def test_q_anl_impl(self):
n_bins = 5
x = self.x[:, 0]
q_groups = quantile(x, n_bins)
s = pd.Series(self.r, index=q_groups)
expected_res = s.groupby(level=0).mean()
calculated_res = er_quantile_analysis(x, n_bins, self.r)
np.testing.assert_array_almost_equal(expected_res.values, calculated_res)
def test_quantile(self):
n = 5000
bins = 10
s = np.random.randn(n)
calculated = quantile(s, bins)
rank = s.argsort().argsort()
bin_size = float(n) / bins
pillars = [int(i * bin_size) for i in range(1, bins + 1)]
starter = 0
for i, r in enumerate(pillars):
self.assertTrue(
np.all(calculated[(rank >= starter) & (rank < r)] == i))
starter = r
def er_quantile_analysis(er: np.ndarray, n_bins: int,
dx_return: np.ndarray) -> np.ndarray:
er = er.flatten()
q_groups = quantile(er, n_bins)
if dx_return.ndim < 2:
dx_return.shape = -1, 1
group_return = agg_mean(q_groups, dx_return).flatten()
total_return = group_return.sum()
ret = group_return.copy()
resid = n_bins - 1
res_weight = 1. / resid
for i, value in enumerate(ret):
ret[i] = (1. + res_weight) * value - res_weight * total_return
return ret
def test_q_anl_impl(self):
n_bins = 5
x = self.x[:, 0]
q_groups = quantile(x, n_bins)
s = pd.Series(self.r, index=q_groups)
grouped_return = s.groupby(level=0).mean().values.flatten()
expected_res = grouped_return.copy()
res = n_bins - 1
res_weight = 1. / res
for i, value in enumerate(expected_res):
expected_res[i] = (1. + res_weight) * value - res_weight * grouped_return.sum()
calculated_res = er_quantile_analysis(x, n_bins, self.r, de_trend=True)
np.testing.assert_array_almost_equal(expected_res, calculated_res)
def er_quantile_analysis(
er: np.ndarray,
n_bins: int,
dx_return: np.ndarray,
benchmark: Optional[np.ndarray] = None,
) -> np.ndarray:
er = er.flatten()
q_groups = quantile(er, n_bins)
if dx_return.ndim < 2:
dx_return.shape = -1, 1
group_return = agg_mean(q_groups, dx_return).flatten()
if benchmark is not None:
b_ret = np.dot(benchmark, dx_return)
b_total = benchmark.sum()
return group_return * b_total - b_ret
else:
return group_return
def basic_info(destsession, session, factor_name, task_id, fac_rets_series,
ic_series, cum_df, total_data, kwargs):
#收益率
fac_rets = fac_rets_series.cumsum().values[-1]
#收益率t值
fac_rets_ttest = ttest_1samp(fac_rets_series, 0)
t_rets = fac_rets_ttest.statistic
ic_mean = ic_series.mean()
ic_std = ic_series.std()
ic_marked = len(ic_series[ic_series.abs() > 0.02]) / len(ic_series)
ir = ic_mean / ic_std
#年化收益率
annualized = cum_df.q0.values[-1] / len(cum_df.q0) * 250
# 换手率
n_bins = 5
last_code = None
diff_count = 0
sum_count = 0
res = []
grouped = total_data.groupby('trade_date')
for k, g in grouped:
er = g['prc_factor'].values
g['group'] = quantile(er.flatten(), n_bins)
res.append(g)
turnover_df = pd.concat(res).reset_index()[['trade_date', 'code', 'group']]
if cum_df.q0.cumsum().values[-1] > cum_df.q4.cumsum().values[-1]:
group_df = turnover_df.set_index('group').loc[0].reset_index()
else:
group_df = turnover_df.set_index('group').loc[4].reset_index()
grouped = group_df.groupby('trade_date')
for k, g in grouped:
if last_code is None:
sum_count = len(g.code.values)
last_code = g.code.values
else:
mix_code = set(g.code.values) & set(last_code)
diff_count += (len(g.code.values) - len(mix_code))
sum_count += len(g.code.values)
turnover_rate = diff_count / sum_count
basic_info = {
'fac_rets': fac_rets,
't_rets': t_rets,
'ic_mean': ic_mean,
'ic_std': ic_std,
'ic_marked': ic_marked,
'ir': ir,
'annualized': annualized,
'turnover_rate': turnover_rate,
'update_time': datetime.datetime.now().strftime('%Y-%m-%d %H:%S:%M'),
'params': str(json.dumps(kwargs)),
'remark': '中证500股票池'
}
basic_df = pd.DataFrame([basic_info])
basic_df['session'] = session
basic_df['factor_name'] = factor_name
basic_df['task_id'] = task_id
update_destdb(destsession, 'basic_info', basic_df, is_trade=False)