def _plot_bayes_cone(returns_train, returns_test,
preds, plot_train_len=None, ax=None):
if ax is None:
ax = plt.gca()
returns_train_cum = cum_returns(returns_train, starting_value=1.)
returns_test_cum = cum_returns(returns_test,
starting_value=returns_train_cum.iloc[-1])
perc = compute_bayes_cone(preds, starting_value=returns_train_cum.iloc[-1])
# Add indices
perc = {k: pd.Series(v, index=returns_test.index) for k, v in perc.items()}
returns_test_cum_rel = returns_test_cum
# Stitch together train and test
returns_train_cum.loc[returns_test_cum_rel.index[0]] = \
returns_test_cum_rel.iloc[0]
# Plotting
if plot_train_len is not None:
returns_train_cum = returns_train_cum.iloc[-plot_train_len:]
returns_train_cum.plot(ax=ax, color='g', label='In-sample')
returns_test_cum_rel.plot(ax=ax, color='r', label='Out-of-sample')
ax.fill_between(returns_test.index, perc[5], perc[95], alpha=.3)
ax.fill_between(returns_test.index, perc[25], perc[75], alpha=.6)
ax.legend(loc='best', frameon=True, framealpha=0.5)
ax.set_title('Bayesian cone')
ax.set_xlabel('')
ax.set_ylabel('Cumulative returns')
return ax
def calculate_metrics(self):
self.benchmark_period_returns = \
cum_returns(self.benchmark_returns).iloc[-1]
self.algorithm_period_returns = \
cum_returns(self.algorithm_returns).iloc[-1]
if not self.algorithm_returns.index.equals(
self.benchmark_returns.index):
message = "Mismatch between benchmark_returns ({bm_count}) and \
algorithm_returns ({algo_count}) in range {start} : {end}"
message = message.format(bm_count=len(self.benchmark_returns),
algo_count=len(self.algorithm_returns),
start=self._start_session,
end=self._end_session)
raise Exception(message)
self.num_trading_days = len(self.benchmark_returns)
self.mean_algorithm_returns = (
self.algorithm_returns.cumsum() /
np.arange(1, self.num_trading_days + 1, dtype=np.float64))
self.benchmark_volatility = annual_volatility(self.benchmark_returns)
self.algorithm_volatility = annual_volatility(self.algorithm_returns)
self.treasury_period_return = choose_treasury(
self.treasury_curves,
self._start_session,
self._end_session,
self.trading_calendar,
)
self.sharpe = sharpe_ratio(self.algorithm_returns, )
# The consumer currently expects a 0.0 value for sharpe in period,
# this differs from cumulative which was np.nan.
# When factoring out the sharpe_ratio, the different return types
# were collapsed into `np.nan`.
# TODO: Either fix consumer to accept `np.nan` or make the
# `sharpe_ratio` return type configurable.
# In the meantime, convert nan values to 0.0
if pd.isnull(self.sharpe):
self.sharpe = 0.0
self.downside_risk = downside_risk(self.algorithm_returns.values)
self.sortino = sortino_ratio(
self.algorithm_returns.values,
_downside_risk=self.downside_risk,
)
self.information = information_ratio(
self.algorithm_returns.values,
self.benchmark_returns.values,
)
self.alpha, self.beta = alpha_beta_aligned(
self.algorithm_returns.values,
self.benchmark_returns.values,
)
self.excess_return = self.algorithm_period_returns - \
self.treasury_period_return
self.max_drawdown = max_drawdown(self.algorithm_returns.values)
self.max_leverage = self.calculate_max_leverage()
def _plot_bayes_cone(returns_train, returns_test,
preds, plot_train_len=None, ax=None):
if ax is None:
ax = plt.gca()
returns_train_cum = cum_returns(returns_train, starting_value=1.)
returns_test_cum = cum_returns(returns_test,
starting_value=returns_train_cum.iloc[-1])
perc = compute_bayes_cone(preds, starting_value=returns_train_cum.iloc[-1])
# Add indices
perc = {k: pd.Series(v, index=returns_test.index) for k, v in perc.items()}
returns_test_cum_rel = returns_test_cum
# Stitch together train and test
returns_train_cum.loc[returns_test_cum_rel.index[0]] = \
returns_test_cum_rel.iloc[0]
# Plotting
if plot_train_len is not None:
returns_train_cum = returns_train_cum.iloc[-plot_train_len:]
returns_train_cum.plot(ax=ax, color='g', label='In-sample')
returns_test_cum_rel.plot(ax=ax, color='r', label='Out-of-sample')
ax.fill_between(returns_test.index, perc[5], perc[95], alpha=.3)
ax.fill_between(returns_test.index, perc[25], perc[75], alpha=.6)
ax.legend(loc='best', frameon=True, framealpha=0.5)
ax.set_title('Bayesian cone')
ax.set_xlabel('')
ax.set_ylabel('Cumulative returns')
return ax
def _cumulative_returns_less_costs(returns, costs):
"""
Compute cumulative returns, less costs.
"""
if costs is None:
return ep.cum_returns(returns)
return ep.cum_returns(returns - costs)
def plot_alpha_curve(return_dict):
"""
:param return_dict: dict, returnName: [returnData, ReturnType]
:return:
"""
strat_return = return_dict['stratReturn'][0]
strat_return_type = return_dict['stratReturn'][1]
benchmark_return = return_dict['benchmarkReturn'][0]
benchmark_return_type = return_dict['benchmarkReturn'][1]
ptf_return = return_dict['ptfReturn'][0]
ptf_return_type = return_dict['ptfReturn'][1]
strat_return = cum_returns(strat_return,
starting_value=1.0) if strat_return_type == ReturnType.NonCumul else strat_return
benchmark_return = cum_returns(benchmark_return,
starting_value=1.0) if benchmark_return_type == ReturnType.NonCumul \
else benchmark_return
ptf_return = cum_returns(ptf_return, starting_value=1.0) if ptf_return_type == ReturnType.NonCumul else ptf_return
data = pd.concat([ptf_return, strat_return, benchmark_return], join_axes=[strat_return.index], axis=1)
# 如果缺失起始数据, 设置为1.0 （起始净值）
data = data.fillna(1.0)
data.columns = [u'策略对冲收益', u'策略未对冲收益', u'指数收益']
ax = data.plot(figsize=(16, 6), title=u'策略收益演示图')
fig_style(ax, [u'策略对冲净值', u'策略未对冲净值', u'指数收益'], x_label=u'交易日', y_label=u'净值',
legend_loc='upper left')
plt.show()
def plot_rolling_returns_multiple(returns_arr, factor_returns=None, logy=False, ax=None, names_arr=None, extra_bm=0):
"""
Plots cumulative rolling returns versus some benchmarks'.
This is based on https://github.com/quantopian/pyfolio/blob/master/pyfolio/plotting.py,
but modified to plot multiple rolling returns on the same graph
Arguments
----------
returns_arr : array of pd.Series. Each element contains daily returns of the strategy, noncumulative.t.
factor_returns : pd.Series, optional
Daily noncumulative returns of the benchmark factor to which betas are
computed. Usually a benchmark such as market returns.
- This is in the same style as returns.
logy : bool, optional
Whether to log-scale the y-axis.
ax : matplotlib.Axes, optional
Axes upon which to plot.
names_arr: array of names for the plots, optional
extra_bm: number of extra benchmarks. These will be assumed to be at the front of returns_array and will be plotted differently
Returns
-------
ax : matplotlib.Axes
The axes that were plotted on.
"""
if ax is None:
ax = plt.gca()
ax.set_xlabel('')
ax.set_ylabel('Cumulative returns')
ax.set_yscale('log' if logy else 'linear')
for i in range(len(returns_arr)):
# pData = perfData[i]
# returns, positions, transactions = pf.utils.extract_rets_pos_txn_from_zipline(pData)
returns = returns_arr[i]
returns.name = 'Portfolio %i' % i if names_arr is None else names_arr[i]
cum_rets = ep.cum_returns(returns, 1.0)
is_cum_returns = cum_rets
if (i == 0 and factor_returns is not None):
cum_factor_returns = ep.cum_returns(factor_returns[cum_rets.index], 1.0)
cum_factor_returns.plot(lw=1, color='gray', label=factor_returns.name, alpha=0.60, ax=ax, style=['-.'])
is_cum_returns.plot(lw=1, alpha=0.6, label=returns.name, ax=ax, style=['-.'] if (i < extra_bm) else None)
# is_cum_returns.plot(lw=1, alpha=0.6, label=returns.name, ax=ax)
years = mdates.YearLocator() # every year
months = mdates.MonthLocator() # every month
major_fmt = mdates.DateFormatter('%b %Y')
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
ax.xaxis.set_major_locator(years)
ax.xaxis.set_major_formatter(major_fmt)
ax.xaxis.set_minor_locator(months)
return ax
def _cumulative_returns_less_costs(returns, costs):
"""
Compute cumulative returns, less costs.
"""
if costs is None:
return ep.cum_returns(returns)
return ep.cum_returns(returns - costs)
def rolling_drawdown(self, returns):
out = np.empty(returns.shape[1:])
returns_1d = returns.ndim == 1
if len(returns) < 1:
out[()] = np.nan
if returns_1d:
out = out.item()
return out
returns_array = np.asanyarray(returns)
cumulative = np.empty((returns.shape[0] + 1, ) + returns.shape[1:],
dtype='float64')
cumulative[0] = start = 100
empyrical.cum_returns(returns_array,
starting_value=start,
out=cumulative[1:])
max_return = np.fmax.accumulate(cumulative, axis=0)
out = (cumulative - max_return) / max_return
out = pd.Series(out[1:])
return out
def test_cumulative(self):
res_a = empyrical.cum_returns(ret['a']).rename('a')
res_b = empyrical.cum_returns(ret['b']).rename('b')
res_c = empyrical.cum_returns(ret['c']).rename('c')
pd.testing.assert_series_equal(ret['a'].vbt.returns.cumulative(),
res_a)
pd.testing.assert_frame_equal(ret.vbt.returns.cumulative(),
pd.concat([res_a, res_b, res_c], axis=1))
def ptf_re_balance(return_dict, margin_prop=0.0, re_balance_freq=FreqType.EOM):
"""
:param return_dict: dict, returnName: [returnData, ReturnType]
:param margin_prop: float, optional, proportion of the init ptf that is allocated to futures account
:param re_balance_freq: str, optional, rebalance frequncy = daily/monthly/yearly
:return: pd.Series, daily cumul returns of hedged ptf
"""
strat_return = return_dict['stratReturn'][0]
strat_return_type = return_dict['stratReturn'][1]
benchmark_return = return_dict['benchmarkReturn'][0]
benchmark_return_type = return_dict['benchmarkReturn'][1]
strat_return = empyrical.cum_returns(strat_return,
starting_value=1.0) if strat_return_type == ReturnType.NonCumul \
else strat_return
benchmark_return = empyrical.cum_returns(benchmark_return,
starting_value=1.0) if benchmark_return_type == ReturnType.NonCumul \
else benchmark_return
pyFinAssert(0 <= margin_prop <= 1.0, ValueError,
" margin prop must be between 0 and 1")
hedged_ptf_return = pd.Series()
# merge strat and index returns together
return_data = pd.concat([strat_return, benchmark_return],
axis=1,
join_axes=[strat_return.index])
return_data.columns = ['strategy', 'benchmark']
return_data.index = pd.to_datetime(return_data.index)
pyFinAssert(return_data.isnull().values.any() == False, ValueError,
" returnData has NaN values")
regroup_total_return = regroup_by_re_balance_freq(return_data,
re_balance_freq)
# first date is a balance date
re_balance_base_nav = 1.0
norm_base_return = return_data.iloc[0]
for name, group in regroup_total_return:
# compute the hedged return
norm_strat_return = group['strategy'] / norm_base_return['strategy']
norm_benchmark_return = group['benchmark'] / norm_base_return[
'benchmark']
hedged_return = (1 + (norm_strat_return - norm_benchmark_return) *
(1 - margin_prop)) * re_balance_base_nav
# update the re_balance base NPV
re_balance_base_nav = hedged_return.iloc[-1]
# update norm base return
norm_base_return = group.iloc[-1]
# merge into ptfValue
hedged_ptf_return = pd.concat([hedged_ptf_return, hedged_return],
axis=0)
hedged_ptf_return.name = 'hedgedPtfReturn'
hedged_ptf_return.index.name = return_data.index.name
return hedged_ptf_return
def plot_returns(perf_attrib_data, cost=None, ax=None):
"""
Plot total, specific, and common returns.
Parameters
----------
perf_attrib_data : pd.DataFrame
df with factors, common returns, and specific returns as columns,
and datetimes as index. Assumes the `total_returns` column is NOT
cost adjusted.
- Example:
momentum reversal common_returns specific_returns
dt
2017-01-01 0.249087 0.935925 1.185012 1.185012
2017-01-02 -0.003194 -0.400786 -0.403980 -0.403980
cost : pd.Series, optional
if present, gets subtracted from `perf_attrib_data['total_returns']`,
and gets plotted separately
ax : matplotlib.axes.Axes
axes on which plots are made. if None, current axes will be used
Returns
-------
ax : matplotlib.axes.Axes
"""
if ax is None:
ax = plt.gca()
returns = perf_attrib_data['total_returns']
total_returns_label = 'Total returns'
if cost is not None:
returns = returns - cost
total_returns_label += ' (adjusted)'
specific_returns = perf_attrib_data['specific_returns']
common_returns = perf_attrib_data['common_returns']
ax.plot(ep.cum_returns(returns), color='b', label=total_returns_label)
ax.plot(ep.cum_returns(specific_returns),
color='g',
label='Cumulative specific returns')
ax.plot(ep.cum_returns(common_returns),
color='r',
label='Cumulative common returns')
if cost is not None:
ax.plot(cost, color='p', label='Cost')
ax.set_title('Time series of cumulative returns')
ax.set_ylabel('Returns')
configure_legend(ax)
return ax
def plot_returns(perf_attrib_data, cost=None, ax=None):
"""
Plot total, specific, and common returns.
Parameters
----------
perf_attrib_data : pd.DataFrame
df with factors, common returns, and specific returns as columns,
and datetimes as index. Assumes the `total_returns` column is NOT
cost adjusted.
- Example:
momentum reversal common_returns specific_returns
dt
2017-01-01 0.249087 0.935925 1.185012 1.185012
2017-01-02 -0.003194 -0.400786 -0.403980 -0.403980
cost : pd.Series, optional
if present, gets subtracted from `perf_attrib_data['total_returns']`,
and gets plotted separately
ax : matplotlib.axes.Axes
axes on which plots are made. if None, current axes will be used
Returns
-------
ax : matplotlib.axes.Axes
"""
if ax is None:
ax = plt.gca()
returns = perf_attrib_data['total_returns']
total_returns_label = 'Total returns'
if cost is not None:
returns = returns - cost
total_returns_label += ' (adjusted)'
specific_returns = perf_attrib_data['specific_returns']
common_returns = perf_attrib_data['common_returns']
ax.plot(ep.cum_returns(returns), color='b', label=total_returns_label)
ax.plot(ep.cum_returns(specific_returns), color='g',
label='Cumulative specific returns')
ax.plot(ep.cum_returns(common_returns), color='r',
label='Cumulative common returns')
if cost is not None:
ax.plot(cost, color='p', label='Cost')
ax.set_title('Time series of cumulative returns')
ax.set_ylabel('Returns')
configure_legend(ax)
return ax
def plot_returns(self):
self.portfolio_total_returns = empyrical.cum_returns(
self.analysis_data.chart_data.returns) * 100
self.benchmark_total_returns = empyrical.cum_returns(
self.analysis_data.chart_data.benchmark_returns) * 100
self.returns_ax.plot(self.portfolio_total_returns)
self.returns_ax.plot(self.benchmark_total_returns)
self.returns_ax.legend(['Strategy', 'SPY'], loc='upper left')
self.returns_ax.set_ylabel('Return')
def plot_returns(returns, factor_returns, transactions=None):
fig, ax = create_trading_figure([0.1, 0.2, 0.8, 0.7], 'DateTime',
'Return ( % )')
ax.xaxis.set_major_formatter(mdates.DateFormatter(fmt='%d-%m-%Y %H:%M'))
fig.autofmt_xdate()
returns = cum_returns(returns, starting_value=1) * 100
factor_returns = cum_returns(factor_returns, starting_value=1) * 100
x = [i.to_pydatetime() for i in returns.index]
ex_x = [x[0]] + x + [x[-1]]
ex_returns = pd.Series([100]).append(returns.append(pd.Series([100])))
ex_factor_returns = pd.Series([100]).append(
factor_returns.append(pd.Series([100])))
ax.fill(ex_x, ex_returns, color=(0, 1, 0, 0.3), zorder=0)
ax.fill(ex_x, ex_factor_returns, color=(1, 0, 0, 0.3), zorder=0)
ax.plot(x, returns, color='g', zorder=2)
ax.plot(x, factor_returns, color='r', zorder=1)
ax.plot(x, [100] * len(x), color='w', zorder=2, linewidth=3)
if transactions is not None:
buys_sells = transactions.iloc[transactions['amount'].nonzero()[0]]
buys = buys_sells['amount'][buys_sells['amount'] > 0]
sells = buys_sells['amount'][buys_sells['amount'] < 0]
ax.scatter(x=[i.to_pydatetime() for i in buys.index],
y=[returns[i] for i in buys.index],
c=[(0, 1, 0, 0.5)] * len(buys),
s=100,
marker='v',
edgecolor=(0, 1, 0, 0.9),
zorder=2)
ax.scatter(x=[i.to_pydatetime() for i in sells.index],
y=[returns[i] for i in sells.index],
c=[(1, 0, 0, 0.5)] * len(buys),
s=100,
marker='^',
edgecolor=(1, 0, 0, 0.9),
zorder=2)
fig.suptitle('Strategy returns %s\n%s to %s' %
(transactions['symbol'].iloc[0],
returns.index[0].strftime('%d-%m-%Y'),
returns.index[-1].strftime('%d-%m-%Y')))
else:
fig.suptitle('Strategy returns\n%s to %s' %
(returns.index[0].strftime('%d-%m-%Y'),
returns.index[-1].strftime('%d-%m-%Y')))
ax.set_xlim(x[0], x[-1])
d1, d2 = max(max(returns), max(factor_returns), 100) - 100, 100 - min(
min(returns), min(factor_returns), 100)
ax.set_ylim(100 - d2 * 1.1, 100 + d1 * 1.1)
plt.show()
def statsAndPlot(self):
logger.info('Statistics and Plots')
for account in self._accounts:
logger.info('Account: %s', account.getName())
logger.info(account.getRecorder().getRecords())
logger.info('Official Report')
logger.info(account.getOfficialRecorder().getRecords())
rec = account.getOfficialRecorder().getRecords()
from pprint import pprint
pprint(recorder.Stats.getStats(rec.returns, rec.benchmark_returns))
r_cum = empyrical.cum_returns(rec.returns)
bc_cum = empyrical.cum_returns(rec.benchmark_returns)
m = pd.DataFrame({'wealth': r_cum, 'benchmark': bc_cum})
m2 = (1 + m) * 100
# m2.plot(figsize=(15, 5), title = 'Wealth Curve')
records = account.getRecorder().getRecords()
ore = account.getOfficialRecorder().getRecords()
lastLine = records.iloc[-1]
lastLine.set_value('wealth', ore.iloc[-1].get_value('wealth'))
lastLine.name += pd.Timedelta(days=1)
records = records.append(lastLine)
records['trades'] = records.order.apply(len)
records['positionsNum'] = records.positions.apply(len)
fig = plt.figure()
ax1 = fig.add_subplot(2, 2, 1)
ax2 = fig.add_subplot(2, 2, 2)
ax3 = fig.add_subplot(2, 2, 3)
m2.plot(figsize=(15, 5), title='Wealth Curve', ax=ax1)
ax1.set_title('Wealth Curve')
for label in ax1.get_xticklabels():
label.set_rotation(20)
ax2.plot(records['positionsNum'])
ax2.set_title('Positions Number')
for label in ax2.get_xticklabels():
label.set_rotation(20)
ts = records['trades'].groupby(pd.TimeGrouper(freq='1M')).sum()
y_pos = np.arange(len(ts))
ax3.bar(y_pos, ts)
ax3.set_xticks(y_pos[::3])
ax3.set_xticklabels(ts.index.strftime('%Y-%m')[::3], rotation=20)
fmt = '%.0f%%' # Format you want the ticks, e.g. '40%'
yticks = mtick.FormatStrFormatter(fmt)
ax3.yaxis.set_major_formatter(yticks)
ax3.set_title('Turnover Rate')
fig.show()
def compute_consistency_score(returns_test, preds):
"""
Compute Bayesian consistency score.
Parameters
----------
returns_test : pd.Series
Observed cumulative returns.
preds : numpy.array
Multiple (simulated) cumulative returns.
Returns
-------
Consistency score
Score from 100 (returns_test perfectly on the median line of the
Bayesian cone spanned by preds) to 0 (returns_test completely
outside of Bayesian cone.)
"""
returns_test_cum = cum_returns(returns_test, starting_value=1.)
cum_preds = np.cumprod(preds + 1, 1)
q = [sp.stats.percentileofscore(cum_preds[:, i],
returns_test_cum.iloc[i],
kind='weak')
for i in range(len(returns_test_cum))]
# normalize to be from 100 (perfect median line) to 0 (completely outside
# of cone)
return 100 - np.abs(50 - np.mean(q)) / .5
def get_max_drawdown(returns):
"""
Determines the maximum drawdown of a strategy.
Parameters
----------
returns : pd.Series
Daily returns of the strategy, noncumulative.
- See full explanation in :func:`~empyrical.cum_returns`.
Returns
-------
float
Maximum drawdown.
Note
-----
See https://en.wikipedia.org/wiki/Drawdown_(economics) for more details.
"""
returns = returns.copy()
df_cum = ep.cum_returns(returns, 1.0)
running_max = np.maximum.accumulate(df_cum)
underwater = df_cum / running_max - 1
return get_max_drawdown_underwater(underwater)
def performance_measures(pct_chg, y):
result = {}
y_init = list(map(reverse_func, y))
predict = pd.Series(index=pct_chg.index, data=y_init)
predict.name = 'label'
df = pd.concat([pct_chg, predict.shift(1)], axis=1)
df['return'] = 0
short_cond = (df['label'] - mid_type) < -epsilon
long_cond = (df['label'] - mid_type) > epsilon
df.loc[long_cond, 'return'] = pct_chg.loc[long_cond]
df.loc[short_cond, 'return'] = -pct_chg[short_cond]
returns = df['return']
if 'Y0' in performance_types:
Y0 = pd.Series(index=pct_chg.index, data=list(y))
result['Y0'] = Y0
if 'Y' in performance_types:
result['Y'] = predict
if 'returns' in performance_types:
result['returns'] = returns
if 'cum_returns' in performance_types:
result['cum_returns'] = empyrical.cum_returns(returns)
if 'annual_return' in performance_types:
result['annual_return'] = empyrical.annual_return(returns)
if 'sharpe_ratio' in performance_types:
result['sharpe_ratio'] = empyrical.sharpe_ratio(returns)
return result
def summarize_paths(samples, cone_std=(1.0, 1.5, 2.0), starting_value=1.0):
"""
Gnerate the upper and lower bounds of an n standard deviation
cone of forecasted cumulative returns.
Parameters
----------
samples : numpy.ndarray
Alternative paths, or series of possible outcomes.
cone_std : list of int/float
Number of standard devations to use in the boundaries of
the cone. If multiple values are passed, cone bounds will
be generated for each value.
Returns
-------
samples : pandas.core.frame.DataFrame
"""
cum_samples = ep.cum_returns(samples.T, starting_value=starting_value).T
cum_mean = cum_samples.mean(axis=0)
cum_std = cum_samples.std(axis=0)
if isinstance(cone_std, (float, int)):
cone_std = [cone_std]
cone_bounds = pd.DataFrame(columns=pd.Float64Index([]))
for num_std in cone_std:
cone_bounds.loc[:, float(num_std)] = cum_mean + cum_std * num_std
cone_bounds.loc[:, float(-num_std)] = cum_mean - cum_std * num_std
return cone_bounds
def compute_consistency_score(returns_test, preds):
"""
Compute Bayesian consistency score.
Parameters
----------
returns_test : pd.Series
Observed cumulative returns.
preds : numpy.array
Multiple (simulated) cumulative returns.
Returns
-------
Consistency score
Score from 100 (returns_test perfectly on the median line of the
Bayesian cone spanned by preds) to 0 (returns_test completely
outside of Bayesian cone.)
"""
returns_test_cum = cum_returns(returns_test, starting_value=1.)
cum_preds = np.cumprod(preds + 1, 1)
q = [
sp.stats.percentileofscore(cum_preds[:, i],
returns_test_cum.iloc[i],
kind='weak')
for i in range(len(returns_test_cum))
]
# normalize to be from 100 (perfect median line) to 0 (completely outside
# of cone)
return 100 - np.abs(50 - np.mean(q)) / .5
def summarize_paths(samples, cone_std=(1., 1.5, 2.), starting_value=1.):
"""
Gnerate the upper and lower bounds of an n standard deviation
cone of forecasted cumulative returns.
Parameters
----------
samples : numpy.ndarray
Alternative paths, or series of possible outcomes.
cone_std : list of int/float
Number of standard devations to use in the boundaries of
the cone. If multiple values are passed, cone bounds will
be generated for each value.
Returns
-------
samples : pandas.core.frame.DataFrame
"""
cum_samples = empyrical.cum_returns(samples.T,
starting_value=starting_value).T
cum_mean = cum_samples.mean(axis=0)
cum_std = cum_samples.std(axis=0)
if isinstance(cone_std, (float, int)):
cone_std = [cone_std]
cone_bounds = pd.DataFrame(columns=pd.Float64Index([]))
for num_std in cone_std:
cone_bounds.loc[:, float(num_std)] = cum_mean + cum_std * num_std
cone_bounds.loc[:, float(-num_std)] = cum_mean - cum_std * num_std
return cone_bounds
def gen_drawdown_table(returns, top=10):
"""
Places top drawdowns in a table.
Parameters
----------
returns : pd.Series
Daily returns of the strategy, noncumulative.
- See full explanation in tears.create_full_tear_sheet.
top : int, optional
The amount of top drawdowns to find (default 10).
Returns
-------
df_drawdowns : pd.DataFrame
Information about top drawdowns.
"""
df_cum = ep.cum_returns(returns, 1.0)
drawdown_periods = get_top_drawdowns(returns, top=top)
df_drawdowns = pd.DataFrame(index=list(range(top)),
columns=[
'Net drawdown in %', 'Peak date',
'Valley date', 'Recovery date', 'Duration'
])
for i, (peak, valley, recovery) in enumerate(drawdown_periods):
if pd.isnull(recovery):
df_drawdowns.loc[i, 'Duration'] = np.nan
else:
df_drawdowns.loc[i, 'Duration'] = len(
pd.date_range(peak, recovery, freq='B'))
# to_pydatetime()疑似是老的API，使用pd.to_datetime替代
# df_drawdowns.loc[i, 'Peak date'] = (peak.to_pydatetime()
# .strftime('%Y-%m-%d'))
# df_drawdowns.loc[i, 'Valley date'] = (valley.to_pydatetime()
# .strftime('%Y-%m-%d'))
# if isinstance(recovery, float):
# df_drawdowns.loc[i, 'Recovery date'] = recovery
# else:
# df_drawdowns.loc[i, 'Recovery date'] = (recovery.to_pydatetime()
# .strftime('%Y-%m-%d'))
df_drawdowns.loc[i, 'Peak date'] = (
pd.to_datetime(peak).strftime('%Y-%m-%d'))
df_drawdowns.loc[i, 'Valley date'] = (
pd.to_datetime(valley).strftime('%Y-%m-%d'))
if isinstance(recovery, float):
df_drawdowns.loc[i, 'Recovery date'] = recovery
else:
df_drawdowns.loc[i, 'Recovery date'] = pd.to_datetime(
peak).strftime('%Y-%m-%d')
df_drawdowns.loc[i, 'Net drawdown in %'] = (
(df_cum.loc[peak] - df_cum.loc[valley]) / df_cum.loc[peak]) * 100
df_drawdowns['Peak date'] = pd.to_datetime(df_drawdowns['Peak date'])
df_drawdowns['Valley date'] = pd.to_datetime(df_drawdowns['Valley date'])
df_drawdowns['Recovery date'] = pd.to_datetime(
df_drawdowns['Recovery date'])
return df_drawdowns
def plot_drawdown_underwater(returns, ax=None, **kwargs):
# Reference from https://github.com/quantopian/pyfolio/blob/master/pyfolio/plotting.py
def percentage(x, pos):
"""
Adds percentage sign to plot ticks.
"""
return '%.0f%%' % x
if ax is None:
ax = plt.gca()
y_axis_formatter = FuncFormatter(percentage)
ax.yaxis.set_major_formatter(FuncFormatter(y_axis_formatter))
df_cum_rets = ep.cum_returns(returns, starting_value=1.0)
running_max = np.maximum.accumulate(df_cum_rets)
underwater = -100 * ((running_max - df_cum_rets) / running_max)
(underwater).plot(ax=ax,
kind='area',
color='coral',
alpha=0.7,
**kwargs)
ax.set_ylabel('Drawdown')
ax.set_title('Underwater plot')
ax.set_xlabel('')
return ax
def get_sec_return_on_date(cls,
start_date,
end_date,
sec_ids,
freq=FreqType.EOD,
field=['close'],
return_type=DfReturnType.DateIndexAndSecIDCol,
is_cumul=False):
"""
:param start_date: str, start date of the query period
:param end_date: str, end date of the query period
:param sec_ids: list of str, sec IDs
:param field: str, filed of data to be queried
:param freq: FreqType
:param return_type
:param is_cumul: return is cumul or not
:return: pd.DataFrame, index = date, col = sec ID
"""
if not w.isconnected():
w.start()
ret = WindMarketDataHandler.get_sec_price_on_date(
start_date, end_date, sec_ids, freq, field, return_type)
ret = ret.pct_change()
if is_cumul:
ret = ret.fillna(0)
ret = cum_returns(ret, starting_value=1.0)
else:
ret = ret.dropna()
return ret
def ensure_cumul_return(func, argname, arg):
if not isinstance(arg, RETURN):
return arg
if arg.type == ReturnType.Cumul:
return arg.data
else:
data_ = cum_returns(arg.data, starting_value=1.0)
return data_
def ensure_cumul_return(func, argname, arg):
if not isinstance(arg, RETURN):
return arg
if arg.type == ReturnType.Cumul:
return arg.data
else:
data_ = cum_returns(arg.data, starting_value=1.0)
return data_
def test_cum_returns(self, returns, starting_value, expected):
cum_returns = empyrical.cum_returns(returns,
starting_value=starting_value)
for i in range(returns.size):
assert_almost_equal(
cum_returns[i],
expected[i],
4)
def cumulative_returns(daily_returns):
"""Cumulative Returns"""
try:
logger.info("Calculating Cumulative Returns...")
cr = empyrical.cum_returns(daily_returns)
return cr
except Exception as exception:
logger.error('Oops! An error Occurred ⚠️')
raise exception
def apply_slippage_penalty(
returns,
txn_daily,
simulate_starting_capital,
backtest_starting_capital,
impact=0.1,
):
"""
Applies quadratic volumeshare slippage model to daily returns based
on the proportion of the observed historical daily bar dollar volume
consumed by the strategy's trades. Scales the size of trades based
on the ratio of the starting capital we wish to test to the starting
capital of the passed backtest data.
Parameters
----------
returns : pd.Series
Time series of daily returns.
txn_daily : pd.Series
Daily transaciton totals, closing price, and daily volume for
each traded name. See price_volume_daily_txns for more details.
simulate_starting_capital : integer
capital at which we want to test
backtest_starting_capital: capital base at which backtest was
origionally run. impact: See Zipline volumeshare slippage model
impact : float
Scales the size of the slippage penalty.
Returns
-------
adj_returns : pd.Series
Slippage penalty adjusted daily returns.
"""
mult = simulate_starting_capital / backtest_starting_capital
simulate_traded_shares = abs(mult * txn_daily.amount)
simulate_traded_dollars = txn_daily.price * simulate_traded_shares
simulate_pct_volume_used = simulate_traded_shares / txn_daily.volume
penalties = (simulate_pct_volume_used**2 * impact *
simulate_traded_dollars)
daily_penalty = penalties.resample("D").sum()
daily_penalty = daily_penalty.reindex(returns.index).fillna(0)
# Since we are scaling the numerator of the penalties linearly
# by capital base, it makes the most sense to scale the denominator
# similarly. In other words, since we aren't applying compounding to
# simulate_traded_shares, we shouldn't apply compounding to pv.
portfolio_value = (
ep.cum_returns(returns, starting_value=backtest_starting_capital) *
mult)
adj_returns = returns - (daily_penalty / portfolio_value)
return adj_returns
def gen_drawdown_table(returns, top=10):
"""
Places top drawdowns in a table.
Parameters
----------
returns : pd.Series
Daily returns of the strategy, noncumulative.
- See full explanation in tears.create_full_tear_sheet.
top : int, optional
The amount of top drawdowns to find (default 10).
Returns
-------
df_drawdowns : pd.DataFrame
Information about top drawdowns.
"""
df_cum = ep.cum_returns(returns, 1.0)
drawdown_periods = get_top_drawdowns(returns, top=top)
df_drawdowns = pd.DataFrame(index=list(range(top)),
columns=[
'Net drawdown in %', 'Peak date',
'Valley date', 'Recovery date', 'Duration'
])
# if there were fewer than the requested number of drawdowns
# (which can happen if an early drawdown never recovers),
# only return the available number of drawdowns (otherwise
# the worst drawdowns plot can mess up the shared X axis)
df_drawdowns = df_drawdowns.dropna(how="all")
for i, (peak, valley, recovery) in enumerate(drawdown_periods):
if pd.isnull(recovery):
df_drawdowns.loc[i, 'Duration'] = np.nan
else:
df_drawdowns.loc[i, 'Duration'] = len(
pd.date_range(peak, recovery, freq='B'))
df_drawdowns.loc[i, 'Peak date'] = (
peak.to_pydatetime().strftime('%Y-%m-%d'))
df_drawdowns.loc[i, 'Valley date'] = (
valley.to_pydatetime().strftime('%Y-%m-%d'))
if isinstance(recovery, float):
df_drawdowns.loc[i, 'Recovery date'] = recovery
else:
df_drawdowns.loc[i, 'Recovery date'] = (
recovery.to_pydatetime().strftime('%Y-%m-%d'))
df_drawdowns.loc[i, 'Net drawdown in %'] = (
(df_cum.loc[peak] - df_cum.loc[valley]) / df_cum.loc[peak]) * 100
df_drawdowns['Peak date'] = pd.to_datetime(df_drawdowns['Peak date'])
df_drawdowns['Valley date'] = pd.to_datetime(df_drawdowns['Valley date'])
df_drawdowns['Recovery date'] = pd.to_datetime(
df_drawdowns['Recovery date'])
return df_drawdowns
def gen_drawdown_table(returns, top=10):
"""
Places top drawdowns in a table.
Parameters
----------
returns : pd.Series
Daily returns of the strategy, noncumulative.
- See full explanation in tears.create_full_tear_sheet.
top : int, optional
The amount of top drawdowns to find (default 10).
Returns
-------
df_drawdowns : pd.DataFrame
Information about top drawdowns.
"""
df_cum = ep.cum_returns(returns, 1.0)
drawdown_periods = get_top_drawdowns(returns, top=top)
df_drawdowns = pd.DataFrame(
index=list(range(top)),
columns=[
"Net drawdown in %",
"Peak date",
"Valley date",
"Recovery date",
"Duration",
],
)
for i, (peak, valley, recovery) in enumerate(drawdown_periods):
if pd.isnull(recovery):
df_drawdowns.loc[i, "Duration"] = np.nan
else:
df_drawdowns.loc[i, "Duration"] = len(
pd.date_range(peak, recovery, freq="B"))
df_drawdowns.loc[i, "Peak date"] = peak.to_pydatetime().strftime(
"%Y-%m-%d")
df_drawdowns.loc[i, "Valley date"] = valley.to_pydatetime().strftime(
"%Y-%m-%d")
if isinstance(recovery, float):
df_drawdowns.loc[i, "Recovery date"] = recovery
else:
df_drawdowns.loc[i, "Recovery date"] = recovery.to_pydatetime(
).strftime("%Y-%m-%d")
df_drawdowns.loc[i, "Net drawdown in %"] = (
(df_cum.loc[peak] - df_cum.loc[valley]) / df_cum.loc[peak]) * 100
df_drawdowns["Peak date"] = pd.to_datetime(df_drawdowns["Peak date"])
df_drawdowns["Valley date"] = pd.to_datetime(df_drawdowns["Valley date"])
df_drawdowns["Recovery date"] = pd.to_datetime(
df_drawdowns["Recovery date"])
return df_drawdowns
def plot_returns(perf_attrib_data, ax=None):
"""
Plot total, specific, and common returns.
Parameters
----------
perf_attrib_data : pd.DataFrame
df with factors, common returns, and specific returns as columns,
and datetimes as index
- Example:
momentum reversal common_returns specific_returns
dt
2017-01-01 0.249087 0.935925 1.185012 1.185012
2017-01-02 -0.003194 -0.400786 -0.403980 -0.403980
ax : matplotlib.axes.Axes
axes on which plots are made. if None, current axes will be used
Returns
-------
ax : matplotlib.axes.Axes
"""
if ax is None:
ax = plt.gca()
returns = perf_attrib_data['total_returns']
specific_returns = perf_attrib_data['specific_returns']
common_returns = perf_attrib_data['common_returns']
ax.plot(ep.cum_returns(returns), color='g', label='Total returns')
ax.plot(ep.cum_returns(specific_returns), color='b',
label='Cumulative specific returns')
ax.plot(ep.cum_returns(common_returns), color='r',
label='Cumulative common returns')
ax.set_title('Time series of cumulative returns')
ax.set_ylabel('Returns')
set_legend_location(ax)
return ax
def gen_drawdown_table(returns, top=10):
"""
Places top drawdowns in a table.
Parameters
----------
returns : pd.Series
Daily returns of the strategy, noncumulative.
- See full explanation in tears.create_full_tear_sheet.
top : int, optional
The amount of top drawdowns to find (default 10).
Returns
-------
df_drawdowns : pd.DataFrame
Information about top drawdowns.
"""
df_cum = empyrical.cum_returns(returns, 1.0)
drawdown_periods = get_top_drawdowns(returns, top=top)
df_drawdowns = pd.DataFrame(index=list(range(top)),
columns=['Net drawdown in %',
'Peak date',
'Valley date',
'Recovery date',
'Duration'])
for i, (peak, valley, recovery) in enumerate(drawdown_periods):
if pd.isnull(recovery):
df_drawdowns.loc[i, 'Duration'] = np.nan
else:
df_drawdowns.loc[i, 'Duration'] = len(pd.date_range(peak,
recovery,
freq='B'))
df_drawdowns.loc[i, 'Peak date'] = (peak.to_pydatetime()
.strftime('%Y-%m-%d'))
df_drawdowns.loc[i, 'Valley date'] = (valley.to_pydatetime()
.strftime('%Y-%m-%d'))
if isinstance(recovery, float):
df_drawdowns.loc[i, 'Recovery date'] = recovery
else:
df_drawdowns.loc[i, 'Recovery date'] = (recovery.to_pydatetime()
.strftime('%Y-%m-%d'))
df_drawdowns.loc[i, 'Net drawdown in %'] = (
(df_cum.loc[peak] - df_cum.loc[valley]) / df_cum.loc[peak]) * 100
df_drawdowns['Peak date'] = pd.to_datetime(df_drawdowns['Peak date'])
df_drawdowns['Valley date'] = pd.to_datetime(df_drawdowns['Valley date'])
df_drawdowns['Recovery date'] = pd.to_datetime(
df_drawdowns['Recovery date'])
return df_drawdowns
def plot_factor_contribution_to_perf(
perf_attrib_data,
ax=None,
title="Cumulative common returns attribution",
):
"""
Plot each factor's contribution to performance.
Parameters
----------
perf_attrib_data : pd.DataFrame
df with factors, common returns, and specific returns as columns,
and datetimes as index
- Example:
momentum reversal common_returns specific_returns
dt
2017-01-01 0.249087 0.935925 1.185012 1.185012
2017-01-02 -0.003194 -0.400786 -0.403980 -0.403980
ax : matplotlib.axes.Axes
axes on which plots are made. if None, current axes will be used
title : str, optional
title of plot
Returns
-------
ax : matplotlib.axes.Axes
"""
if ax is None:
ax = plt.gca()
factors_to_plot = perf_attrib_data.drop(
["total_returns", "common_returns", "tilt_returns", "timing_returns"],
axis="columns",
errors="ignore",
)
factors_cumulative = pd.DataFrame()
for factor in factors_to_plot:
factors_cumulative[factor] = ep.cum_returns(factors_to_plot[factor])
for col in factors_cumulative:
ax.plot(factors_cumulative[col])
ax.axhline(0, color="k")
configure_legend(ax, change_colors=True)
ax.set_ylabel("Cumulative returns by factor")
ax.set_title(title)
return ax
def getAssetReturn(assets, start_date, end_date):
chg_pct = getDailyIndexData(assets,
start_date,
end_date,
"CHG_PCT",
mode=1)
cum_return = dict()
for asset in assets:
cum_return[asset] = empyrical.cum_returns(chg_pct.loc[:, asset],
starting_value=1.0)
cum_return = pd.DataFrame(data=cum_return)
return cum_return
def plot_returns(returns, specific_returns, common_returns, ax=None):
"""
Plot total, specific, and common returns.
Parameters
----------
returns : pd.Series
total returns, indexed by datetime
specific_returns : pd.Series
specific returns, indexed by datetime
commons_returns : pd.Series
common returns, indexed by datetime
ax : matplotlib.axes.Axes
axes on which plots are made. if None, current axes will be used
Returns
-------
ax : matplotlib.axes.Axes
"""
if ax is None:
ax = plt.gca()
ax.plot(ep.cum_returns(returns), color='g', label='Total returns')
ax.plot(ep.cum_returns(specific_returns),
color='b',
label='Cumulative specific returns')
ax.plot(ep.cum_returns(common_returns),
color='r',
label='Cumulative common returns')
ax.set_title('Time Series of cumulative returns')
ax.set_ylabel('Returns')
set_legend_location(ax)
return ax
def apply_slippage_penalty(returns, txn_daily, simulate_starting_capital,
backtest_starting_capital, impact=0.1):
"""
Applies quadratic volumeshare slippage model to daily returns based
on the proportion of the observed historical daily bar dollar volume
consumed by the strategy's trades. Scales the size of trades based
on the ratio of the starting capital we wish to test to the starting
capital of the passed backtest data.
Parameters
----------
returns : pd.Series
Time series of daily returns.
txn_daily : pd.Series
Daily transaciton totals, closing price, and daily volume for
each traded name. See price_volume_daily_txns for more details.
simulate_starting_capital : integer
capital at which we want to test
backtest_starting_capital: capital base at which backtest was
origionally run. impact: See Zipline volumeshare slippage model
impact : float
Scales the size of the slippage penalty.
Returns
-------
adj_returns : pd.Series
Slippage penalty adjusted daily returns.
"""
mult = simulate_starting_capital / backtest_starting_capital
simulate_traded_shares = abs(mult * txn_daily.amount)
simulate_traded_dollars = txn_daily.price * simulate_traded_shares
simulate_pct_volume_used = simulate_traded_shares / txn_daily.volume
penalties = simulate_pct_volume_used**2 \
* impact * simulate_traded_dollars
daily_penalty = penalties.resample('D').sum()
daily_penalty = daily_penalty.reindex(returns.index).fillna(0)
# Since we are scaling the numerator of the penalties linearly
# by capital base, it makes the most sense to scale the denominator
# similarly. In other words, since we aren't applying compounding to
# simulate_traded_shares, we shouldn't apply compounding to pv.
portfolio_value = empyrical.cum_returns(
returns, starting_value=backtest_starting_capital) * mult
adj_returns = returns - (daily_penalty / portfolio_value)
return adj_returns
def get_top_drawdowns(returns, top=10):
"""
Finds top drawdowns, sorted by drawdown amount.
Parameters
----------
returns : pd.Series
Daily returns of the strategy, noncumulative.
- See full explanation in tears.create_full_tear_sheet.
top : int, optional
The amount of top drawdowns to find (default 10).
Returns
-------
drawdowns : list
List of drawdown peaks, valleys, and recoveries. See get_max_drawdown.
"""
returns = returns.copy()
df_cum = ep.cum_returns(returns, 1.0)
running_max = np.maximum.accumulate(df_cum)
underwater = df_cum / running_max - 1
# SVK start
#print('returns', returns)
#print('uw', underwater)
#import pickle as pkl
#pkl.dump([returns, underwater], open('dump.obj', "wb" ))
#print('count', top)
# SVK end
drawdowns = []
for t in range(top):
#print('iter', t) # SVK
peak, valley, recovery = get_max_drawdown_underwater(underwater)
#print(peak, valley, recovery) # SVK
# Slice out draw-down period
if not pd.isnull(recovery):
underwater.drop(underwater[peak:recovery].index[1:-1],
inplace=True)
else:
# drawdown has not ended yet
underwater = underwater.loc[:peak]
drawdowns.append((peak, valley, recovery))
if (len(returns) == 0) or (len(underwater) == 0):
break
return drawdowns
def plot_factor_contribution_to_perf(
perf_attrib_data,
ax=None,
title='Cumulative common returns attribution',
):
"""
Plot each factor's contribution to performance.
Parameters
----------
perf_attrib_data : pd.DataFrame
df with factors, common returns, and specific returns as columns,
and datetimes as index
- Example:
momentum reversal common_returns specific_returns
dt
2017-01-01 0.249087 0.935925 1.185012 1.185012
2017-01-02 -0.003194 -0.400786 -0.403980 -0.403980
ax : matplotlib.axes.Axes
axes on which plots are made. if None, current axes will be used
title : str, optional
title of plot
Returns
-------
ax : matplotlib.axes.Axes
"""
if ax is None:
ax = plt.gca()
factors_to_plot = perf_attrib_data.drop(
['total_returns', 'common_returns'], axis='columns', errors='ignore'
)
factors_cumulative = ep.cum_returns(factors_to_plot)
for col in factors_cumulative:
ax.plot(factors_cumulative[col])
ax.axhline(0, color='k')
configure_legend(ax, change_colors=True)
ax.set_ylabel('Cumulative returns by factor')
ax.set_title(title)
return ax
def get_top_drawdowns(returns, top=10):
"""
Finds top drawdowns, sorted by drawdown amount.
Parameters
----------
returns : pd.Series
Daily returns of the strategy, noncumulative.
- See full explanation in tears.create_full_tear_sheet.
top : int, optional
The amount of top drawdowns to find (default 10).
Returns
-------
drawdowns : list
List of drawdown peaks, valleys, and recoveries. See get_max_drawdown.
"""
returns = returns.copy()
df_cum = empyrical.cum_returns(returns, 1.0)
running_max = np.maximum.accumulate(df_cum)
underwater = df_cum / running_max - 1
drawdowns = []
for t in range(top):
peak, valley, recovery = get_max_drawdown_underwater(underwater)
# Slice out draw-down period
if not pd.isnull(recovery):
underwater.drop(underwater[peak: recovery].index[1:-1],
inplace=True)
else:
# drawdown has not ended yet
underwater = underwater.loc[:peak]
drawdowns.append((peak, valley, recovery))
if (len(returns) == 0) or (len(underwater) == 0):
break
return drawdowns
def cum_returns(returns, starting_value=0):
"""
Compute cumulative returns from simple returns.
Parameters
----------
returns : pd.Series
Daily returns of the strategy, noncumulative.
- See full explanation in tears.create_full_tear_sheet.
starting_value : float, optional
The starting returns (default 1).
Returns
-------
pandas.Series
Series of cumulative returns.
Notes
-----
For increased numerical accuracy, convert input to log returns
where it is possible to sum instead of multiplying.
"""
return empyrical.cum_returns(returns, starting_value=starting_value)
def update(self, dt, algorithm_returns, benchmark_returns, leverage):
# Keep track of latest dt for use in to_dict and other methods
# that report current state.
self.latest_dt = dt
dt_loc = self.cont_index.get_loc(dt)
self.latest_dt_loc = dt_loc
self.algorithm_returns_cont[dt_loc] = algorithm_returns
self.algorithm_returns = self.algorithm_returns_cont[:dt_loc + 1]
self.num_trading_days = len(self.algorithm_returns)
if self.create_first_day_stats:
if len(self.algorithm_returns) == 1:
self.algorithm_returns = np.append(0.0, self.algorithm_returns)
self.algorithm_cumulative_returns[dt_loc] = cum_returns(
self.algorithm_returns
)[-1]
algo_cumulative_returns_to_date = \
self.algorithm_cumulative_returns[:dt_loc + 1]
self.mean_returns_cont[dt_loc] = \
algo_cumulative_returns_to_date[dt_loc] / self.num_trading_days
self.mean_returns = self.mean_returns_cont[:dt_loc + 1]
self.annualized_mean_returns_cont[dt_loc] = \
self.mean_returns_cont[dt_loc] * 252
self.annualized_mean_returns = \
self.annualized_mean_returns_cont[:dt_loc + 1]
if self.create_first_day_stats:
if len(self.mean_returns) == 1:
self.mean_returns = np.append(0.0, self.mean_returns)
self.annualized_mean_returns = np.append(
0.0, self.annualized_mean_returns)
self.benchmark_returns_cont[dt_loc] = benchmark_returns
self.benchmark_returns = self.benchmark_returns_cont[:dt_loc + 1]
if self.create_first_day_stats:
if len(self.benchmark_returns) == 1:
self.benchmark_returns = np.append(0.0, self.benchmark_returns)
self.benchmark_cumulative_returns[dt_loc] = cum_returns(
self.benchmark_returns
)[-1]
benchmark_cumulative_returns_to_date = \
self.benchmark_cumulative_returns[:dt_loc + 1]
self.mean_benchmark_returns_cont[dt_loc] = \
benchmark_cumulative_returns_to_date[dt_loc] / \
self.num_trading_days
self.mean_benchmark_returns = self.mean_benchmark_returns_cont[:dt_loc]
self.annualized_mean_benchmark_returns_cont[dt_loc] = \
self.mean_benchmark_returns_cont[dt_loc] * 252
self.annualized_mean_benchmark_returns = \
self.annualized_mean_benchmark_returns_cont[:dt_loc + 1]
self.algorithm_cumulative_leverages_cont[dt_loc] = leverage
self.algorithm_cumulative_leverages = \
self.algorithm_cumulative_leverages_cont[:dt_loc + 1]
if self.create_first_day_stats:
if len(self.algorithm_cumulative_leverages) == 1:
self.algorithm_cumulative_leverages = np.append(
0.0,
self.algorithm_cumulative_leverages)
if not len(self.algorithm_returns) and len(self.benchmark_returns):
message = "Mismatch between benchmark_returns ({bm_count}) and \
algorithm_returns ({algo_count}) in range {start} : {end} on {dt}"
message = message.format(
bm_count=len(self.benchmark_returns),
algo_count=len(self.algorithm_returns),
start=self.start_session,
end=self.end_session,
dt=dt
)
raise Exception(message)
self.update_current_max()
self.benchmark_volatility[dt_loc] = annual_volatility(
self.benchmark_returns
)
self.algorithm_volatility[dt_loc] = annual_volatility(
self.algorithm_returns
)
# caching the treasury rates for the minutely case is a
# big speedup, because it avoids searching the treasury
# curves on every minute.
# In both minutely and daily, the daily curve is always used.
treasury_end = dt.replace(hour=0, minute=0)
if np.isnan(self.daily_treasury[treasury_end]):
treasury_period_return = choose_treasury(
self.treasury_curves,
self.start_session,
treasury_end,
self.trading_calendar,
)
self.daily_treasury[treasury_end] = treasury_period_return
self.treasury_period_return = self.daily_treasury[treasury_end]
self.excess_returns[dt_loc] = (
self.algorithm_cumulative_returns[dt_loc] -
self.treasury_period_return)
self.alpha[dt_loc], self.beta[dt_loc] = alpha_beta_aligned(
self.algorithm_returns,
self.benchmark_returns,
)
self.sharpe[dt_loc] = sharpe_ratio(
self.algorithm_returns,
)
self.downside_risk[dt_loc] = downside_risk(
self.algorithm_returns
)
self.sortino[dt_loc] = sortino_ratio(
self.algorithm_returns,
_downside_risk=self.downside_risk[dt_loc]
)
self.information[dt_loc] = information_ratio(
self.algorithm_returns,
self.benchmark_returns,
)
self.max_drawdown = max_drawdown(
self.algorithm_returns
)
self.max_drawdowns[dt_loc] = self.max_drawdown
self.max_leverage = self.calculate_max_leverage()
self.max_leverages[dt_loc] = self.max_leverage
def calculate_metrics(self):
self.benchmark_period_returns = \
cum_returns(self.benchmark_returns).iloc[-1]
self.algorithm_period_returns = \
cum_returns(self.algorithm_returns).iloc[-1]
if not self.algorithm_returns.index.equals(
self.benchmark_returns.index
):
message = "Mismatch between benchmark_returns ({bm_count}) and \
algorithm_returns ({algo_count}) in range {start} : {end}"
message = message.format(
bm_count=len(self.benchmark_returns),
algo_count=len(self.algorithm_returns),
start=self._start_session,
end=self._end_session
)
raise Exception(message)
self.num_trading_days = len(self.benchmark_returns)
self.mean_algorithm_returns = (
self.algorithm_returns.cumsum() /
np.arange(1, self.num_trading_days + 1, dtype=np.float64)
)
self.benchmark_volatility = annual_volatility(self.benchmark_returns)
self.algorithm_volatility = annual_volatility(self.algorithm_returns)
self.treasury_period_return = choose_treasury(
self.treasury_curves,
self._start_session,
self._end_session,
self.trading_calendar,
)
self.sharpe = sharpe_ratio(
self.algorithm_returns,
)
# The consumer currently expects a 0.0 value for sharpe in period,
# this differs from cumulative which was np.nan.
# When factoring out the sharpe_ratio, the different return types
# were collapsed into `np.nan`.
# TODO: Either fix consumer to accept `np.nan` or make the
# `sharpe_ratio` return type configurable.
# In the meantime, convert nan values to 0.0
if pd.isnull(self.sharpe):
self.sharpe = 0.0
self.downside_risk = downside_risk(
self.algorithm_returns.values
)
self.sortino = sortino_ratio(
self.algorithm_returns.values,
_downside_risk=self.downside_risk,
)
self.information = information_ratio(
self.algorithm_returns.values,
self.benchmark_returns.values,
)
self.alpha, self.beta = alpha_beta_aligned(
self.algorithm_returns.values,
self.benchmark_returns.values,
)
self.excess_return = self.algorithm_period_returns - \
self.treasury_period_return
self.max_drawdown = max_drawdown(self.algorithm_returns.values)
self.max_leverage = self.calculate_max_leverage()
def risk_metric_period(cls,
start_session,
end_session,
algorithm_returns,
benchmark_returns,
algorithm_leverages):
"""
Creates a dictionary representing the state of the risk report.
Parameters
----------
start_session : pd.Timestamp
Start of period (inclusive) to produce metrics on
end_session : pd.Timestamp
End of period (inclusive) to produce metrics on
algorithm_returns : pd.Series(pd.Timestamp -> float)
Series of algorithm returns as of the end of each session
benchmark_returns : pd.Series(pd.Timestamp -> float)
Series of benchmark returns as of the end of each session
algorithm_leverages : pd.Series(pd.Timestamp -> float)
Series of algorithm leverages as of the end of each session
Returns
-------
risk_metric : dict[str, any]
Dict of metrics that with fields like:
{
'algorithm_period_return': 0.0,
'benchmark_period_return': 0.0,
'treasury_period_return': 0,
'excess_return': 0.0,
'alpha': 0.0,
'beta': 0.0,
'sharpe': 0.0,
'sortino': 0.0,
'period_label': '1970-01',
'trading_days': 0,
'algo_volatility': 0.0,
'benchmark_volatility': 0.0,
'max_drawdown': 0.0,
'max_leverage': 0.0,
}
"""
algorithm_returns = algorithm_returns[
(algorithm_returns.index >= start_session) &
(algorithm_returns.index <= end_session)
]
# Benchmark needs to be masked to the same dates as the algo returns
benchmark_returns = benchmark_returns[
(benchmark_returns.index >= start_session) &
(benchmark_returns.index <= algorithm_returns.index[-1])
]
benchmark_period_returns = ep.cum_returns(benchmark_returns).iloc[-1]
algorithm_period_returns = ep.cum_returns(algorithm_returns).iloc[-1]
alpha, beta = ep.alpha_beta_aligned(
algorithm_returns.values,
benchmark_returns.values,
)
sharpe = ep.sharpe_ratio(algorithm_returns)
# The consumer currently expects a 0.0 value for sharpe in period,
# this differs from cumulative which was np.nan.
# When factoring out the sharpe_ratio, the different return types
# were collapsed into `np.nan`.
# TODO: Either fix consumer to accept `np.nan` or make the
# `sharpe_ratio` return type configurable.
# In the meantime, convert nan values to 0.0
if pd.isnull(sharpe):
sharpe = 0.0
sortino = ep.sortino_ratio(
algorithm_returns.values,
_downside_risk=ep.downside_risk(algorithm_returns.values),
)
rval = {
'algorithm_period_return': algorithm_period_returns,
'benchmark_period_return': benchmark_period_returns,
'treasury_period_return': 0,
'excess_return': algorithm_period_returns,
'alpha': alpha,
'beta': beta,
'sharpe': sharpe,
'sortino': sortino,
'period_label': end_session.strftime("%Y-%m"),
'trading_days': len(benchmark_returns),
'algo_volatility': ep.annual_volatility(algorithm_returns),
'benchmark_volatility': ep.annual_volatility(benchmark_returns),
'max_drawdown': ep.max_drawdown(algorithm_returns.values),
'max_leverage': algorithm_leverages.max(),
}
# check if a field in rval is nan or inf, and replace it with None
# except period_label which is always a str
return {
k: (
None
if k != 'period_label' and not np.isfinite(v) else
v
)
for k, v in iteritems(rval)
}