def testDefaultsExist(self):
"""Check defaults are accessible."""
# Summon the defaults.
col_names = semantics.DataFrameNameMapping()
groups = semantics.GroupSemantics()
periods = semantics.PeriodSemantics()
# Check default col names available.
self.assertIsInstance(col_names.geo, str)
self.assertIsInstance(col_names.group, str)
self.assertIsInstance(col_names.period, str)
self.assertIsInstance(col_names.response, str)
self.assertIsInstance(col_names.cost, str)
self.assertIsInstance(col_names.incr_response, str)
self.assertIsInstance(col_names.incr_cost, str)
# Check default data semantics available.
self.assertIsInstance(groups.control, int)
self.assertIsInstance(groups.treatment, int)
self.assertIsInstance(groups.unassigned, int)
self.assertIsInstance(periods.pre, int)
self.assertIsInstance(periods.test, int)
self.assertIsInstance(periods.cooldown, int)
self.assertIsInstance(periods.unassigned, int)
def fit(self, data_frame, target, **kwargs):
"""Fit the TBR model to the supplied data frame.
See optional kwargs for interpretation of the data frame.
Args:
data_frame: a pandas.DataFrame. Should contain the columns and indices
corresponding to the **kwargs information below. Only one of response
or cost need be present, corresponding to the supplied `target`. Must be
indexed by date.
target: `str`. The name of the column to be analysed.
**kwargs: optional column/index names for the data and related semantics:
key_geo='geo' - geo data frame index name.
key_period='period' - experimental period column name.
key_group='group' - group assignment column name.
key_cost='cost' - cost column name.
key_response='response' - response column name.
key_date='date' - date index name.
key_incr_cost='_incr_cost' - incremental cost column name.
key_incr_response='_incr_response' - incremental response column name.
group_control=1 - value representing the control group in the data.
group_treatment=2 - value representing the treatment group in the data.
period_pre=0 - value representing the pre-test period in the data.
period_test=1 - value representing the test period in the data.
period_cool=2 - value representing the cooldown period in the data.
"""
# Set the target of the analysis.
self.target = target
# Extract any column / index name information supplied by the user.
user_df_names = utils.kwarg_subdict('key_', **kwargs)
self.df_names = semantics.DataFrameNameMapping(**user_df_names)
# Extract any semantics for control / treatment supplied by user.
user_group_semantics = utils.kwarg_subdict('group_', **kwargs)
self.groups = semantics.GroupSemantics(**user_group_semantics)
# Extract any semantics for experimental period supplied by user.
user_period_semantics = utils.kwarg_subdict('period_', **kwargs)
self.periods = semantics.PeriodSemantics(**user_period_semantics)
# Set up the analysis data.
self._construct_analysis_data(data_frame)
# Fit pre-period models for response and for cost.
self._fit_pre_period_model()
def setUp(self):
"""This method will be run before each of the test methods in the class."""
super(TBRiROASTest, self).setUp()
# Load the salesandcost dataset.
csv_path = 'matched_markets/csv/'
csv_dir = os.path.join("", csv_path)
self.data = salesandcost.example_data_formatted(csv_dir)
# Data frame names for the salesandcost example.
self.key_response = 'sales'
self.key_cost = 'cost'
self.key_group = 'geo.group'
self.key_period = 'period'
self.key_geo = 'geo'
self.key_date = 'date'
# Semantics for groups and periods.
self.groups = semantics.GroupSemantics()
self.periods = semantics.PeriodSemantics()
def fit(self, data_frame, **kwargs):
"""Fit the TBRiROAS model to the supplied data frame.
See optional kwargs for interpretation of the data frame.
Args:
data_frame: a pandas.DataFrame. Should contain the columns and indices
corresponding to the **kwargs information below. Must be indexed by date.
**kwargs: optional column/index names for the data and related semantics:
key_geo='geo' - geo data frame index name.
key_period='period' - experimental period column name.
key_group='group' - group assignment column name.
key_cost='cost' - cost column name.
key_response='response' - response column name.
key_date='date' - date index name.
key_incr_cost='_incr_cost' - incremental cost column name.
key_incr_response='_incr_response' - incremental response column name.
group_control=1 - value representing the control group in the data.
group_treat=2 - value representing the treatment group in the data.
period_pre=0 - value representing the pre-test period in the data.
period_test=1 - value representing the test period in the data.
period_cool=2 - value representing the cooldown period in the data.
"""
# Extract any column / index name information supplied by the user
user_df_names = utils.kwarg_subdict('key_', **kwargs)
self.df_names = semantics.DataFrameNameMapping(**user_df_names)
# Extract any semantics for control / treatment supplied by user
user_group_semantics = utils.kwarg_subdict('group_', **kwargs)
self.groups = semantics.GroupSemantics(**user_group_semantics)
# Extract any semantics for experimental period supplied by user
user_period_semantics = utils.kwarg_subdict('period_', **kwargs)
self.periods = semantics.PeriodSemantics(**user_period_semantics)
# Fit seprate TBR models for response and cost
self.tbr_response.fit(data_frame, self.df_names.response, **kwargs)
self.tbr_cost.fit(data_frame, self.df_names.cost, **kwargs)
def __init__(
self,
n_control,
n_treat,
time_pre,
time_test, # no cooldown as yet
hetresp,
hetcost,
beta,
hetsked,
sig_resp,
sig_cost,
noise_treat_only=False,
seed=None,
**kwargs):
"""Creates a data simulator.
Args:
n_control: int. The number of control geos.
n_treat: int. The number of treatment geos.
time_pre: int. The number of pre-test period ticks.
time_test: int. The number of test period ticks.
hetresp: float. The degree of mean response variable heterogeneity.
hetcost: float. The degree of mean cost variable heterogeneity.
beta: float. The iROAS coefficient to be used.
hetsked: float. The degree of heteroskedasticity in cost and response.
sig_resp: float. The noise level in the response variable.
sig_cost: float. The noise level in the cost variable.
noise_treat_only: bool. Whether to add noise only in the treatment period.
seed: int. Sets the seed of the random number generator.
**kwargs: optional sematics for the produced data frame.
"""
# Constants.
self.n_control = n_control
self.n_treat = n_treat
self.time_pre = time_pre
self.time_test = time_test
self.time_total = time_pre + time_test
# Model parameters.
self.hetresp = hetresp
self.hetcost = hetcost
self.beta = beta
self.hetsked = hetsked
self.sig_resp = sig_resp
self.sig_cost = sig_cost
# Derived facts.
self.n_total = self.n_treat + self.n_control
self.col_len = self.n_total * self.time_total
# Extract any column / index name information supplied by the user.
user_df_names = utils.kwarg_subdict('key_', **kwargs)
self._df_names = semantics.DataFrameNameMapping(**user_df_names)
# Options
self.noise_treat_only = noise_treat_only
# Extract any semantics for control / treatment supplied by user.
user_group_semantics = utils.kwarg_subdict('group_', **kwargs)
self._groups = semantics.GroupSemantics(**user_group_semantics)
# Extract any semantics for experimental period supplied by user.
user_period_semantics = utils.kwarg_subdict('period_', **kwargs)
self._periods = semantics.PeriodSemantics(**user_period_semantics)
if seed is None:
seed = np.random.randint(0, 2**32)
self._rng = np.random.RandomState(seed=seed)
def fit(self, data_frame, target=None, **kwargs):
"""Runs the TBR diagnostics suite.
This method executes the following diagnostics: (1) detect and remove the
disrupted geos; (2) detect and remove the outlier time points (3)
correlation test and (4) the structural stability (A/A) test removing part
of the pre-test period. The results of these diagnostics are stored in the
_test_results attribute. The resulting modified data frame is stored in the
_data attribute and accessible via the get_data() method.
Note. This method makes a copy of the original data_frame, and it doesn't
modify the original.
See optional kwargs for interpretation of the data frame.
Args:
data_frame: (pandas.DataFrame) Should contain the columns and indices
corresponding to the **kwargs information below. Only one of response
need be present, corresponding to the supplied `target`. Must be
indexed by date.
target: (str) name of the target metric (data frame column). If not
specified, the column specified as key_response will be assumed.
**kwargs: optional column/index names for the data and related semantics:
key_geo (string) column name for geo (default: 'geo').
key_period (string) column name for period (default: 'period').
key_group (string) column name for group (default: 'group').
key_response (string) response column name (default: 'response').
key_date (string) date index name (default: 'date').
group_control (int) control group id (default: 1).
group_treat (int) treatment group id (default: 2).
period_pre (int) pre-test period id (default: 0).
period_test (int) test period id (default: 1).
period_cool (int) cooldown period id (default: 2).
"""
self._data = data_frame.copy()
user_df_names = utils.kwarg_subdict('key_', **kwargs)
self._df_names = semantics.DataFrameNameMapping(**user_df_names)
user_group_semantics = utils.kwarg_subdict('group_', **kwargs)
self._groups = semantics.GroupSemantics(**user_group_semantics)
user_period_semantics = utils.kwarg_subdict('period_', **kwargs)
self._periods = semantics.PeriodSemantics(**user_period_semantics)
if target is None:
target = self._df_names.response
self._target = target
remove_geos = self._detect_noisy_geos(iqr_coef=1.5, max_threshold=0.5)
self._diagnostics['noisy_geos'] = remove_geos
if remove_geos:
exclude = self._data[self._df_names.geo].isin(remove_geos)
self._data = self._data[~exclude]
self._create_analysis_data()
remove_dates = self._detect_outliers(max_prob=0.1)
self._diagnostics['outlier_dates'] = remove_dates
if remove_dates:
exclude_dates = self._data[self._df_names.date].isin(remove_dates)
self._data = self._data[~exclude_dates]
self._create_analysis_data()
self._diagnostics['corr_test'] = self._correlation_test(
min_cor=0.5, prefer_cor=0.8, credible_level=0.95)
