def test_dlt_map_univariate(make_weekly_data):
train_df, test_df, coef = make_weekly_data
dlt = DLT(
response_col='response',
date_col='week',
seasonality=52,
num_warmup=50,
verbose=False,
estimator='stan-map'
)
dlt.fit(train_df)
init_call = dlt._model.get_init_values()
assert isinstance(init_call, DLTInitializer)
assert init_call.s == 52
init_values = init_call()
assert init_values['init_sea'].shape == (51,)
predict_df = dlt.predict(test_df)
expected_columns = ['week', 'prediction']
expected_shape = (51, len(expected_columns))
expected_num_parameters = 12
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert len(dlt._posterior_samples) == expected_num_parameters
def test_base_dlt_init(estimator):
dlt = DLT(estimator=estimator)
is_fitted = dlt.is_fitted()
model_data_input = dlt.get_training_data_input()
model_param_names = dlt._model.get_model_param_names()
init_values = dlt._model.get_init_values()
# model is not yet fitted
assert not is_fitted
# should only be initialized and not set
assert not model_data_input
# model param names should already be set
assert model_param_names
# callable is not implemented yet
assert not init_values
def test_dlt_aggregated_with_regression(make_weekly_data, estimator, regressor_signs, point_method):
train_df, test_df, coef = make_weekly_data
dlt = DLT(
response_col='response',
date_col='week',
regressor_col=train_df.columns.tolist()[2:],
regressor_sign=regressor_signs,
seasonality=52,
num_warmup=50,
verbose=False,
estimator=estimator
)
dlt.fit(train_df, point_method=point_method)
predict_df = dlt.predict(test_df)
regression_out = dlt.get_regression_coefs()
num_regressors = regression_out.shape[0]
expected_columns = ['week', 'prediction']
expected_shape = (51, len(expected_columns))
expected_regression_shape = (6, 3)
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert regression_out.shape == expected_regression_shape
assert num_regressors == len(train_df.columns.tolist()[2:])
predict_df = dlt.predict(test_df, decompose=True)
assert any(predict_df['regression'].values)
def test_dlt_map_global_trend(make_weekly_data, global_trend_option):
train_df, test_df, coef = make_weekly_data
dlt = DLT(
response_col='response',
date_col='week',
seasonality=52,
global_trend_option=global_trend_option,
estimator='stan-map'
)
dlt.fit(train_df)
predict_df = dlt.predict(test_df)
expected_columns = ['week', 'prediction']
expected_shape = (51, len(expected_columns))
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
def test_dlt_non_seasonal_fit(make_weekly_data, estimator):
train_df, test_df, coef = make_weekly_data
dlt = DLT(
response_col='response',
date_col='week',
estimator=estimator,
num_warmup=50,
)
dlt.fit(train_df)
predict_df = dlt.predict(test_df)
expected_columns = ['week', 'prediction_5', 'prediction', 'prediction_95']
expected_shape = (51, len(expected_columns))
expected_num_parameters = 10
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert len(dlt._posterior_samples) == expected_num_parameters
def test_dlt_fixed_sm_input(make_weekly_data, level_sm_input, seasonality_sm_input, slope_sm_input):
train_df, test_df, coef = make_weekly_data
dlt = DLT(
response_col='response',
date_col='week',
regressor_col=train_df.columns.tolist()[2:],
level_sm_input=level_sm_input,
seasonality_sm_input=seasonality_sm_input,
slope_sm_input=slope_sm_input,
seasonality=52,
num_warmup=50,
verbose=False,
estimator='stan-map'
)
dlt.fit(train_df)
predict_df = dlt.predict(test_df)
regression_out = dlt.get_regression_coefs()
num_regressors = regression_out.shape[0]
expected_columns = ['week', 'prediction']
expected_shape = (51, len(expected_columns))
expected_regression_shape = (6, 3)
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert regression_out.shape == expected_regression_shape
assert num_regressors == len(train_df.columns.tolist()[2:])
def test_dlt_map_reproducibility(make_weekly_data, seasonality):
train_df, test_df, coef = make_weekly_data
dlt1 = DLT(
response_col='response',
date_col='week',
prediction_percentiles=[5, 95],
seasonality=seasonality,
estimator='stan-map'
)
# first fit and predict
dlt1.fit(train_df)
posteriors1 = copy(dlt1._point_posteriors['map'])
prediction1 = dlt1.predict(test_df)
# second fit and predict
# note a new instance must be created to reset the seed
# note both fit and predict contain random generation processes
dlt2 = DLT(
response_col='response',
date_col='week',
prediction_percentiles=[5, 95],
seasonality=seasonality,
estimator='stan-map'
)
dlt2.fit(train_df)
posteriors2 = copy(dlt2._point_posteriors['map'])
prediction2 = dlt2.predict(test_df)
# assert same posterior keys
assert set(posteriors1.keys()) == set(posteriors2.keys())
# assert posterior draws are reproducible
for k, v in posteriors1.items():
assert np.allclose(posteriors1[k], posteriors2[k])
# assert prediction is reproducible
assert np.allclose(prediction1['prediction'].values, prediction2['prediction'].values)
def test_dlt_map_single_regressor(iclaims_training_data):
df = iclaims_training_data
df['claims'] = np.log(df['claims'])
regressor_col = ['trend.unemploy']
dlt = DLT(
response_col='claims',
date_col='week',
regressor_col=regressor_col,
seasonality=52,
seed=8888,
estimator='stan-map'
)
dlt.fit(df)
predicted_df = dlt.predict(df)
expected_num_parameters = 13
expected_columns = ['week', 'prediction']
assert predicted_df.shape[0] == df.shape[0]
assert predicted_df.columns.tolist() == expected_columns
assert len(dlt._posterior_samples) == expected_num_parameters
def test_dlt_is_fitted(iclaims_training_data, estimator, keep_samples, point_method):
df = iclaims_training_data
df['claims'] = np.log(df['claims'])
regressor_col = ['trend.unemploy']
dlt = DLT(
response_col='claims',
date_col='week',
regressor_col=regressor_col,
seasonality=52,
seed=8888,
num_warmup=50,
num_sample=50,
verbose=False,
estimator=estimator
)
dlt.fit(df, keep_samples=keep_samples, point_method=point_method)
is_fitted = dlt.is_fitted()
# still True when keep_samples is False
assert is_fitted
def test_dlt_predict_seed(make_weekly_data, estimator, random_seed):
train_df, test_df, coef = make_weekly_data
args = {
'response_col': 'response',
'date_col': 'week',
'seasonality': 52,
'n_bootstrap_draws': 100,
'verbose': False,
'estimator': estimator,
}
if estimator == 'stan-mcmc':
args.update({'num_warmup': 50, 'num_sample': 100})
elif estimator == 'pyro-svi':
args.update({'num_steps': 10})
lgt = DLT(**args)
lgt.fit(train_df)
predict_df1 = lgt.predict(test_df, seed=random_seed)
predict_df2 = lgt.predict(test_df, seed=random_seed)
assert all(predict_df1['prediction'].values == predict_df2['prediction'].values)
def test_invalid_regressor(make_weekly_data, regressor_signs, invalid_input):
train_df, test_df, coef = make_weekly_data
regressor_col = train_df.columns.tolist()[2:]
# make invalid values
train_df[regressor_col[0]][36] = invalid_input
expected_flag = False
try:
dlt = DLT(
response_col='response',
date_col='week',
regressor_col=regressor_col,
regressor_sign=regressor_signs,
prediction_percentiles=[5, 95],
seasonality=52,
num_warmup=50,
verbose=False,
estimator='stan-map'
)
dlt.fit(train_df)
except ModelException:
expected_flag = True
assert expected_flag
def test_dlt_missing(iclaims_training_data, estimator):
df = iclaims_training_data
missing_idx = np.array([10, 20, 30, 40, 41, 42, 43, 44, df.shape[0] - 1])
df.loc[missing_idx, 'claims'] = np.nan
dlt = DLT(
response_col='claims',
date_col='week',
seasonality=52,
verbose=False,
estimator=estimator
)
dlt.fit(df)
predicted_df = dlt.predict(df)
if estimator == 'stan-map':
expected_columns = ['week', 'prediction']
elif estimator == 'stan-mcmc':
expected_columns = ['week', 'prediction_5', 'prediction', 'prediction_95']
assert all(~np.isnan(predicted_df['prediction']))
assert predicted_df.columns.tolist() == expected_columns
assert predicted_df.shape[0] == df.shape[0]
def test_dlt_grid_tuning(make_weekly_data, param_grid):
train_df, test_df, coef = make_weekly_data
args = {
'response_col': 'response',
'date_col': 'week',
'seasonality': 52,
'estimator': 'stan-map',
}
dlt = DLT(**args)
best_params, tuned_df = grid_search_orbit(param_grid,
model=dlt,
df=train_df,
min_train_len=80, incremental_len=20, forecast_len=20,
metrics=None, criteria=None, verbose=True)
assert best_params[0].keys() == param_grid.keys()
assert set(tuned_df.columns.to_list()) == set(list(param_grid.keys()) + ['metrics'])
assert tuned_df.shape == (9, 3)
def test_plot_predicted_data(iclaims_training_data):
df = iclaims_training_data
df['claims'] = np.log(df['claims'])
regressor_col = ['trend.unemploy', 'trend.filling', 'trend.job']
dlt = DLT(date_col='week',
response_col='claims',
regressor_col=regressor_col,
seasonality=52,
estimator='stan-map')
bt = BackTester(model=dlt,
df=df,
min_train_len=100,
incremental_len=100,
forecast_len=20)
bt.fit_predict()
predicted_df = bt.get_predicted_df()
# test plotting
_ = plot_bt_predictions(predicted_df, metrics=smape, ncol=2, include_vline=True)
def test_dlt_full_with_regression(make_weekly_data, regressor_signs):
train_df, test_df, coef = make_weekly_data
dlt = DLT(
response_col='response',
date_col='week',
regressor_col=train_df.columns.tolist()[2:],
regressor_sign=regressor_signs,
prediction_percentiles=[5, 95],
seasonality=52,
num_warmup=50,
verbose=False,
estimator='stan-mcmc'
)
dlt.fit(train_df)
init_call = dlt._model.get_init_values()
assert isinstance(init_call, DLTInitializer)
init_values = init_call()
assert init_values['init_sea'].shape == (51,)
if regressor_signs.count('+') > 0:
assert init_values['pr_beta'].shape == (regressor_signs.count('+'),)
if regressor_signs.count('-') > 0:
assert init_values['nr_beta'].shape == (regressor_signs.count('-'),)
if regressor_signs.count('=') > 0:
assert init_values['rr_beta'].shape == (regressor_signs.count('='),)
predict_df = dlt.predict(test_df)
regression_out = dlt.get_regression_coefs()
num_regressors = regression_out.shape[0]
expected_columns = ['week', 'prediction_5', 'prediction', 'prediction_95']
expected_shape = (51, len(expected_columns))
expected_regression_shape = (6, 3)
assert predict_df.shape == expected_shape
assert predict_df.columns.tolist() == expected_columns
assert regression_out.shape == expected_regression_shape
assert num_regressors == len(train_df.columns.tolist()[2:])
assert np.sum(regression_out['coefficient'].values >= 0) <= \
regressor_signs.count('+') + regressor_signs.count('=')
assert np.sum(regression_out['coefficient'].values <= 0) <= \
regressor_signs.count('-') + regressor_signs.count('=')
def test_dlt_prediction_percentiles(iclaims_training_data, prediction_percentiles):
df = iclaims_training_data
dlt = DLT(
response_col='claims',
date_col='week',
seasonality=52,
num_warmup=50,
num_sample=50,
seed=8888,
prediction_percentiles=prediction_percentiles,
estimator='stan-mcmc'
)
if not prediction_percentiles:
p_labels = ['_5', '', '_95']
else:
p_labels = ['_5', '_10', '', '_95']
dlt.fit(df)
predicted_df = dlt.predict(df)
expected_columns = ['week'] + ["prediction" + p for p in p_labels]
assert predicted_df.columns.tolist() == expected_columns
assert predicted_df.shape[0] == df.shape[0]
predicted_df = dlt.predict(df, decompose=True)
predicted_components = [
'prediction',
PredictionKeys.TREND.value,
PredictionKeys.SEASONALITY.value,
PredictionKeys.REGRESSION.value]
expected_columns = ['week']
for pc in predicted_components:
for p in p_labels:
expected_columns.append(pc + p)
assert predicted_df.columns.tolist() == expected_columns
assert predicted_df.shape[0] == df.shape[0]
def test_dlt_full_reproducibility(make_weekly_data, estimator, regressor_signs, seasonality):
train_df, test_df, coef = make_weekly_data
dlt_first = DLT(
response_col='response',
date_col='week',
regressor_col=train_df.columns.tolist()[2:],
regressor_sign=regressor_signs,
prediction_percentiles=[5, 95],
seasonality=seasonality,
num_warmup=50,
verbose=False,
estimator=estimator
)
# first fit and predict
dlt_first.fit(train_df)
posteriors_first = copy(dlt_first._posterior_samples)
predict_df_first = dlt_first.predict(test_df)
regression_out_first = dlt_first.get_regression_coefs()
# second fit and predict
# note a new instance must be created to reset the seed
# note both fit and predict contain random generation processes
dlt_second = DLT(
response_col='response',
date_col='week',
regressor_col=train_df.columns.tolist()[2:],
regressor_sign=regressor_signs,
prediction_percentiles=[5, 95],
seasonality=seasonality,
num_warmup=50,
verbose=False,
estimator=estimator
)
dlt_second.fit(train_df)
posteriors_second = copy(dlt_second._posterior_samples)
predict_df_second = dlt_second.predict(test_df)
regression_out_second = dlt_second.get_regression_coefs()
# assert same posterior keys
assert set(posteriors_first.keys()) == set(posteriors_second.keys())
# assert posterior draws are reproducible
for k, v in posteriors_first.items():
assert np.allclose(posteriors_first[k], posteriors_second[k])
# assert identical regression columns
# this is also checked in posterior samples, but an extra layer just in case
# since this one very commonly retrieved by end users
assert regression_out_first.equals(regression_out_second)
# assert prediction is reproducible
assert predict_df_first.equals(predict_df_second)
def test_dlt_mixed_signs_and_order(iclaims_training_data, regressor_signs):
df = iclaims_training_data
df['claims'] = np.log(df['claims'])
raw_regressor_col = ['trend.unemploy', 'trend.filling', 'trend.job']
new_regressor_col = [raw_regressor_col[idx] for idx in [1, 2, 0]]
new_regressor_signs = [regressor_signs[idx] for idx in [1, 2, 0]]
# mixiing ordering of cols in df of prediction
new_df = df[['claims', 'week'] + new_regressor_col]
dlt = DLT(
response_col='claims',
date_col='week',
regressor_col=raw_regressor_col,
regressor_sign=regressor_signs,
seasonality=52,
seed=8888,
estimator='stan-map'
)
dlt.fit(df)
predicted_df_v1 = dlt.predict(df)
predicted_df_v2 = dlt.predict(new_df)
# mixing ordering of signs
dlt_new = DLT(
response_col='claims',
date_col='week',
regressor_col=new_regressor_col,
regressor_sign=new_regressor_signs,
seasonality=52,
seed=8888,
estimator='stan-map'
)
dlt_new.fit(df)
predicted_df_v3 = dlt_new.predict(df)
predicted_df_v4 = dlt_new.predict(new_df)
pred_v1 = predicted_df_v1['prediction'].values
pred_v2 = predicted_df_v2['prediction'].values
pred_v3 = predicted_df_v3['prediction'].values
pred_v4 = predicted_df_v4['prediction'].values
# they should be all identical; ordering of signs or columns in prediction show not matter
assert np.allclose(pred_v1, pred_v2, atol=1e-2)
assert np.allclose(pred_v1, pred_v3, atol=1e-2)
assert np.allclose(pred_v1, pred_v4, atol=1e-2)