def test_fit_execute_multi_target(self, data):
f1, target, model = self._create_base_components()
target1 = QuarterlyTarget(data_key='quarterly',
col='marketcap',
quarter_shift=-1)
pipeline = Pipeline(data=data,
feature=f1,
target=[target, target1],
model=model,
out_name=None)
res = pipeline.fit(tickers, metric=median_absolute_relative_error)
assert type(res) == dict
assert res['metric_y_0'] > 0
assert res['metric_y_1'] > 0
df = pipeline.execute(tickers)
assert type(df) == pd.DataFrame
assert df['y_0'].mean() > 0
assert df['y_1'].mean() > 0
assert (df['y_0'] == df['y_1']).min() == False
pipeline = Pipeline(data=data,
feature=f1,
target=[target, target],
model=model,
out_name=None)
res = pipeline.fit(tickers, metric=median_absolute_relative_error)
assert type(res) == dict
assert res['metric_y_0'] > 0
assert res['metric_y_1'] > 0
df = pipeline.execute(tickers)
assert (df['y_0'] == df['y_1']).min() == True
def FairMarketcapDiffSF1(pretrained=True) -> Pipeline:
'''
Model is used to evaluate quarter-to-quarter(q2q) company
fundamental progress. Model uses
:class:`~ml_investment.features.QuarterlyDiffFeatures`
(q2q results progress, e.g. 30% revenue increase,
decrease in debt by 15% etc),
:class:`~ml_investment.features.BaseCompanyFeatures`,
:class:`~ml_investment.features.QuarterlyFeatures`
:class:`~ml_investment.features.CommoditiesAggQuarterFeatures`
and trying to predict real q2q marketcap difference(
:class:`~ml_investment.targets.QuarterlyDiffTarget` ).
So model prediction may be interpreted as "fair" marketcap
change according this q2q fundamental change.
:mod:`~ml_investment.data_loaders.sf1`
is used for loading data.
Note:
SF1 dataset is paid, so for using this model you need to subscribe
and paste quandl token to `~/.ml_investment/secrets.json`
``quandl_api_key``
Parameters
----------
pretrained:
use pretreined weights or not. If so,
`fair_marketcap_diff_sf1.pickle` will be downloaded.
Downloading directory path can be changed in
`~/.ml_investment/config.json` ``models_path``
'''
_check_download_data()
data = _create_data()
feature = _create_feature()
target = _create_target()
model = _create_model()
pipeline = Pipeline(feature=feature,
target=target,
model=model,
data=data,
out_name=OUT_NAME)
core_path = '{}/{}.pickle'.format(config['models_path'], OUT_NAME)
if pretrained:
if not os.path.exists(core_path):
urlretrieve(URL, core_path)
pipeline.load_core(core_path)
return pipeline
def test_fit_execute_simple(self, data):
f1, target, model = self._create_base_components()
pipeline = Pipeline(data=data,
feature=f1,
target=target,
model=model,
out_name=None)
res = pipeline.fit(tickers, metric=median_absolute_relative_error)
assert type(res) == dict
assert res['metric_y_0'] > 0
df = pipeline.execute(tickers)
assert type(df) == pd.DataFrame
assert df['y_0'].mean() > 0
def FairMarketcapSF1(pretrained=True) -> Pipeline:
'''
Model is used to estimate fair company marketcap for several last quarters.
Pipeline uses features from
:class:`~ml_investment.features.BaseCompanyFeatures`,
:class:`~ml_investment.features.QuarterlyFeatures`,
:class:`~ml_investment.features.DailyAggQuarterFeatures`,
:class:`~ml_investment.features.CommoditiesAggQuarterFeatures`
and trained to predict real market capitalizations
( using :class:`~ml_investment.targets.QuarterlyTarget` ).
Since some companies are overvalued and some are undervalued,
the model makes an average "fair" prediction.
:mod:`~ml_investment.data_loaders.sf1` and
:mod:`~ml_investment.data_loaders.quandl_commodities`
is used for loading data.
Note:
SF1 dataset is paid, so for using this model you need to subscribe
and paste quandl token to `~/.ml_investment/secrets.json`
``quandl_api_key``
Parameters
----------
pretrained:
use pretreined weights or not. If so, `fair_marketcap_sf1.pickle`
will be downloaded. Downloading directory path can be changed in
`~/.ml_investment/config.json` ``models_path``
'''
_check_download_data()
data = _create_data()
feature = _create_feature()
target = _create_target()
model = _create_model()
pipeline = Pipeline(feature=feature,
target=target,
model=model,
data=data,
out_name=OUT_NAME)
core_path = '{}/{}.pickle'.format(config['models_path'], OUT_NAME)
if pretrained:
if not os.path.exists(core_path):
urlretrieve(URL, core_path)
pipeline.load_core(core_path)
return pipeline
def test_fit_execute_multi_names(self, data):
f1, target, model = self._create_base_components()
pipeline = Pipeline(data=data,
feature=f1,
target=[target, target],
model=model,
out_name=['name1', 'name2'])
res = pipeline.fit(tickers, metric=median_absolute_relative_error)
assert type(res) == dict
assert res['metric_name1'] > 0
assert res['metric_name2'] > 0
df = pipeline.execute(tickers)
assert type(df) == pd.DataFrame
assert df['name1'].mean() > 0
assert df['name2'].mean() > 0
assert (df['name1'] == df['name2']).min() == True
def FairMarketcapDiffYahoo(pretrained=True) -> Pipeline:
'''
Model is used to evaluate quarter-to-quarter(q2q) company
fundamental progress. Model uses
:class:`~ml_investment.features.QuarterlyDiffFeatures`
(q2q results progress, e.g. 30% revenue increase,
decrease in debt by 15% etc),
:class:`~ml_investment.features.BaseCompanyFeatures`,
:class:`~ml_investment.features.QuarterlyFeatures`
and trying to predict smoothed real q2q marketcap difference(
:class:`~ml_investment.targets.DailySmoothedQuarterlyDiffTarget` ).
So model prediction may be interpreted as "fair" marketcap
change according this q2q fundamental change.
:mod:`~ml_investment.data_loaders.yahoo` and
:mod:`~ml_investment.data_loaders.daily_bars`
are used for loading data.
Parameters
----------
pretrained:
use pretreined weights or not. If so,
`fair_marketcap_diff_yahoo.pickle` will be downloaded.
Downloading directory path can be changed in
`~/.ml_investment/config.json` ``models_path``
'''
_check_download_data()
data = _create_data()
feature = _create_feature()
target = _create_target()
model = _create_model()
pipeline = Pipeline(feature=feature,
target=target,
model=model,
data=data,
out_name=OUT_NAME)
core_path = '{}/{}.pickle'.format(config['models_path'], OUT_NAME)
if pretrained:
if not os.path.exists(core_path):
print('Downloading pretrained model')
urlretrieve(URL, core_path)
pipeline.load_core(core_path)
return pipeline
def test_export_load(self, data, tmpdir):
f1, target, model = self._create_base_components()
pipeline = Pipeline(data=data,
feature=f1,
target=target,
model=model,
out_name=None)
res = pipeline.fit(tickers, metric=median_absolute_relative_error)
df = pipeline.execute(tickers)
pipeline.export_core('{}/pipeline'.format(str(tmpdir)))
pipeline.load_core('{}/pipeline.pickle'.format(str(tmpdir)))
df1 = pipeline.execute(tickers)
np.testing.assert_array_equal(df['y_0'].values, df1['y_0'].values)
def MarketcapDownStdSF1(pretrained=True) -> Pipeline:
'''
Model is used to predict future down-std value.
Pipeline consist of time-series model training(
:class:`~ml_investment.models.TimeSeriesOOFModel` )
and validation on real marketcap down-std values(
:class:`~ml_investment.targets.DailyAggTarget` ).
Model prediction may be interpreted as "risk" for the next quarter.
:mod:`~ml_investment.data_loaders.sf1`
is used for loading data.
Note:
SF1 dataset is paid, so for using this model you need to subscribe
and paste quandl token to `~/.ml_investment/secrets.json`
``quandl_api_key``
Parameters
----------
pretrained:
use pretreined weights or not. If so,
`marketcap_down_std_sf1.pickle` will be downloaded.
Downloading directory path can be changed in
`~/.ml_investment/config.json` ``models_path``
'''
_check_download_data()
data = _create_data()
feature = _create_feature()
target = _create_target()
model = _create_model()
pipeline = Pipeline(feature=feature,
target=target,
model=model,
data=data,
out_name=OUT_NAME)
core_path = '{}/{}.pickle'.format(config['models_path'], OUT_NAME)
if pretrained:
if not os.path.exists(core_path):
urlretrieve(URL, core_path)
pipeline.load_core(core_path)
return pipeline
def test_fit_execute_multi_target_metric(self, data):
f1, target, model = self._create_base_components()
target1 = QuarterlyTarget(data_key='quarterly',
col='marketcap',
quarter_shift=-1)
pipeline = Pipeline(data=data,
feature=f1,
target=[target, target1],
model=model,
out_name=None)
res = pipeline.fit(tickers,
metric=[
median_absolute_relative_error,
mean_absolute_relative_error
])
assert type(res) == dict
assert res['metric_y_0'] > 0
assert res['metric_y_1'] > 0
assert res['metric_y_0'] < res['metric_y_1']
def FairMarketcapYahoo(pretrained=True) -> Pipeline:
'''
Model is used to estimate fair company marketcap for `last` quarter.
Pipeline uses features from
:class:`~ml_investment.features.BaseCompanyFeatures`,
:class:`~ml_investment.features.QuarterlyFeatures`
and trained to predict real market capitalizations
( using :class:`~ml_investment.targets.QuarterlyTarget` ).
Since some companies are overvalued and some are undervalued,
the model makes an average "fair" prediction.
:mod:`~ml_investment.data_loaders.yahoo`
is used for loading data.
Parameters
----------
pretrained:
use pretreined weights or not. If so, `fair_marketcap_yahoo.pickle`
will be downloaded. Downloading directory path can be changed in
`~/.ml_investment/config.json` ``models_path``
'''
_check_download_data()
data = _create_data()
feature = _create_feature()
target = _create_target()
model = _create_model()
pipeline = Pipeline(feature=feature,
target=target,
model=model,
data=data,
out_name=OUT_NAME)
core_path = '{}/{}.pickle'.format(config['models_path'], OUT_NAME)
if pretrained:
if not os.path.exists(core_path):
print('Downloading pretrained model')
urlretrieve(URL, core_path)
pipeline.load_core(core_path)
return pipeline
def test_fit_execute_multi_target_model(self, data):
f1, target, model = self._create_base_components()
target1 = QuarterlyTarget(data_key='quarterly',
col='marketcap',
quarter_shift=-1)
model1 = GroupedOOFModel(ctb.CatBoostRegressor(verbose=False),
group_column='ticker',
fold_cnt=4)
pipeline = Pipeline(data=data,
feature=f1,
target=[target, target],
model=[model, model1],
out_name=None)
res = pipeline.fit(tickers, metric=median_absolute_relative_error)
assert type(res) == dict
assert res['metric_y_0'] > 0
assert res['metric_y_1'] > 0
df = pipeline.execute(tickers)
assert type(df) == pd.DataFrame
assert df['y_0'].mean() > 0
assert df['y_1'].mean() > 0
assert (df['y_0'] == df['y_1']).min() == False
def test_fit_execute_simple(self, data):
columns = ['revenue', 'netinc', 'ncf', 'ebitda', 'debt', 'fcf']
f1 = QuarterlyFeatures(data_key='quarterly',
columns=columns,
quarter_counts=[2, 10],
max_back_quarter=1)
target1 = QuarterlyTarget(data_key='quarterly',
col='marketcap',
quarter_shift=0)
target2 = QuarterlyTarget(data_key='quarterly',
col='marketcap',
quarter_shift=-1)
model = lgbm.sklearn.LGBMRegressor()
pipeline1 = Pipeline(data=data,
feature=f1,
target=target1,
model=model,
out_name='p1')
pipeline2 = Pipeline(data=data,
feature=f1,
target=target2,
model=model,
out_name='p2')
pipeline3 = LoadingPipeline(data['quarterly'],
['ticker', 'date', 'marketcap'])
merge1 = MergePipeline(pipeline_list=[pipeline1, pipeline2, pipeline3],
execute_merge_on=['ticker', 'date'])
merge1.fit(tickers)
df_m1 = merge1.execute(tickers)
pipeline1.fit(tickers)
pipeline2.fit(tickers)
merge2 = MergePipeline(pipeline_list=[pipeline1, pipeline2, pipeline3],
execute_merge_on=['ticker', 'date'])
df1 = pipeline1.execute(tickers)
df2 = pipeline2.execute(tickers)
df3 = pipeline3.execute(tickers)
df_m2 = merge1.execute(tickers)
df_m3 = merge1.execute(tickers, 2)
assert type(df_m1) == pd.DataFrame
assert type(df_m2) == pd.DataFrame
assert len(df_m1) == len(df1)
assert len(df_m2) == len(df1)
np.testing.assert_array_equal(
df_m1.columns, ['ticker', 'date', 'p1', 'p2', 'marketcap'])
np.testing.assert_array_equal(
df_m2.columns, ['ticker', 'date', 'p1', 'p2', 'marketcap'])
np.testing.assert_array_equal(
df_m3.columns, ['ticker', 'date', 'p1', 'p2', 'marketcap'])
np.testing.assert_array_equal(df1['p1'], df_m1['p1'])
np.testing.assert_array_equal(df2['p2'], df_m1['p2'])
np.testing.assert_array_equal(df_m1['p1'], df_m2['p1'])
np.testing.assert_array_equal(df_m1['p2'], df_m2['p2'])
np.testing.assert_array_equal(df_m2['p1'], df_m3['p1'])
np.testing.assert_array_equal(df_m2['p2'], df_m3['p2'])