在本例中，使用公式选股并进行回测
In [1]:
%matplotlib inline
import os
import datetime as dt
import pandas as pd
import numpy as np
from PyFin.api import *
from alphamind.api import *
from matplotlib import pyplot as plt
In [50]:
freq = "1w"
universe = Universe('zz800')
engine = SqlEngine(os.environ['DB_URI'])
start_date = '2017-05-31'
end_date = '2019-05-31'
benchmark_code = 300
rebalance_dates = makeSchedule(start_date, end_date, freq, 'china.sse', BizDayConventions.Preceding)
In [3]:
formula = CSTopN(LAST('ETOP'), 3, groups='sw1')
In [4]:
%%time
factors = engine.fetch_factor_range(universe, {'alpha': formula}, dates=rebalance_dates)
industries = engine.fetch_industry_range(universe, dates=rebalance_dates)
In [5]:
%%time
chosen_stks = factors[factors.alpha == 1.]
codes = chosen_stks.code.unique()
stk_returns = engine.fetch_dx_return_range(universe, start_date, end_date)
In [6]:
%%time
ret_dates = np.unique(stk_returns.trade_date)
rebalance_dates = chosen_stks.trade_date.unique()
rtn_groups = stk_returns.groupby('trade_date')
industry_groups = industries.groupby('trade_date')
stks_groups = chosen_stks.groupby('trade_date')
portfolio_rtns = []
portfolios = pd.DataFrame(columns=['code', 'weight', 'industry'])
for ref_date, g in rtn_groups:
ref_date = ref_date.to_datetime64()
nearest_date = rebalance_dates[rebalance_dates <= ref_date][-1]
this_codes = stks_groups.get_group(nearest_date)['code'].tolist()
next_day_rtn = g.loc[g.code.isin(this_codes), 'dx'].values
portfolio_nd_rtn = np.mean(next_day_rtn)
portfolio_rtns.append(portfolio_nd_rtn)
if ref_date in rebalance_dates:
this_industry = industry_groups.get_group(nearest_date).set_index('code')
this_df = pd.DataFrame(data={'code': this_codes,
'weight': np.ones(len(this_codes)) * 1. / len(this_codes),
'industry': this_industry.loc[this_codes, 'industry'].values})
this_df.index =[ref_date] * len(this_codes)
portfolios = portfolios.append(this_df)
portfolio_rtns = pd.DataFrame({'ret': portfolio_rtns}, index=ret_dates)
In [7]:
from PyFin.Math.Accumulators import MovingDrawdown
In [8]:
%%time
mv = MovingDrawdown(window=1000, x='ret')
dds = []
for r in portfolio_rtns['ret'].values:
mv.push({'ret': r})
dds.append(mv.value)
portfolio_rtns['dd.'] = dds
In [9]:
portfolio_rtns['ret'].cumsum().plot(figsize=(14, 7))
portfolio_rtns['dd.'].plot(kind='area')
plt.legend(['ret', 'dd.'])
plt.ylim([-0.6, 0.3])
Out[9]:
In [19]:
print("最大回撤： {0:.2f}%".format(-np.min(portfolio_rtns['dd.']) * 100.))
print("年化收益： {0:.2f}%".format(250 * np.mean(portfolio_rtns['ret']) * 100.))
In [95]:
%%time
benchmark = engine.fetch_dx_return_index_range(benchmark_code, start_date, end_date)[['trade_date', 'dx']].set_index('trade_date')['dx']
In [96]:
excess_ret = portfolio_rtns['ret'] - benchmark
print("年化超额收益： {0:.2f}%".format(250 * np.mean(excess_ret) * 100.))
In [102]:
%%time
p_rets = portfolio_rtns['ret'].values
b_rets = benchmark.values
x = np.concatenate([np.ones((len(b_rets),1)), b_rets.reshape((-1, 1))], axis=1)
alpha, beta = np.linalg.solve(x.T @ x, x.T @ p_rets)
print("年化alpha： {0:.2f}%".format(250 * alpha * 100.))
print("beta： {0:.2f}".format(beta))
In [98]:
portfolio_rtns['ret'].resample('M').sum().plot(figsize=(14, 7), kind='bar')
plt.legend(['ret'])
Out[98]:
In [99]:
portfolios.reset_index().groupby(['index', 'industry']).count()
Out[99]:
In [103]:
%%time
p_groups = portfolios.groupby(level=0)
previous_date = None
previous_pos = None
turn_over = []
for ref_date, g in p_groups:
g = g.set_index('code')
if previous_date:
diff_pos = g['weight'].subtract(previous_pos, fill_value=0)
turn_over.append(diff_pos.abs().sum())
previous_date = ref_date
previous_pos = g['weight']
turn_over = pd.Series(turn_over, index=portfolios.index.unique()[1:])
In [104]:
turn_over.resample('M').sum().plot(kind='bar', figsize=(16, 7))
Out[104]:
In [ ]: