def test_prob_alive_is_close_to_Hardie_paper_table_6(self, donations):
"""Table 6: P(Alive in 2002) as a Function of Recency and Frequency"""
bbtf = estimation.BetaGeoBetaBinomFitter()
bbtf.fit(
donations['frequency'],
donations['recency'],
donations['periods'],
donations['weights'],
)
bbtf.data['prob_alive'] = bbtf.conditional_probability_alive(
1, donations['frequency'], donations['recency'],
donations['periods'])
# Expected probabilities for last year 1995-0 repeat, 1999-2 repeat, 2001-6 repeat
expected = np.array([0.11, 0.59, 0.93])
prob_list = np.zeros(3)
prob_list[0] = (bbtf.data[(bbtf.data['frequency'] == 0)
& (bbtf.data['recency'] == 0)]['prob_alive'])
prob_list[1] = (bbtf.data[(bbtf.data['frequency'] == 2)
& (bbtf.data['recency'] == 4)]['prob_alive'])
prob_list[2] = (bbtf.data[(bbtf.data['frequency'] == 6)
& (bbtf.data['recency'] == 6)]['prob_alive'])
npt.assert_array_almost_equal(expected, prob_list, decimal=2)
def test_fit_with_index(self, donations):
bbtf = estimation.BetaGeoBetaBinomFitter()
index = range(len(donations), 0, -1)
bbtf.fit(donations['frequency'],
donations['recency'],
donations['periods'],
donations['weights'],
index=index)
assert (bbtf.data.index == index).all() == True
bbtf = estimation.BetaGeoBetaBinomFitter()
bbtf.fit(donations['frequency'],
donations['recency'],
donations['periods'],
donations['weights'],
index=None)
assert (bbtf.data.index == index).all() == False
def test_params_out_is_close_to_Hardie_paper(self):
bbtf = estimation.BetaGeoBetaBinomFitter()
bbtf.fit(
donations['frequency'],
donations['recency'],
donations['n'],
donations['n_custs'],
)
expected = np.array([1.204, 0.750, 0.657, 2.783])
npt.assert_array_almost_equal(expected, np.array(bbtf._unload_params('alpha','beta','gamma','delta')),
decimal=2)
def test_expected_purchases_in_n_periods_returns_same_value_as_Hardie_excel_sheet(self):
"""Total expected from Hardie's In-Sample Fit sheet."""
bbtf = estimation.BetaGeoBetaBinomFitter()
bbtf.fit(
donations['frequency'],
donations['recency'],
donations['n'],
donations['n_custs'],
)
expected = np.array([3454.9, 1253.1]) # Cells C18 and C24
estimated = bbtf.expected_number_of_transactions_in_first_n_periods(6).loc[[0,6]].values.flatten()
npt.assert_almost_equal(expected, estimated, decimal=0)
def test_params_same_from_sim_data(self, bbgb_params):
sim_data = beta_geometric_beta_binom_model(N=6, size=100000, **bbgb_params)
bbtf = estimation.BetaGeoBetaBinomFitter()
grouped_data = sim_data.groupby(['frequency', 'recency', 'n_periods'])['customer_id'].count()
grouped_data = grouped_data.reset_index().rename(columns={'customer_id': 'weights'})
bbtf.fit(grouped_data['frequency'],
grouped_data['recency'],
grouped_data['n_periods'],
grouped_data['weights'])
npt.assert_allclose(
np.asarray(list(bbgb_params.values())).astype(float),
np.asarray(bbtf._unload_params('alpha', 'beta', 'gamma', 'delta')).astype(float),
atol=0.1, rtol=1e-2)
def test_fit_with_and_without_weights(self, donations):
exploded_dataset = pd.DataFrame(
columns=['frequency', 'recency', 'periods'])
for _, row in donations.iterrows():
exploded_dataset = exploded_dataset.append(
pd.DataFrame(
[[row['frequency'], row['recency'], row['periods']]] *
row['weights'],
columns=['frequency', 'recency', 'periods']))
exploded_dataset = exploded_dataset.astype(np.int64)
assert exploded_dataset.shape[0] == donations['weights'].sum()
bbtf_noweights = estimation.BetaGeoBetaBinomFitter()
bbtf_noweights.fit(
exploded_dataset['frequency'],
exploded_dataset['recency'],
exploded_dataset['periods'],
)
bbtf = estimation.BetaGeoBetaBinomFitter()
bbtf.fit(
donations['frequency'],
donations['recency'],
donations['periods'],
donations['weights'],
)
npt.assert_array_almost_equal(
np.array(
bbtf_noweights._unload_params('alpha', 'beta', 'gamma',
'delta')),
np.array(bbtf._unload_params('alpha', 'beta', 'gamma', 'delta')),
decimal=4)
def test_params_same_from_sim_data(self):
sim_data = beta_geometric_beta_binom_model(N=6,
size=100000,
**bbgb_params)
bbtf = estimation.BetaGeoBetaBinomFitter()
grouped_data = sim_data.groupby(['frequency', 'recency',
'n'])['customer_id'].count()
grouped_data = grouped_data.reset_index().rename(
columns={'customer_id': 'n_custs'})
bbtf.fit(grouped_data['frequency'], grouped_data['recency'],
grouped_data['n'], grouped_data['n_custs'])
assert (
(np.array(list(bbgb_params.values())) -
np.array(bbtf._unload_params('alpha', 'beta', 'gamma', 'delta')))
< 0.1).all()
def test_conditional_expectation_returns_same_value_as_Hardie_excel_sheet(self):
"""
Total from Hardie's Conditional Expectations (II) sheet.
http://brucehardie.com/notes/010/BGBB_2011-01-20_XLSX.zip
"""
bbtf = estimation.BetaGeoBetaBinomFitter()
bbtf.fit(
donations['frequency'],
donations['recency'],
donations['n'],
donations['n_custs'],
)
pred_purchases = bbtf.conditional_expected_number_of_purchases_up_to_time(5) * donations['n_custs']
expected = 12884.2 # Sum of column F Exp Tot
npt.assert_almost_equal(expected, pred_purchases.sum(), decimal=0)