def _run_with_data(self, data, pre_period, post_period, model_args, alpha,
estimation):
# Zoom in on data in modeling range
if data.shape[1] == 1: # no exogenous values provided
raise ValueError("data contains no exogenous variables")
data_modeling = data.copy()
df_pre = data_modeling.loc[pre_period[0]:pre_period[1], :]
df_post = data_modeling.loc[post_period[0]:post_period[1], :]
# Standardize all variables
orig_std_params = (0, 1)
if model_args["standardize_data"]:
sd_results = standardize_all_variables(data_modeling, pre_period,
post_period)
df_pre = sd_results["data_pre"]
df_post = sd_results["data_post"]
orig_std_params = sd_results["orig_std_params"]
# Construct model and perform inference
model = construct_model(df_pre, model_args)
self.model = model
trained_model = model_fit(model, estimation, model_args["niter"])
self.model = trained_model
inferences = compile_posterior_inferences(trained_model, data, df_pre,
df_post, None, alpha,
orig_std_params, estimation)
# "append" to 'CausalImpact' object
self.inferences = inferences["series"]
def _run_with_data(self, data, pre_period, post_period, model_args, alpha,
estimation):
# Zoom in on data in modeling range
if data.shape[1] == 1: # no exogenous values provided
raise ValueError("data contains no exogenous variables")
non_null = pd.isnull(data.iloc[:, 1]).to_numpy().nonzero()
first_non_null = non_null[0]
if first_non_null.size > 0:
pre_period[0] = max(pre_period[0], data.index[first_non_null[0]])
data_modeling = data.copy()
df_pre = data_modeling.loc[pre_period[0]:pre_period[1], :]
df_post = data_modeling.loc[post_period[0]:post_period[1], :]
# Standardize all variables
orig_std_params = (0, 1)
if model_args["standardize_data"]:
sd_results = standardize_all_variables(data_modeling, pre_period,
post_period)
df_pre = sd_results["data_pre"]
df_post = sd_results["data_post"]
orig_std_params = sd_results["orig_std_params"]
# Construct model and perform inference
ucm_model = construct_model(self, df_pre, model_args)
res = model_fit(self, ucm_model, estimation, model_args["niter"])
inferences = compile_posterior_inferences(res, data, df_pre, df_post, None,
alpha, orig_std_params,
estimation)
# "append" to 'CausalImpact' object
self.inferences = inferences["series"]
self.model = ucm_model
def _run_with_ucm(self, ucm_model, post_period_response, alpha, model_args,
estimation):
""" Runs an impact analysis on top of a ucm model.
Args:
ucm_model: Model as returned by UnobservedComponents(),
in which the data during the post-period was set to NA
post_period_response: observed data during the post-intervention
period
alpha: tail-probabilities of posterior intervals"""
# Guess <pre_period> and <post_period> from the observation vector
# These will be needed for plotting period boundaries in plot().
#raise NotImplementedError()
"""
try:
indices = infer_period_indices_from_data(y)
except ValueError:
raise ValueError("ucm_model must have been fitted on data where " +
"the values in the post-intervention period " +
"have been set to NA")
"""
df_pre = ucm_model.data.orig_endog[:-len(post_period_response)]
df_pre = pd.DataFrame(df_pre)
post_period_response = pd.DataFrame(post_period_response)
data = pd.DataFrame(
np.concatenate([df_pre.values, post_period_response.values]))
orig_std_params = (0, 1)
fitted_model = model_fit(ucm_model, estimation, model_args["niter"])
# Compile posterior inferences
inferences = compile_posterior_inferences(fitted_model, data, df_pre,
None, post_period_response,
alpha, orig_std_params,
estimation)
obs_inter = pre_len = fitted_model.model.nobs - len(
post_period_response)
self.params["pre_period"] = [0, obs_inter - 1]
self.params["post_period"] = [obs_inter, -1]
self.data = pd.concat([df_pre, post_period_response])
self.inferences = inferences["series"]
self.model = fitted_model