('accuracy is: %0.5f'%accuracy)
print(lgbm_model.best_params_)
print(lgbm_model.best_score_)
gbdt = GradientBoostingClassifier()
gbdt_params_grid = {'max_depth':[4,6,8],'min_samples_split':[10,20,30]}
gbdt_model = GridSearchCV(estimator=gbdt,param_grid=gbdt_params_grid)
gbdt_model.fit(lgbm_x_train,lgbm_y_train)
gbdt_test_res = gbdt_model.predict(lgbm_x_test)
gbdt_accuracy = accuracy_score(lgbm_y_test,gbdt_test_res)
print('accuracy is: %0.5f'%gbdt_accuracy)
print(gbdt_model.best_params_)
print(gbdt_model.best_score_)
fund_for_pre = fund_profit_class_data_neu[keys[3]] #取一期的截面数据验证
columns_for_pre = fund_for_pre.columns
x_for_pre = fund_for_pre[col_s]
y_for_pre = fund_for_pre[columns_for_pre[-1]]
prediction = lgbm_model.predict(x_for_pre)
accuracy_for_pre = accuracy_score(y_for_pre,prediction)
print(accuracy_for_pre)
print(len(prediction))
本社区仅针对特定人员开放
查看需注册登录并通过风险意识测评
5秒后跳转登录页面...
移动端课程