def test_ma_fit_durbin_levinson(self):
np.random.seed(12345)
simulated_ma1 = simulate_arma(theta=[0], sigma=0.5, simulations=10000)
ma1 = ARMA(simulated_ma1)
ma1.fit_ma(q=1, method="durbin_levinson")
self.assertAlmostEqual(ma1.model.get_theta(1), 0, 1)
self.assertAlmostEqual(ma1.model.get_sigma_sq(), 0.25, 1)
def test_likelihood_kalman(self):
ma_model = PureARMA(theta=[0.5])
ma_data = [0, 1]
ma_ts = ARMA(ma_data, subtract_mean=False)
likelihood = (2 * np.pi) ** -1 * np.exp(-0.5)
self.assertAlmostEqual(ma_ts.get_likelihood(model=ma_model, method="kalman"), likelihood)
def test_turning_point_test(self):
data = np.array([1, 0.5, 0, 1, 0.4, 0.2, 3])
seq = ARMA(data)
model = PureARMA()
turning_test_result = abs(10 / 3 - 3) / ((16 * 7 - 29) / 90)
self.assertEqual(seq.turning_point_test(model=model), turning_test_result)
self.assertEqual(seq.turning_point_test(residuals=data), turning_test_result)
def test_difference_sign_test(self):
data = np.array([1, 0.5, 0, 1, 0.4, 0.2, 3])
seq = ARMA(data)
model = PureARMA()
diff_sign_result = 1 / (8 / 12)
self.assertEqual(seq.difference_sign_test(model=model), diff_sign_result)
self.assertEqual(seq.difference_sign_test(residuals=data), diff_sign_result)
def test_ar_fit_durbin_levinson(self):
np.random.seed(12345)
simulated_ar3 = simulate_arma(phi=[0.5, 0.1, 0.2], sigma=0.5)
ar3 = ARMA(simulated_ar3)
ar3.fit_ar(p=3, method="durbin_levinson")
self.assertAlmostEqual(ar3.model.get_phi(1), 0.5, 1)
self.assertAlmostEqual(ar3.model.get_phi(2), 0.1, 1)
self.assertAlmostEqual(ar3.model.get_phi(3), 0.2, 1)
self.assertAlmostEqual(ar3.model.get_sigma_sq(), 0.25, 1)
def test_one_step_kalman(self):
ma_model = PureARMA(theta=[0.5], sigma_sq=2)
ma_data = [1, 1.5, 0.5, -2]
ma = ARMA(ma_data, subtract_mean=False)
zeros = np.zeros(20)
arma_model = PureARMA(phi=[0.2, -0.6, 0.3], theta=[0.3, -0.4], sigma_sq=0.25)
arma = ARMA(zeros)
ar_data = [1, 2, 4]
ar_model = PureARMA(phi=[0.5, -0.5], sigma_sq=2)
ar = ARMA(ar_data, subtract_mean=False)
for k in range(21):
self.assertEqual(arma.get_one_step_predictor(k, arma_model, method="kalman"), 0)
self.assertAlmostEqual(ma.get_one_step_predictor(0, ma_model, method="kalman"), 0)
self.assertAlmostEqual(ma.get_one_step_predictor(1, ma_model, method="kalman"), 0.5)
self.assertAlmostEqual(ma.get_one_step_predictor(2, ma_model, method="kalman"), 0.5)
self.assertAlmostEqual(ma.get_one_step_predictor(3, ma_model, method="kalman"), 0)
self.assertAlmostEqual(ma.get_one_step_predictor(4, ma_model, method="kalman"), -1)
self.assertAlmostEqual(ar.get_one_step_predictor(0, ar_model, method="kalman"), 0)
self.assertAlmostEqual(ar.get_one_step_predictor(1, ar_model, method="kalman"), 0.5)
self.assertAlmostEqual(ar.get_one_step_predictor(2, ar_model, method="kalman"), 0.5)
self.assertAlmostEqual(ar.get_one_step_predictor(3, ar_model, method="kalman"), 1)
def test_weighted_sum_squared_residuals(self):
data = [-2, 0, 2]
arma = ARMA(data)
model = PureARMA()
ma_model = PureARMA(theta=[-0.9], sigma_sq=2)
ma_data = [-2.58, 1.62, -0.96, 2.62, -1.36]
ma = ARMA(ma_data, subtract_mean=False)
self.assertAlmostEqual(arma.get_weighted_sum_squared_residuals(model), 8)
self.assertAlmostEqual(ma.get_weighted_sum_squared_residuals(ma_model), 8.02, 1)
def test_aicc_methods(self):
np.random.seed(12345)
simulated_arma = simulate_arma(phi=[0.2, 0.5], theta=[0.2], simulations=50)
arma = ARMA(simulated_arma)
arma_model = PureARMA(phi=[0.2, 0.5], theta=[0.2], sigma_sq=0.25)
aicc_kalman = arma.get_aicc(model=arma_model, method="kalman")
aicc_innovations = arma.get_aicc(model=arma_model, method="innovations_algo")
# test <0.5% difference between methods
self.assertTrue(abs(aicc_innovations - aicc_kalman) / aicc_innovations < 0.005)
def test_loglikelihood(self):
model_arma11 = PureARMA([0.2], [0.4], sigma_sq=1)
zeros = np.zeros(6)
non_zeros = [0, 0, 0, 0, 0, 1]
zero_ts = ARMA(zeros)
non_zero_ts = ARMA(non_zeros, subtract_mean=False)
loglikelihood_zeros = np.log((2 * np.pi) ** -3 * (1.375 * 1.0436 * 1.0067 * 1.0011 * 1.0002) ** -0.5)
loglikelihood_non_zeros = np.log(np.exp(-0.5 * (1 / 1.0002))) + loglikelihood_zeros
self.assertAlmostEqual(zero_ts.get_loglikelihood(model_arma11), loglikelihood_zeros, 4)
self.assertAlmostEqual(non_zero_ts.get_loglikelihood(model_arma11), loglikelihood_non_zeros, 3)
def test_aic_innovation(self):
model_arma11 = PureARMA([0.2], [0.4], sigma_sq=1)
zeros = np.zeros(6)
non_zeros = [0, 0, 0, 0, 0, 1]
zero_ts = ARMA(zeros)
non_zero_ts = ARMA(non_zeros, subtract_mean=False)
loglikelihood_zeros = np.log((2 * np.pi) ** -3 * (1.375 * 1.0436 * 1.0067 * 1.0011 * 1.0002) ** -0.5)
loglikelihood_non_zeros = np.log(np.exp(-0.5 * (1 / 1.0002))) + loglikelihood_zeros
aicc_zeros = -2 * loglikelihood_zeros + 6
aicc_nonzeros = -2 * loglikelihood_non_zeros + 6
self.assertAlmostEqual(zero_ts.get_aic(model_arma11, method="innovations_algo"), aicc_zeros, 4)
self.assertAlmostEqual(non_zero_ts.get_aic(model_arma11, method="innovations_algo"), aicc_nonzeros, 3)
def test_transform_mean(self):
data = [-1, 0.5, 0, -0.5, 1]
empty = []
mean_zero = ARMA([1, -1])
pos_mean = ARMA([2])
neg_fraction_mean = ARMA([-0.5])
self.assertListEqual(mean_zero._transform(data).tolist(), data)
self.assertListEqual(pos_mean._transform(data).tolist(), [-3, -1.5, -2, -2.5, -1])
self.assertListEqual(neg_fraction_mean._transform(data).tolist(), [-0.5, 1, 0.5, 0, 1.5])
self.assertListEqual(mean_zero._transform(empty).tolist(), empty)
self.assertListEqual(pos_mean._transform(empty).tolist(), empty)
self.assertListEqual(neg_fraction_mean._transform(empty).tolist(), empty)
def test_FPE(self):
short_data = np.zeros(10)
long_data = np.ones(100000)
low_var_model = PureARMA(phi=[0.3, 0.2], sigma_sq=0.1)
high_var_model = PureARMA(phi=[0.1, 0.4, -0.4], sigma_sq=5)
noise_model = PureARMA(sigma_sq=2)
short_ts = ARMA(short_data)
long_ts = ARMA(long_data)
short_ts.model = low_var_model
long_ts.model = low_var_model
self.assertAlmostEqual(short_ts.get_fpe(), 0.1 * 12 / 8)
self.assertAlmostEqual(short_ts.get_fpe(model=high_var_model), 5 * 13 / 7)
self.assertAlmostEqual(short_ts.get_fpe(model=noise_model), 2)
self.assertAlmostEqual(long_ts.get_fpe(), 0.1 * 100002 / 99998)
self.assertAlmostEqual(long_ts.get_fpe(model=high_var_model), 5 * 100003 / 99997)
self.assertAlmostEqual(long_ts.get_fpe(model=noise_model), 2)
def test_likelihood_inno(self):
ma_model = PureARMA(theta=[-0.9], sigma_sq=1)
ma_data = [-2.58, 1.62, -0.96, 2.62, -1.36]
ma = ARMA(ma_data, subtract_mean=False)
model_arma11 = PureARMA([0.2], [0.4], sigma_sq=1)
zeros = np.zeros(6)
non_zeros = [0, 0, 0, 0, 0, 1]
zero_ts = ARMA(zeros)
non_zero_ts = ARMA(non_zeros, subtract_mean=False)
likelihood_zeros = (2 * np.pi) ** -3 * (1.375 * 1.0436 * 1.0067 * 1.0011 * 1.0002) ** -0.5
likelihood_non_zeros = likelihood_zeros * np.exp(-0.5 * (1 / 1.0002))
# self.assertAlmostEqual(ma.get_likelihood(ma_model), 0.0035943790355147075, 4)
self.assertAlmostEqual(zero_ts.get_likelihood(model_arma11, method="innovations_algo"), likelihood_zeros)
self.assertAlmostEqual(
non_zero_ts.get_likelihood(model_arma11, method="innovations_algo"), likelihood_non_zeros, 6
)
def test_sample_covariance_matrix(self):
integers = [1, 2, 3, 4, 5]
integer_ts = ARMA(integers)
cov_mat_1 = np.matrix([2])
cov_mat_2 = np.matrix([[2, 0.8], [0.8, 2]])
cov_mat_3 = np.matrix([[2, 0.8, -0.2], [0.8, 2, 0.8], [-0.2, 0.8, 2]])
cov_mat_4 = np.matrix([[2, 0.8, -0.2, -0.8], [0.8, 2, 0.8, -0.2], [-0.2, 0.8, 2, 0.8], [-0.8, -0.2, 0.8, 2]])
cov_mat_5 = np.matrix(
[
[2, 0.8, -0.2, -0.8, -0.8],
[0.8, 2, 0.8, -0.2, -0.8],
[-0.2, 0.8, 2, 0.8, -0.2],
[-0.8, -0.2, 0.8, 2, 0.8],
[-0.8, -0.8, -0.2, 0.8, 2],
]
)
np.testing.assert_almost_equal(integer_ts.sample_covariance_matrix(1), cov_mat_1)
np.testing.assert_almost_equal(integer_ts.sample_covariance_matrix(2), cov_mat_2)
np.testing.assert_almost_equal(integer_ts.sample_covariance_matrix(3), cov_mat_3)
np.testing.assert_almost_equal(integer_ts.sample_covariance_matrix(4), cov_mat_4)
np.testing.assert_almost_equal(integer_ts.sample_covariance_matrix(5), cov_mat_5)
def test_sample_acf(self):
integers = [1, 2, 3, 4, 5]
integer_ts = ARMA(integers)
self.assertAlmostEqual(integer_ts.sample_acf(0), 1, 10)
self.assertAlmostEqual(integer_ts.sample_acf(-1), 0.4, 10)
self.assertAlmostEqual(integer_ts.sample_acf(2), -0.1, 10)
self.assertAlmostEqual(integer_ts.sample_acf(-3), -0.4, 10)
self.assertAlmostEqual(integer_ts.sample_acf(4), -0.4, 10)
def ran_a_test():
a = ARMA([-0.5, 0.11], [], 0.5, "uni")
time_series = [a.generater.next() for i in range(2000)]
p = ArPredicter(len(a.alphas), max_x = max(time_series),)
errors = []
for index, x in enumerate(time_series):
if index < p.min_ob:
p.predict_and_fit(x)
elif random.random() > missing_percent:
rec_x = p.predict_and_fit(x)
errors.append(rec_x - x)
else:
p.predict_and_fit('*')
mse = sum(map(lambda x: x**2, errors)) / len(errors)
result['errors'].append(errors)
result['mses'].append(mse)
def run_a_test():
#a = ARMA([0.3, -0.4, 0.4, -0.5, 0.6], [], 0.3, noise_type)
a = ARMA([-0.5, 0.11], [], 0.5, noise_type)
time_series = [a.generater.next() for i in range(2000)]
if_missing = [
1 if random.random() > missing_rate or time_series.index(t) < 6
else 0 for t in time_series
]
time_series = [
t if flag == 1 else 0 for t, flag in zip(time_series, if_missing)
]
missing_indexs = [
index for index, flag in enumerate(if_missing) if flag == 0
]
p = ArPredicter(a.p, time_series, missing_indexs)
result['errors'].append(p.errors)
result['mses'].append(p.mse)
def main(_):
flags_obj = tf.flags.FLAGS
euler_graph = tf_euler.dataset.get_dataset(flags_obj.dataset)
euler_graph.load_graph()
dims = [flags_obj.hidden_dim, flags_obj.hidden_dim]
if flags_obj.run_mode == 'train':
metapath = [euler_graph.train_edge_type]
else:
metapath = [euler_graph.all_edge_type]
num_steps = int((euler_graph.total_size + 1) // flags_obj.batch_size *
flags_obj.num_epochs)
model = ARMA(dims,
metapath,
euler_graph.feature_idx,
euler_graph.feature_dim,
euler_graph.label_idx,
euler_graph.label_dim,
num_layers=flags_obj.layers,
K=flags_obj.K)
params = {
'train_node_type': euler_graph.train_node_type[0],
'batch_size': flags_obj.batch_size,
'optimizer': flags_obj.optimizer,
'learning_rate': flags_obj.learning_rate,
'log_steps': flags_obj.log_steps,
'model_dir': flags_obj.model_dir,
'id_file': euler_graph.id_file,
'infer_dir': flags_obj.model_dir,
'total_size': euler_graph.total_size,
'total_step': num_steps
}
config = tf.estimator.RunConfig(log_step_count_steps=None)
model_estimator = NodeEstimator(model, params, config)
if flags_obj.run_mode == 'train':
model_estimator.train()
elif flags_obj.run_mode == 'evaluate':
model_estimator.evaluate()
elif flags_obj.run_mode == 'infer':
model_estimator.infer()
else:
raise ValueError('Run mode not exist!')
def test_difference_sign(self):
const = np.ones(10)
decreasing = np.array([2, 1, 0.5, 0, -0.2])
increasing = np.array([-2, -1, 0, 0.1, 0.2, 0.4])
seq = np.array([1, 0.5, 0, 1, 0.4, 0.2, 3])
short = np.array([0, 1, 0])
very_short = np.array([0, 1])
singleton = np.array([0])
empty = np.array([])
arma = ARMA([0])
self.assertEqual(arma._differene_sign(const), 0)
self.assertEqual(arma._differene_sign(decreasing), 0)
self.assertEqual(arma._differene_sign(increasing), 5)
self.assertEqual(arma._differene_sign(seq), 2)
self.assertEqual(arma._differene_sign(short), 1)
self.assertEqual(arma._differene_sign(very_short), 1)
self.assertEqual(arma._differene_sign(singleton), 0)
self.assertEqual(arma._differene_sign(empty), 0)
def test_turning_pts(self):
const = np.ones(10)
decreasing = np.array([2, 1, 0.5, 0, -0.2])
increasing = np.array([-2, -1, 0, 0.1, 0.2, 0.4])
seq = np.array([1, 0.5, 0, 1, 0.4, 0.2, 3])
short = np.array([0, 1, 0])
very_short = np.array([0, 1])
singleton = np.array([0])
empty = np.array([])
arma = ARMA([0])
self.assertEqual(arma._turning_points(const), 0)
self.assertEqual(arma._turning_points(decreasing), 0)
self.assertEqual(arma._turning_points(increasing), 0)
self.assertEqual(arma._turning_points(seq), 3)
self.assertEqual(arma._turning_points(short), 1)
self.assertEqual(arma._turning_points(very_short), 0)
self.assertEqual(arma._turning_points(singleton), 0)
self.assertEqual(arma._turning_points(empty), 0)
def test_ar_fit_yule_walker(self):
zeros = np.zeros(100)
zero_ts = ARMA(zeros)
np.random.seed(12345)
simulated_ar3 = simulate_arma(phi=[0.5, 0.1, 0.2], sigma=0.5)
ar3 = ARMA(simulated_ar3)
zero_ts.fit_ar(p=0, method="yule_walker")
self.assertEqual(len(zero_ts.model.get_params()[0]), 0)
ar3.fit_ar(p=3, method="yule_walker")
self.assertAlmostEqual(ar3.model.get_params()[0][0], 0.5, 1)
self.assertAlmostEqual(ar3.model.get_params()[0][1], 0.1, 1)
self.assertAlmostEqual(ar3.model.get_params()[0][2], 0.2, 1)
def test_sample_autocovariance(self):
zeros = np.zeros(100)
zero_ts = ARMA(zeros, subtract_mean=True)
ones = np.ones(10)
ones_ts_mean_corrected = ARMA(ones, subtract_mean=True)
ones_ts_mean_uncorrected = ARMA(ones, subtract_mean=False)
integers = [1, 2, 3, 4, 5]
integer_ts = ARMA(integers)
for k in range(100):
self.assertEqual(zero_ts.sample_autocovariance(k), 0)
for k in range(10):
self.assertEqual(ones_ts_mean_corrected.sample_autocovariance(k), 0)
self.assertEqual(ones_ts_mean_corrected.sample_autocovariance(-k), 0)
for k in range(10):
self.assertEqual(ones_ts_mean_uncorrected.sample_autocovariance(-k), 0)
self.assertEqual(ones_ts_mean_uncorrected.sample_autocovariance(k), 0)
self.assertAlmostEqual(integer_ts.sample_autocovariance(0), 2, 10)
self.assertAlmostEqual(integer_ts.sample_autocovariance(1), 0.8, 10)
self.assertAlmostEqual(integer_ts.sample_autocovariance(2), -0.2, 10)
self.assertAlmostEqual(integer_ts.sample_autocovariance(3), -0.8, 10)
self.assertAlmostEqual(integer_ts.sample_autocovariance(4), -0.8, 10)
def test_sample_pacf(self):
integers = [1, 2, 3, 4, 5]
integer_ts = ARMA(integers)
self.assertAlmostEqual(integer_ts.sample_pacf(1), 0.4)
self.assertAlmostEqual(integer_ts.sample_pacf(2), -13 / 42)
def test_one_step_predictions(self):
ma_model = PureARMA(theta=[-0.9], sigma_sq=1)
ma_data = [-2.58, 1.62, -0.96, 2.62, -1.36]
ma = ARMA(ma_data, subtract_mean=False)
self.assertEqual(ma.get_one_step_predictor(0, ma_model), 0)
self.assertAlmostEqual(ma.get_one_step_predictor(1, ma_model, method="innovations_algo"), 1.28, 2)
self.assertAlmostEqual(ma.get_one_step_predictor(2, ma_model, method="innovations_algo"), -0.22, 2)
self.assertAlmostEqual(ma.get_one_step_predictor(3, ma_model, method="innovations_algo"), 0.55, 2)
self.assertAlmostEqual(ma.get_one_step_predictor(4, ma_model, method="innovations_algo"), -1.63, 2)
self.assertAlmostEqual(ma.get_one_step_predictor(5, ma_model, method="innovations_algo"), -0.22, 2)
arma_data = [-1.1, 0.514, 0.116, -0.845, 0.872, -0.467, -0.977, -1.699, -1.228, -1.093]
arma_model = PureARMA(phi=[0.2], theta=[0.4], sigma_sq=1)
arma = ARMA(arma_data, subtract_mean=False)
self.assertEqual(arma.get_one_step_predictor(0, arma_model, method="innovations_algo"), 0)
self.assertAlmostEqual(arma.get_one_step_predictor(1, arma_model, method="innovations_algo"), -0.534, 1)
self.assertAlmostEqual(arma.get_one_step_predictor(2, arma_model, method="innovations_algo"), 0.5068, 1)
self.assertAlmostEqual(arma.get_one_step_predictor(3, arma_model, method="innovations_algo"), -0.1321, 1)
self.assertAlmostEqual(arma.get_one_step_predictor(4, arma_model, method="innovations_algo"), -0.4539, 1)
self.assertAlmostEqual(arma.get_one_step_predictor(5, arma_model, method="innovations_algo"), 0.7046, 1)
data = [1, 2, 3, 4, 5]
empty_model = PureARMA()
empty = ARMA(data)
empty.model = empty_model
for k in range(6):
self.assertEqual(empty.get_one_step_predictor(k, method="innovations_algo"), 0)
def test_no_fitting_exceptions(self):
np.random.seed(12345)
simulated_arma = simulate_arma(phi=[0.2, 0.5], theta=[0.2], simulations=10)
arma = ARMA(simulated_arma)
for method in arma._implemented_arma_methods:
arma.fit_arma(p=2, q=1, method=method)
def test_reduced_likelihood(self):
ma_model = PureARMA(theta=[-0.9], sigma_sq=0.5)
ma_data = [-2.58, 1.62, -0.96, 2.62, -1.36]
ma = ARMA(ma_data, subtract_mean=False)
self.assertAlmostEqual(ma.get_reduced_likelihood(ma_model), 0.7387, 2)
