def test_get_causal_impact_results_statsmodels_or_tensorflow():
"""
tests get_causal_impact_results function with the statsmodels
or the tensorflow implementation (depends on which one is installed)
"""
if PYTHON_IMPLEMENTATION_USED == "pycausalimpact":
ci_implementation = "statsmodels"
print("Using pycausalimpact (statsmodels implementation)")
elif PYTHON_IMPLEMENTATION_USED == "tfcausalimpact":
ci_implementation = "tensorflow"
print("Using tfcausalimpact (tensorflow implementation)")
else:
raise ValueError(
"Neither tfcausalimpact nor pycausalmodel seem to be installed")
data = generate_test_dataset()
test_experiment_config = ExperimentConfig(**dict(
data=data,
experiment_name="experiment1",
correlation_metric="mutual_info",
time_var="t",
y_var="y",
training_period=[0, 69],
evaluation_period=[70, 99],
percentile_correlation_cut=0.5,
max_number_of_columns=4,
))
most_relevant_vars = get_most_relevant_variables(test_experiment_config)
# TODO: Don't add time_var? Is the index relevant?
data = data[[test_experiment_config.y_var] + most_relevant_vars]
data = impute_missing_values(data, impute_x_vars=True, impute_y_vars=True)
experiment_output = get_causal_impact_results(
processed_data=data,
ci_implementation=ci_implementation,
experiment_config=test_experiment_config,
)
results = experiment_output.results_summary
cumulative_abs_effects = results.loc["abs_effect", "cumulative"]
assert (
190 > cumulative_abs_effects > 100
), f"Cumulative absolute results {cumulative_abs_effects} aren't what is expected!"
def test_passing_causal_model_config_to_get_causal_impact_results_from_tensorflow(
):
data = generate_test_dataset()
custom_alpha = 0.02
causal_model_config = TFCausalImpactConfig(**dict(
model=None,
alpha=custom_alpha,
# Model args dictionary
prior_level_sd=0.6,
fit_method="vi",
nseasons=30, # monthly seasonality.
season_duration=7,
))
test_experiment_config = ExperimentConfig(**dict(
data=data,
experiment_name="experiment1",
correlation_metric="mutual_info",
time_var="t",
y_var="y",
training_period=[0, 69],
evaluation_period=[70, 99],
percentile_correlation_cut=0.5,
max_number_of_columns=4,
causal_model_config=causal_model_config,
))
most_relevant_vars = get_most_relevant_variables(test_experiment_config)
data = data[[test_experiment_config.y_var] + most_relevant_vars]
data = impute_missing_values(data, impute_x_vars=True, impute_y_vars=True)
experiment_output = get_causal_impact_results(
processed_data=data,
ci_implementation="tensorflow",
experiment_config=test_experiment_config,
)
results = experiment_output.results_summary
cumulative_abs_effects = results.loc["abs_effect", "cumulative"]
assert (experiment_output.trained_model.alpha == custom_alpha
), f"Alpha should be {custom_alpha}"
assert (
190 > cumulative_abs_effects > 100
), f"Cumulative absolute results {cumulative_abs_effects} aren't what is expected!"
def test_get_most_relevant_variables():
np.random.seed(12345)
df = generate_test_dataset()
test_experiment_config = ExperimentConfig(**dict(
data=df,
experiment_name="experiment1",
correlation_metric="mutual_info",
time_var="t",
y_var="y",
training_period=[0, 69],
evaluation_period=[70, 99],
percentile_correlation_cut=0.5,
max_number_of_columns=4,
))
most_relevant_columns = get_most_relevant_variables(test_experiment_config)
assert set(["x1", "x2"]) == set(
most_relevant_columns), f"got {set(most_relevant_columns)} instead"
def test_get_results_from_list_of_experiments():
np.random.seed(134)
data = generate_test_dataset()
# Generate List of experiment configs
# In this case, data is shared, but it doesn't have to be
experiment_config1 = ExperimentConfig(**dict(
data=data,
experiment_name="experiment1",
correlation_metric="mutual_info",
time_var="t",
y_var="y",
training_period=[0, 69],
evaluation_period=[70, 99],
percentile_correlation_cut=0.5,
max_number_of_columns=4,
available_x_columns=["x1", "x2"],
))
experiment_config2 = ExperimentConfig(**dict(
data=data,
experiment_name="experiment2",
correlation_metric="mutual_info",
time_var="t",
y_var="x2",
training_period=[0, 60],
evaluation_period=[61, 99],
percentile_correlation_cut=0.5,
max_number_of_columns=4,
available_x_columns=["x1", "x3"],
))
experiment_configs = [experiment_config1, experiment_config2]
all_experiments_results = get_results_from_list_of_experiments(
experiment_configs, ci_implementation="tensorflow")
cumulative_abs_effects = all_experiments_results[
"experiment1"].results_summary.loc["abs_effect", "cumulative"]
assert (
190 > cumulative_abs_effects > 100
), f"Cumulative absolute results {cumulative_abs_effects} aren't what is expected!"
def test_get_causal_impact_results_r():
"""
tests get_causal_impact_results function with R implementation
"""
ci_implementation = "r"
data = generate_test_dataset()
test_experiment_config = ExperimentConfig(**dict(
data=data,
experiment_name="experiment1",
correlation_metric="mutual_info",
time_var="t",
y_var="y",
training_period=[0, 69],
evaluation_period=[70, 99],
percentile_correlation_cut=0.5,
max_number_of_columns=4,
))
most_relevant_vars = get_most_relevant_variables(test_experiment_config)
# TODO: Don't add time_var? Is the index relevant?
data = data[
[test_experiment_config.time_var, test_experiment_config.y_var] +
most_relevant_vars]
data = impute_missing_values(data, impute_x_vars=True, impute_y_vars=True)
experiment_output = get_causal_impact_results(
processed_data=data,
ci_implementation=ci_implementation,
experiment_config=test_experiment_config,
)
results = experiment_output.results_summary
cumulative_abs_effects = results.loc["abs_effect", "cumulative"]
assert (
190 > cumulative_abs_effects > 100
), f"Cumulative absolute results {cumulative_abs_effects} aren't what is expected!"
#########
#
# Example of package usage
#
########
import CausalImpactExplainer as cie
import numpy as np
from CausalImpactExplainer.utils import generate_test_dataset
np.random.seed(134)
data = generate_test_dataset()
test_experiment_config = cie.ExperimentConfig(**dict(
data=data,
experiment_name="experiment1",
correlation_metric="mutual_info",
time_var="t",
y_var="y",
training_period=[0, 69],
evaluation_period=[70, 99],
percentile_correlation_cut=0.5,
max_number_of_columns=4,
))
most_relevant_columns = cie.get_most_relevant_variables(test_experiment_config)
print(most_relevant_columns)
data = data[[test_experiment_config.y_var] + most_relevant_columns]
data = cie.impute_missing_values(data, impute_x_vars=True, impute_y_vars=True)