def test_autoreg_append_deterministic(append_data):
y = append_data.y
y_oos = append_data.y_oos
y_both = append_data.y_both
x = append_data.x
x_oos = append_data.x_oos
x_both = append_data.x_both
terms = [TimeTrend(constant=True, order=1), Seasonality(12)]
dp = DeterministicProcess(y.index, additional_terms=terms)
res = AutoReg(y, lags=3, trend="n", deterministic=dp).fit()
res_append = res.append(y_oos, refit=True)
res_direct = AutoReg(y_both,
lags=3,
trend="n",
deterministic=dp.apply(y_both.index)).fit()
assert_allclose(res_append.params, res_direct.params)
res_np = AutoReg(np.asarray(y), lags=3, trend="n", deterministic=dp).fit()
res_append_np = res_np.append(np.asarray(y_oos))
assert_allclose(res_np.params, res_append_np.params)
res = AutoReg(y, exog=x, lags=3, trend="n", deterministic=dp).fit()
res_append = res.append(y_oos, exog=x_oos, refit=True)
res_direct = AutoReg(
y_both,
exog=x_both,
lags=3,
trend="n",
deterministic=dp.apply(y_both.index),
).fit()
assert_allclose(res_append.params, res_direct.params)
def test_drop_two_consants(time_index):
tt = TimeTrend(constant=True, order=1)
dp = DeterministicProcess(
time_index, constant=True, additional_terms=[tt], drop=True
)
assert dp.in_sample().shape[1] == 2
dp2 = DeterministicProcess(time_index, additional_terms=[tt], drop=True)
pd.testing.assert_frame_equal(dp.in_sample(), dp2.in_sample())
def test_deterministic(reset_randomstate):
y = pd.Series(np.random.normal(size=200))
terms = [TimeTrend(constant=True, order=1), Seasonality(12)]
dp = DeterministicProcess(y.index, additional_terms=terms)
m = AutoReg(y, trend="n", seasonal=False, lags=2, deterministic=dp)
res = m.fit()
m2 = AutoReg(y, trend="ct", seasonal=True, lags=2, period=12)
res2 = m2.fit()
assert_almost_equal(np.asarray(res.params), np.asarray(res2.params))
with pytest.warns(RuntimeWarning, match="When using deterministic, trend"):
AutoReg(y, trend="ct", seasonal=False, lags=2, deterministic=dp)
def test_drop():
index = pd.RangeIndex(0, 200)
dummy = DummyTerm()
str(dummy)
assert dummy != TimeTrend()
dp = DeterministicProcess(index, additional_terms=[dummy], drop=True)
in_samp = dp.in_sample()
assert in_samp.shape == (200, 4)
oos = dp.out_of_sample(37)
assert oos.shape == (37, 4)
assert list(oos.columns) == list(in_samp.columns)
valid = ("const", "trend", "dummy", "normal")
for valid_col in valid:
assert sum([1 for col in oos if valid_col in col]) == 1
def test_additional_terms(time_index):
add_terms = [TimeTrend(True, order=1)]
dp = DeterministicProcess(time_index, additional_terms=add_terms)
dp2 = DeterministicProcess(time_index, constant=True, order=1)
pd.testing.assert_frame_equal(dp.in_sample(), dp2.in_sample())
with pytest.raises(ValueError,
match="One or more terms in additional_terms"):
DeterministicProcess(time_index,
additional_terms=add_terms + add_terms)
with pytest.raises(ValueError,
match="One or more terms in additional_terms"):
DeterministicProcess(time_index,
constant=True,
order=1,
additional_terms=add_terms)
# This is equivalent to using the integer values 58 and 70.
det_proc.range(58, 70)
# ## Advanced Construction
#
# Deterministic processes with features not supported directly through the
# constructor can be created using `additional_terms` which accepts a list
# of `DetermisticTerm`. Here we create a deterministic process with two
# seasonal components: day-of-week with a 5 day period and an annual
# captured through a Fourier component with a period of 365.25 days.
from statsmodels.tsa.deterministic import Fourier, Seasonality, TimeTrend
index = pd.period_range("2020-03-01", freq="D", periods=2 * 365)
tt = TimeTrend(constant=True)
four = Fourier(period=365.25, order=2)
seas = Seasonality(period=7)
det_proc = DeterministicProcess(index, additional_terms=[tt, seas, four])
det_proc.in_sample().head(28)
# ## Custom Deterministic Terms
#
# The `DetermisticTerm` Abstract Base Class is designed to be subclassed
# to help users write custom deterministic terms. We next show two
# examples. The first is a broken time trend that allows a break after a
# fixed number of periods. The second is a "trick" deterministic term that
# allows exogenous data, which is not really a deterministic process, to be
# treated as if was deterministic. This lets use simplify gathering the
# terms needed for forecasting.
#
def test_time_trend_smoke(index, forecast_index):
tt = TimeTrend(True, 2)
tt.in_sample(index)
steps = 83 if forecast_index is None else len(forecast_index)
warn = None
if type(index) is NumericIndex and np.any(np.diff(index) != 1):
warn = UserWarning
with pytest.warns(warn):
tt.out_of_sample(steps, index, forecast_index)
str(tt)
hash(tt)
assert isinstance(tt.order, int)
assert isinstance(tt._constant, bool)
assert TimeTrend.from_string("ctt") == tt
assert TimeTrend.from_string("ct") != tt
assert TimeTrend.from_string("t") != tt
assert TimeTrend.from_string("n") != tt
assert Seasonality(12) != tt
tt0 = TimeTrend(False, 0)
tt0.in_sample(index)
str(tt0)
def test_time_trend(index):
tt = TimeTrend(constant=True)
const = tt.in_sample(index)
assert const.shape == (index.shape[0], 1)
assert np.all(const == 1)
pd.testing.assert_index_equal(const.index, index)
warn = None
if type(index) is NumericIndex and np.any(np.diff(index) != 1):
warn = UserWarning
with pytest.warns(warn):
const_fcast = tt.out_of_sample(23, index)
assert np.all(const_fcast == 1)
tt = TimeTrend(constant=False)
empty = tt.in_sample(index)
assert empty.shape == (index.shape[0], 0)
tt = TimeTrend(constant=False, order=2)
t2 = tt.in_sample(index)
assert t2.shape == (index.shape[0], 2)
assert list(t2.columns) == ["trend", "trend_squared"]
tt = TimeTrend(constant=True, order=2)
final = tt.in_sample(index)
expected = pd.concat([const, t2], axis=1)
pd.testing.assert_frame_equal(final, expected)
tt = TimeTrend(constant=True, order=2)
short = tt.in_sample(index[:-50])
with pytest.warns(warn):
remainder = tt.out_of_sample(50, index[:-50])
with pytest.warns(warn):
direct = tt.out_of_sample(steps=50,
index=index[:-50],
forecast_index=index[-50:])
combined = pd.concat([short, remainder], axis=0)
if isinstance(index, (pd.DatetimeIndex, pd.RangeIndex)):
pd.testing.assert_frame_equal(combined, final)
combined = pd.concat([short, direct], axis=0)
pd.testing.assert_frame_equal(combined, final, check_index_type=False)
def test_invalid_formcast_index(index):
tt = TimeTrend(order=4)
with pytest.raises(ValueError, match="The number of values in forecast_"):
tt.out_of_sample(10, index, pd.RangeIndex(11))