def test_ktrlite_dual_seas(make_daily_data, seasonality_fs_order):
train_df, _, _ = make_daily_data
ktrlite = KTRLite(
response_col='response',
date_col='date',
seasonality=[7, 365.25],
seasonality_fs_order=seasonality_fs_order,
estimator='stan-map',
n_bootstrap_draws=-1,
)
ktrlite.fit(train_df)
predict_df = ktrlite.predict(train_df)
expected_columns = ['date', 'prediction']
expected_shape = (train_df.shape[0], len(expected_columns))
expected_num_parameters = 6
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert len(ktrlite._posterior_samples) == expected_num_parameters
smape_val = smape(train_df['response'].values,
predict_df['prediction'].values)
assert smape_val <= SMAPE_TOLERANCE
def test_ktrlite_predict_decompose(make_daily_data):
train_df, test_df, coef = make_daily_data
ktrlite = KTRLite(
response_col='response',
date_col='date',
seasonality=[7, 365.25],
seasonality_fs_order=[2, 5],
estimator='stan-map',
n_bootstrap_draws=1e4,
)
ktrlite.fit(train_df)
predict_df = ktrlite.predict(test_df, decompose=True)
expected_columns = [
'date', 'prediction_5', 'prediction', 'prediction_95', 'trend_5',
'trend', 'trend_95', 'seasonality_7_5', 'seasonality_7',
'seasonality_7_95', 'seasonality_365.25_5', 'seasonality_365.25',
'seasonality_365.25_95'
]
expected_shape = (364, len(expected_columns))
expected_num_parameters = 6
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert len(ktrlite._posterior_samples) == expected_num_parameters
def test_ktrlite_level_segments(make_daily_data, level_segments):
train_df, test_df, coef = make_daily_data
ktrlite = KTRLite(
response_col='response',
date_col='date',
level_segments=level_segments,
estimator='stan-map',
n_bootstrap_draws=-1,
)
ktrlite.fit(train_df)
predict_df = ktrlite.predict(test_df)
expected_columns = ['date', 'prediction']
expected_shape = (364, len(expected_columns))
expected_num_parameters = 4
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert len(ktrlite._posterior_samples) == expected_num_parameters
knots_df = ktrlite.get_level_knots()
levels_df = ktrlite.get_levels()
assert knots_df.shape[0] in [level_segments + 1, level_segments + 2]
assert levels_df.shape[0] == ktrlite.get_training_meta()['num_of_obs']
def test_ktrlite_level_knot_distance(make_daily_data, level_knot_distance):
train_df, test_df, coef = make_daily_data
ktrlite = KTRLite(
response_col='response',
date_col='date',
level_knot_distance=level_knot_distance,
estimator='stan-map',
n_bootstrap_draws=1e4,
)
ktrlite.fit(train_df)
predict_df = ktrlite.predict(test_df)
expected_columns = ['date', 'prediction_5', 'prediction', 'prediction_95']
expected_shape = (364, len(expected_columns))
expected_num_parameters = 4
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert len(ktrlite._posterior_samples) == expected_num_parameters
def test_ktrlite_hourly_data(ca_hourly_electricity_data):
train_df, test_df = ca_hourly_electricity_data
ktrlite = KTRLite(
response_col='SDGE',
date_col='Dates',
seasonality=[24, 7, 365.25],
seasonality_fs_order=[3, 3, 5],
estimator='stan-map',
n_bootstrap_draws=-1,
)
ktrlite.fit(train_df)
predict_df = ktrlite.predict(train_df)
expected_columns = ['Dates', 'prediction']
expected_shape = (train_df.shape[0], len(expected_columns))
expected_num_parameters = 6
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert len(ktrlite._posterior_samples) == expected_num_parameters
smape_val = smape(train_df['SDGE'].values, predict_df['prediction'].values)
assert smape_val <= SMAPE_TOLERANCE
def test_ktrlite_seas_segments(make_daily_data, seas_segments):
train_df, test_df, coef = make_daily_data
ktrlite = KTRLite(
response_col='response',
date_col='date',
seasonality=[7, 365.25],
seasonality_fs_order=[2, 5],
level_segments=10,
seasonality_segments=seas_segments,
estimator='stan-map',
n_bootstrap_draws=-1,
)
ktrlite.fit(train_df)
predict_df = ktrlite.predict(test_df)
expected_columns = ['date', 'prediction']
expected_shape = (364, len(expected_columns))
expected_num_parameters = 6
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert len(ktrlite._posterior_samples) == expected_num_parameters
def test_backtester_ktr_and_missing_val(make_daily_data, missing_flag):
train_df, test_df, _ = make_daily_data
df = pd.concat([train_df, test_df], axis=0, ignore_index=True)
if missing_flag:
# create a missing value in testing
df.loc[df.shape[0] - 3, 'response'] = np.nan
# create a missing value in training
df.loc[10, 'response'] = np.nan
ktr = KTRLite(date_col='date',
response_col='response',
seasonality=[365.25],
verbose=False)
bt = BackTester(
model=ktr,
df=df,
n_splits=3,
incremental_len=100,
forecast_len=20,
)
bt.fit_predict()
predicted_df = bt.get_predicted_df()
assert set(predicted_df['split_key'].tolist()) == {0, 1, 2}
bt_score_df = bt.score(include_training_metrics=False)
num_testing_metrics = 6
expected_shape = (num_testing_metrics, 3)
assert bt_score_df.shape == expected_shape
testing_metrics_df = bt_score_df[~bt_score_df['is_training_metric']]
# rmsse is the only one not working for null values; otherwise, they should have valid values
if missing_flag:
metric_vals = testing_metrics_df.loc[
testing_metrics_df['metric_name'] != 'rmsse',
'metric_values'].values
assert np.all(~np.isnan(metric_vals))
missing_metric_val = testing_metrics_df.loc[
testing_metrics_df['metric_name'] == 'rmsse',
'metric_values'].values
assert np.all(np.isnan(missing_metric_val))
else:
metric_vals = testing_metrics_df['metric_values'].values
assert np.all(~np.isnan(metric_vals))