def test_oob_sarimax():
xreg = rs.rand(wineind.shape[0], 2)
fit = ARIMA(order=(1, 1, 1),
seasonal_order=(0, 1, 1, 12),
out_of_sample_size=15).fit(y=wineind, exogenous=xreg)
fit_no_oob = ARIMA(order=(1, 1, 1),
seasonal_order=(0, 1, 1, 12),
out_of_sample_size=0,
suppress_warnings=True).fit(y=wineind[:-15],
exogenous=xreg[:-15, :])
# now assert some of the same things here that we did in the former test
oob = fit.oob()
# compare scores:
scoring = get_callable(fit_no_oob.scoring, VALID_SCORING)
no_oob_preds = fit_no_oob.predict(n_periods=15, exogenous=xreg[-15:, :])
assert np.allclose(oob, scoring(wineind[-15:], no_oob_preds), rtol=1e-2)
# show params are still the same
assert np.allclose(fit.params(), fit_no_oob.params(), rtol=1e-2)
# show we can add the new samples and get the exact same forecasts
xreg_test = rs.rand(5, 2)
fit_no_oob.add_new_observations(wineind[-15:], xreg[-15:, :])
assert np.allclose(fit.predict(5, xreg_test),
fit_no_oob.predict(5, xreg_test),
rtol=1e-2)
# Show we can get a confidence interval out here
preds, conf = fit.predict(5, xreg_test, return_conf_int=True)
assert all(isinstance(a, np.ndarray) for a in (preds, conf))
def test_with_oob():
# show we can fit with CV (kinda)
arima = ARIMA(order=(2, 1, 2),
suppress_warnings=True,
scoring='mse',
out_of_sample_size=10).fit(y=hr)
oob = arima.oob()
assert not np.isnan(oob) # show this works
# Assert the predictions give the expected MAE/MSE
oob_preds = arima.oob_preds_
assert oob_preds.shape[0] == 10
scoring = val.get_scoring_metric('mse')
assert scoring(hr[-10:], oob_preds) == oob
# show we can fit if ooss < 0 and oob will be nan
arima = ARIMA(order=(2, 1, 2),
suppress_warnings=True,
out_of_sample_size=-1).fit(y=hr)
assert np.isnan(arima.oob())
# This will raise since n_steps is not an int
with pytest.raises(TypeError):
arima.predict(n_periods="5")
# But that we CAN forecast with an int...
_ = arima.predict(n_periods=5) # noqa: F841
# Show we fail if cv > n_samples
with pytest.raises(ValueError):
ARIMA(order=(2, 1, 2), out_of_sample_size=1000).fit(hr)
def test_basic_arima():
arima = ARIMA(order=(0, 0, 0), suppress_warnings=True)
preds = arima.fit_predict(y) # fit/predict for coverage
# test some of the attrs
assert_almost_equal(arima.aic(), 11.201308403566909, decimal=5)
assert_almost_equal(arima.aicc(), 11.74676, decimal=5)
assert_almost_equal(arima.bic(), 13.639060053303311, decimal=5)
# get predictions
expected_preds = np.array([
0.44079876, 0.44079876, 0.44079876, 0.44079876, 0.44079876, 0.44079876,
0.44079876, 0.44079876, 0.44079876, 0.44079876
])
# generate predictions
assert_array_almost_equal(preds, expected_preds)
# Make sure we can get confidence intervals
expected_intervals = np.array([[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139]])
_, intervals = arima.predict(n_periods=10,
return_conf_int=True,
alpha=0.05)
assert_array_almost_equal(intervals, expected_intervals)
def test_more_elaborate():
# show we can fit this with a non-zero order
arima = ARIMA(order=(2, 1, 2), suppress_warnings=True).fit(y=hr)
_try_get_attrs(arima)
# can we fit this same arima with a made-up exogenous array?
xreg = rs.rand(hr.shape[0], 4)
arima = ARIMA(order=(2, 1, 2), suppress_warnings=True).fit(y=hr,
exogenous=xreg)
_try_get_attrs(arima)
# pickle this for the __get/setattr__ coverage.
# since the only time this is tested is in parallel in auto.py,
# this doesn't actually get any coverage proof...
fl = 'some_temp_file.pkl'
with open(fl, 'wb') as p:
pickle.dump(arima, p)
# show we can predict with this even though it's been pickled
new_xreg = rs.rand(5, 4)
_preds = arima.predict(n_periods=5, exogenous=new_xreg)
# now unpickle
with open(fl, 'rb') as p:
other = pickle.load(p)
# show we can still predict, compare
_other_preds = other.predict(n_periods=5, exogenous=new_xreg)
assert_array_almost_equal(_preds, _other_preds)
# now remove the pickle file
os.unlink(fl)
# now show that since we fit the ARIMA with an exogenous array,
# we need to provide one for predictions otherwise it breaks.
with pytest.raises(ValueError):
arima.predict(n_periods=5, exogenous=None)
# show that if we DO provide an exogenous and it's the wrong dims, we
# also break things down.
with pytest.raises(ValueError):
arima.predict(n_periods=5, exogenous=rs.rand(4, 4))
def test_with_seasonality1():
fit = ARIMA(order=(1, 1, 1),
seasonal_order=(0, 1, 1, 12),
suppress_warnings=True).fit(y=wineind)
_try_get_attrs(fit)
# R code AIC result is ~3004
assert abs(fit.aic() - 3004) < 100 # show equal within 100 or so
# R code AICc result is ~3005
assert abs(fit.aicc() - 3005) < 100 # show equal within 100 or so
# R code BIC result is ~3017
assert abs(fit.bic() - 3017) < 100 # show equal within 100 or so
# show we can predict in-sample
fit.predict_in_sample()
# test with SARIMAX confidence intervals
fit.predict(n_periods=10, return_conf_int=True, alpha=0.05)
def test_oob_for_issue_29(self, d, cv, exog):
model = ARIMA(order=(2, d, 0),
out_of_sample_size=cv).fit(self.dta, exogenous=exog)
# If exogenous is defined, we need to pass n_periods of
# exogenous rows to the predict function. Otherwise we'll
# just leave it at None
if exog is not None:
xr = exog[:3, :]
else:
xr = None
_, _ = model.predict(n_periods=3, return_conf_int=True, exogenous=xr)
def test_oob_sarimax():
xreg = rs.rand(wineind.shape[0], 2)
fit = ARIMA(order=(1, 1, 1),
seasonal_order=(0, 1, 1, 12),
maxiter=5,
out_of_sample_size=15).fit(y=wineind, exogenous=xreg)
fit_no_oob = ARIMA(order=(1, 1, 1),
seasonal_order=(0, 1, 1, 12),
out_of_sample_size=0,
maxiter=5,
suppress_warnings=True).fit(y=wineind[:-15],
exogenous=xreg[:-15, :])
# now assert some of the same things here that we did in the former test
oob = fit.oob()
# compare scores:
scoring = val.get_scoring_metric(fit_no_oob.scoring)
no_oob_preds = fit_no_oob.predict(n_periods=15, exogenous=xreg[-15:, :])
assert np.allclose(oob, scoring(wineind[-15:], no_oob_preds), rtol=1e-2)
# show params are no longer the same
assert not np.allclose(fit.params(), fit_no_oob.params(), rtol=1e-2)
# show we can add the new samples and get the exact same forecasts
xreg_test = rs.rand(5, 2)
fit_no_oob.update(wineind[-15:], xreg[-15:, :])
assert np.allclose(fit.predict(5, xreg_test),
fit_no_oob.predict(5, xreg_test),
rtol=1e-2)
# And also the params should be close now after updating
assert np.allclose(fit.params(), fit_no_oob.params())
# Show we can get a confidence interval out here
preds, conf = fit.predict(5, xreg_test, return_conf_int=True)
assert all(isinstance(a, np.ndarray) for a in (preds, conf))
def test_oob_for_issue_29():
dta = sm.datasets.sunspots.load_pandas().data
dta.index = pd.Index(sm.tsa.datetools.dates_from_range('1700', '2008'))
del dta["YEAR"]
xreg = np.random.RandomState(1).rand(dta.shape[0], 3)
# Try for cv on/off, various D levels, and various Xregs
for d in (0, 1):
for cv in (0, 3):
for exog in (xreg, None):
# surround with try/except so we can log the failing combo
try:
model = ARIMA(order=(2, d, 0),
out_of_sample_size=cv).fit(dta,
exogenous=exog)
# If exogenous is defined, we need to pass n_periods of
# exogenous rows to the predict function. Otherwise we'll
# just leave it at None
if exog is not None:
xr = exog[:3, :]
else:
xr = None
_, _ = model.predict(n_periods=3,
return_conf_int=True,
exogenous=xr)
except Exception as ex:
print("Failing combo: d=%i, cv=%i, exog=%r" %
(d, cv, exog))
# Statsmodels can be fragile with ARMA coefficient
# computation. If we encounter that, pass:
# ValueError: The computed initial MA coefficients are
# not invertible. You should induce invertibility,
# choose a different model order, or ...
if "invertibility" in str(ex):
pass
else:
raise
def test_basic_arma():
arma = ARIMA(order=(0, 0, 0), suppress_warnings=True)
preds = arma.fit_predict(y) # fit/predict for coverage
# No OOB, so assert none
assert arma.oob_preds_ is None
# test some of the attrs
assert_almost_equal(arma.aic(), 11.201, decimal=3) # equivalent in R
# intercept is param 0
intercept = arma.params()[0]
assert_almost_equal(intercept, 0.441, decimal=3) # equivalent in R
assert_almost_equal(arma.aicc(), 11.74676, decimal=5)
assert_almost_equal(arma.bic(), 13.639060053303311, decimal=5)
# get predictions
expected_preds = np.array([0.44079876, 0.44079876, 0.44079876,
0.44079876, 0.44079876, 0.44079876,
0.44079876, 0.44079876, 0.44079876,
0.44079876])
# generate predictions
assert_array_almost_equal(preds, expected_preds)
# Make sure we can get confidence intervals
expected_intervals = np.array([
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139],
[-0.10692387, 0.98852139]
])
_, intervals = arma.predict(n_periods=10, return_conf_int=True,
alpha=0.05)
assert_array_almost_equal(intervals, expected_intervals)
def test_with_oob():
# show we can fit with CV (kinda)
arima = ARIMA(order=(2, 1, 2),
suppress_warnings=True,
out_of_sample_size=10).fit(y=hr)
assert not np.isnan(arima.oob()) # show this works
# show we can fit if ooss < 0 and oob will be nan
arima = ARIMA(order=(2, 1, 2), suppress_warnings=True,
out_of_sample_size=-1).fit(y=hr)
assert np.isnan(arima.oob())
# This will raise since n_steps is not an int
assert_raises(TypeError, arima.predict, n_periods="5")
# But that we CAN forecast with an int...
_ = arima.predict(n_periods=5) # noqa: F841
# Show we fail if cv > n_samples
assert_raises(ValueError,
ARIMA(order=(2, 1, 2), out_of_sample_size=1000).fit, hr)
def ragged_fill_series(
series,
function=np.nanmean,
backup_fill_method=np.nanmean,
est_series=None,
fitted_arma=None,
arma_full_series=None,
):
"""Filling in the ragged ends of a series, adhering to the periodicity of the series. If there is only one observation and periodicity cannot be determined, series will be returned unchanged.
parameters:
:series: list/pandas Series: the series to fill the ragged edges of. Missings should be np.nans
:function: the function to fill nas with (e.g. np.nanmean, etc.). Use "ARMA" for ARMA filling
:backup_fill_method: function: which function to fill ragged edges with in case ARMA can't be estimated
:est_series: list/pandas Series: optional, the series to calculate the fillna and/or ARMA function on. Should not have nas filled in yet by any method. E.g. a train set. If None, will calculated based on itself.
:fitted_arma: optional, fitted ARMA model if available to avoid reestimating every time in the `gen_ragged_X` function
:arma_full_series: optional, for_full_arma_dataset output of `gen_dataset` function. Fitting the ARMA model on the full series history rather than just the series provided
output:
:return: pandas Series with filled ragged edges
"""
result = pd.Series(series).copy()
if est_series is None:
est_series = result.copy()
# periodicity of the series, to see which to fill in
nonna_bools = ~pd.isna(series)
nonna_indices = list(
nonna_bools.index[nonna_bools]) # existing indices with values
# if there is only one non-na observation, can't determine periodicity or position in full series, don't fill anything
if len(nonna_indices) > 1:
periodicity = int(
(pd.Series(result[~pd.isna(result)].index) -
(pd.Series(result[~pd.isna(result)].index)).shift()
).mode()[0]) # how often data comes (quarterly, monthly, etc.)
last_nonna = result.index[result.notna()][-1]
fill_indices = nonna_indices + [
int(nonna_indices[-1] + periodicity * i)
for i in range(1, (len(series) - last_nonna))
] # indices to be filled in, including only the correct periodicity
fill_indices = [x for x in fill_indices if x in series.index
] # cut down on the indices if went too long
if function == "ARMA":
# estimate the model if not given
if fitted_arma is None:
fitted_arma = estimate_arma(est_series)
# instantiate model with previously estimated parameters (i.e. on train set)
arma = ARIMA(order=fitted_arma.order)
arma.set_params(**fitted_arma.get_params())
# refit the model on the full series to this point
if arma_full_series is not None:
y = list(arma_full_series[~pd.isna(arma_full_series)])
present = list(result[~pd.isna(result)])
# limit the series to the point where actuals are
end_index = 0
for i in range(len(present), len(y) + 1):
if list(y[(i - len(present)):i]) == list(present):
end_index = i
y = y[:end_index]
# refit model on just this series
else:
y = list(result[~pd.isna(result)]) # refit the model on data
present = y.copy()
# can fail if not enough datapoints for order of ARMA process
try:
arma.fit(y, error_action="ignore")
preds = arma.predict(n_periods=int(len(series) - last_nonna))
fills = list(present) + list(preds)
fills = fills[:len(fill_indices)]
except:
fills = list(result[~pd.isna(result)]) + [
backup_fill_method(est_series)
] * (len(series) - last_nonna)
fills = fills[:len(fill_indices)]
result[fill_indices] = fills
else:
fills = list(result[~pd.isna(result)]) + [function(est_series)] * (
len(series) - last_nonna)
fills = fills[:len(fill_indices)]
result[fill_indices] = fills
return result, fitted_arma
def test_oob_for_issue_28():
# Continuation of above: can we do one with an exogenous array, too?
xreg = rs.rand(hr.shape[0], 4)
arima = ARIMA(order=(2, 1, 2),
suppress_warnings=True,
out_of_sample_size=10).fit(y=hr, exogenous=xreg)
oob = arima.oob()
assert not np.isnan(oob)
# Assert that the endog shapes match. First is equal to the original,
# and the second is the differenced array, with original shape - d.
assert np.allclose(arima.arima_res_.data.endog, hr, rtol=1e-2)
assert arima.arima_res_.model.endog.shape[0] == hr.shape[0] - 1
# Now assert the same for exog
assert np.allclose(arima.arima_res_.data.exog, xreg, rtol=1e-2)
assert arima.arima_res_.model.exog.shape[0] == xreg.shape[0] - 1
# Compare the OOB score to an equivalent fit on data - 10 obs, but
# without any OOB scoring, and we'll show that the OOB scoring in the
# first IS in fact only applied to the first (train - n_out_of_bag)
# samples
arima_no_oob = ARIMA(order=(2, 1, 2),
suppress_warnings=True,
out_of_sample_size=0).fit(y=hr[:-10],
exogenous=xreg[:-10, :])
scoring = get_callable(arima_no_oob.scoring, VALID_SCORING)
preds = arima_no_oob.predict(n_periods=10, exogenous=xreg[-10:, :])
assert np.allclose(oob, scoring(hr[-10:], preds), rtol=1e-2)
# Show that the model parameters are exactly the same
xreg_test = rs.rand(5, 4)
assert np.allclose(arima.params(), arima_no_oob.params(), rtol=1e-2)
# Now assert on the forecast differences.
with_oob_forecasts = arima.predict(n_periods=5, exogenous=xreg_test)
no_oob_forecasts = arima_no_oob.predict(n_periods=5, exogenous=xreg_test)
assert_raises(AssertionError, assert_array_almost_equal,
with_oob_forecasts, no_oob_forecasts)
# But after we update the no_oob model with the latest data, we should
# be producing the same exact forecasts
# First, show we'll fail if we try to add observations with no exogenous
assert_raises(ValueError, arima_no_oob.add_new_observations, hr[-10:],
None)
# Also show we'll fail if we try to add mis-matched shapes of data
assert_raises(ValueError, arima_no_oob.add_new_observations, hr[-10:],
xreg_test)
# Show we fail if we try to add observations with a different dim exog
assert_raises(ValueError, arima_no_oob.add_new_observations, hr[-10:],
xreg_test[:, :2])
# Actually add them now, and compare the forecasts (should be the same)
arima_no_oob.add_new_observations(hr[-10:], xreg[-10:, :])
assert np.allclose(with_oob_forecasts,
arima_no_oob.predict(n_periods=5, exogenous=xreg_test),
rtol=1e-2)
class ARIMAModel(BaseModel):
def __init__(self):
"""
Initialize Model
"""
self.seasonal = True
self.metric = 'mse'
self.model = None
self.model_init = False
def _build(self, **config):
"""
build the models and initialize.
:param config: hyperparameters for the model
"""
p = config.get('p', 2)
d = config.get('d', 0)
q = config.get('q', 2)
self.seasonal = config.get('seasonality_mode', True)
P = config.get('P', 1)
D = config.get('D', 0)
Q = config.get('Q', 1)
m = config.get('m', 7)
self.metric = config.get('metric', self.metric)
order = (p, d, q)
if not self.seasonal:
seasonal_order = (0, 0, 0, 0)
else:
seasonal_order = (P, D, Q, m)
self.model = ARIMA(order=order,
seasonal_order=seasonal_order,
suppress_warnings=True)
def fit_eval(self, data, validation_data, **config):
"""
Fit on the training data from scratch.
:param data: A 1-D numpy array as the training data
:param validation_data: A 1-D numpy array as the evaluation data
:return: the evaluation metric value
"""
if not self.model_init:
# Estimating differencing term (d) and seasonal differencing term (D)
kpss_diffs = ndiffs(data, alpha=0.05, test='kpss', max_d=6)
adf_diffs = ndiffs(data, alpha=0.05, test='adf', max_d=6)
d = max(adf_diffs, kpss_diffs)
D = 0 if not self.seasonal else nsdiffs(data, m=7, max_D=12)
config.update(d=d, D=D)
self._build(**config)
self.model_init = True
self.model.fit(data)
val_metric = self.evaluate(x=None,
target=validation_data,
metrics=[self.metric])[0].item()
return {self.metric: val_metric}
def predict(self, x=None, horizon=24, update=False, rolling=False):
"""
Predict horizon time-points ahead the input x in fit_eval
:param x: ARIMA predicts the horizon steps foreward from the training data.
So x should be None as it is not used.
:param horizon: the number of steps forward to predict
:param update: whether to update the original model
:param rolling: whether to use rolling prediction
:return: predicted result of length horizon
"""
if x is not None:
raise ValueError("x should be None")
if update and not rolling:
raise Exception(
"We don't support updating model without rolling prediction currently"
)
if self.model is None:
raise Exception(
"Needs to call fit_eval or restore first before calling predict"
)
if not update and not rolling:
forecasts = self.model.predict(n_periods=horizon)
elif rolling:
if not update:
self.save("tmp.pkl")
forecasts = []
for step in range(horizon):
fc = self.model.predict(n_periods=1).item()
forecasts.append(fc)
# Updates the existing model with a small number of MLE steps for rolling prediction
self.model.update(fc)
if not update:
self.restore("tmp.pkl")
os.remove("tmp.pkl")
return forecasts
def evaluate(self, target, x=None, metrics=['mse'], rolling=False):
"""
Evaluate on the prediction results and y. We predict horizon time-points ahead the input x
in fit_eval before evaluation, where the horizon length equals the second dimension size of
y.
:param target: target for evaluation.
:param x: ARIMA predicts the horizon steps foreward from the training data.
So x should be None as it is not used.
:param metrics: a list of metrics in string format
:param rolling: whether to use rolling prediction
:return: a list of metric evaluation results
"""
if x is not None:
raise ValueError("We don't support input x currently")
if target is None:
raise ValueError("Input invalid target of None")
if self.model is None:
raise Exception(
"Needs to call fit_eval or restore first before calling evaluate"
)
forecasts = self.predict(horizon=len(target), rolling=rolling)
return [Evaluator.evaluate(m, target, forecasts) for m in metrics]
def save(self, checkpoint_file):
if self.model is None:
raise Exception(
"Needs to call fit_eval or restore first before calling save")
with open(checkpoint_file, 'wb') as fout:
pickle.dump(self.model, fout)
def restore(self, checkpoint_file):
with open(checkpoint_file, 'rb') as fin:
self.model = pickle.load(fin)
self.model_init = True
class Model:
def select_data(self):
# Merge columns into a single dataframe of observed values, based on date
dataset = files.data_main.join(files.data_exo.set_index('date'),
on='date').dropna()
# Select part of the precipitation dataframe that corresponds to the forecast
obs_end = dataset.tail(1)['date'].values[0]
exo_prev = files.data_exo[(files.data_exo['date'] > obs_end)]
# Select predict dates
self.dates_prev = exo_prev['date']
# Reshape
endo_obs = np.array(dataset['endo_value'])
self.endo_obs = endo_obs.reshape(-1, 1)
exo_obs = np.array(dataset['exo_value'])
self.exo_obs = exo_obs.reshape(-1, 1)
exo_prev = np.array(exo_prev['exo_value'])
self.exo_prev = exo_prev.reshape(-1, 1)
def normalize(self):
# Calculate lambda only if doesn't have zero values
n_zeros = len(self.endo_obs[self.endo_obs <= 0])
if n_zeros == 0:
self.endo_obs2, self.lambda_boxcox = boxcox(self.endo_obs)
else:
self.lambda_boxcox = -999
# Limit lambda values
if abs(self.lambda_boxcox[0]) > 1:
self.endo_obs2 = self.endo_obs
self.lambda_boxcox = -999
#print(self.endo_obs2, self.lambda_boxcox)
def run_auto(self):
self.arima_model = auto_arima(self.endo_obs2,
start_p=0,
start_d=0,
start_q=0,
max_p=3,
max_d=1,
max_q=3,
start_P=0,
start_Q=0,
D=1,
seasonal=False,
m=1,
exogeneous=self.exo_obs,
trace=True,
error_action='ignore',
suppress_warnings=True,
stepwise=True)
#print(model.arima_model.summary())
# Compile parameters to list
self.parameters = [
self.arima_model.order, self.arima_model.seasonal_order,
self.lambda_boxcox[0],
self.arima_model.aic()
]
print(self.parameters)
return (self.arima_model)
def run_auto_arimax(self):
lower_aic = float(99999)
best_pdq = [0, 0, 0]
param = list(itertools.product(range(0, 4), range(0, 2), range(0, 4)))
for pdq in param:
#print(pdq)
try:
self.arima_model = ARIMA(order=pdq,
suppress_warnings=True).fit(
y=self.endo_obs2,
exogenous=self.exo_obs)
if self.arima_model.aic() < lower_aic:
lower_aic = self.arima_model.aic()
best_pdq = tuple(self.arima_model.order)
except:
continue
#print(model.arima_model.summary())
# Compile parameters to list
self.parameters = [best_pdq, self.lambda_boxcox[0], lower_aic]
print(self.parameters)
return (self.arima_model)
def run_auto_sarimax(self):
lower_aic = float(99999)
best_pdq = [0, 0, 0]
best_spdq = [0, 0, 0]
param = list(itertools.product(range(0, 4), range(0, 2), range(0, 4)))
m = 1 # frequency
param_seasonal = [(x[0], x[1], x[2], m) for x in list(
itertools.product(range(0, 4), range(0, 2), range(0, 4)))]
for pdq in param:
for spdq in param_seasonal:
try:
mod = sm.tsa.statespace.SARIMAX(
self.endo_obs2,
exog=self.exo_obs,
order=pdq,
seasonal_order=spdq,
enforce_stationarity=False,
enforce_invertibility=False)
self.arima_model = mod.fit(disp=0)
print('ARIMA{}x{}{} - AIC:{}'.format(
pdq, spdq, m, self.arima_model.aic))
if self.arima_model.aic() < lower_aic:
lower_aic = self.arima_model.aic()
best_pdq = tuple(pdq)
best_spdq = tuple(spdq)
except:
continue
#print(model.arima_model.summary())
# Compile parameters to list
self.parameters = [
best_pdq, best_spdq, self.lambda_boxcox[0], lower_aic
]
print(self.parameters)
return (self.arima_model)
def forecast(self):
#self.predict = self.arima_model.predict(n_periods=self.exo_prev.shape[0], exogenous=self.exo_prev)
self.predict = self.arima_model.predict(
n_periods=self.exo_prev.shape[0],
exogenous=self.exo_prev,
return_conf_int=True,
alpha=0.7)
def renormalize(self):
if self.lambda_boxcox == float(-999):
self.predict_mean = self.predict[0]
self.predict_down = self.predict[1][:, 0]
self.predict_up = self.predict[1][:, 1]
else:
self.predict_mean = inv_boxcox(self.predict[0], self.lambda_boxcox)
self.predict_down = inv_boxcox(self.predict[1][:, 0],
self.lambda_boxcox)
self.predict_up = inv_boxcox(self.predict[1][:, 1],
self.lambda_boxcox)
# Join predict dates with values into a dataframe
df_final = pd.DataFrame(self.predict_mean, self.dates_prev)
df_final.columns = ['endo_value']
return (df_final)
# Plot Residuals and fitted values
# plt.figure()
# fitted_values = arima.predict_in_sample()
# plt.plot(df.index[:train_len - 1], fitted_values,
# color='C0', label="Fitted values")
# plt.plot(pd.to_datetime(df.index), data, color='C1', label="Data")
# plt.plot(df.index[:train_len - 1], arima.resid(),
# color='C2', label="Residuals")
# plt.gca().grid(which='both', axis='x', linestyle='--')
# plt.title("Residuals and fitted values")
# plt.legend()
print("SSE: {}".format((arima.resid()**2).sum()))
# Plot fitted values and forecasts
predictions = arima.predict(n_periods=test.shape[0])
fitted_values = arima.predict_in_sample()
plt.figure()
plt.plot(df.index[train_len:], test, '--', color='C0', label="test set")
plt.plot(df.index[train_len:],
predictions,
'--',
color='C1',
label="forecasted values")
plt.plot(df.index[:train_len], train, color='C0', label="train set")
plt.plot(df.index[:train_len - 1],
fitted_values,
color='C1',
label="fitted values")
plt.legend()
plt.title("Fitted values and forecasts")
def predict_arima(df):
time_in=current_milli_time()
try:
forecast_in = open("forecast.pickle","rb")
future_forecast = pickle.load(forecast_in)
forecast_in.append(df)
error=[]
"""
Calculate errors
"""
if len(df) < len(future_forecast):
error=df["memory_used"] - future_forecast[:len(df)]["memory_used"]
elif len(df) > len(future_forecast):
error=df[0:len(future_forecast)]["memory_used"]- future_forecast["memory_used"]
else:
error=df["memory_used"]-future_forecast["memory_used"]
overestimation=[x for x in error if x<0]
overestimation=sum(overestimation)/len(overestimation)
underestimation=[x for x in error if x>=0]
underestimation=sum(underestimation)/len(underestimation)
print("UNDERESTIMATION ERROR: "+underestimation)
print("OVERESTIMATION ERROR: "+overestimation)
print("Mean Absolute Error in Last iteration "+str(error))
"""
Overestimation & Underestimation errors
"""
except Exception as e:
print("RMSE To be computed")
# Do Nothing
try:
pm.plot_pacf(df,show=False).savefig('pacf.png')
pm.plot_acf(df,show=False).savefig('acf.png')
except:
print("Data points insufficient for ACF & PACF")
try:
pickle_in = open("arima.pickle","rb")
arima_data = pickle.load(pickle_in)
arima_data.append(df)
#df=arima_data
except Exception as e:
arima_data_out = open("arima.pickle","wb")
pickle.dump([], arima_data_out)
arima_data_out = open("arima.pickle","wb")
pickle.dump(df, arima_data_out)
arima_data_out.close()
'''
tests
'''
nd=1
nsd=1
try:
adf_test=ADFTest(alpha=0.05)
p_val, should_diff = adf_test.is_stationary(df["memory_used"])
nd = ndiffs(df, test='adf')
logging.info(nd)
nsd = nsdiffs(df,12)
logging.info(nd)
except:
nd=1
print("Exception on tests")
ch_test=CHTest(12)
try:
nsd=ch_test.estimate_seasonal_differencing_term(df)
except Exception as e:
print(e)
logging.error(e)
'''
ARIMA MODEL
'''
'''
Find p,q dynamically
'''
acf_lags=acf(df["memory_used"])
acf_lags_threshold=[x for x in acf_lags if x>=getThreshold()]
p=len(acf_lags_threshold) if len(acf_lags_threshold)<=4 else 4
pacf_lags=pacf(df["memory_used"])
pacf_lags_threshold=[x for x in pacf_lags if x>=getThreshold()]
q=len(pacf_lags_threshold) if len(pacf_lags_threshold)<=1 else 1
d=nd
train, test = train_test_split(df,shuffle=False, test_size=0.3)
# If data is seasonal set the values of P,D,Q in seasonal order
stepwise_model = ARIMA(
order=(p,d,q),
seasonal_order=(0,nsd,0,12),
suppress_warnings=True,
scoring='mse'
)
x=str(p)+" "+str(nd)+" "+str(q)
print("Model with p="+str(q)+" d="+str(d)+" q="+str(q))
try:
stepwise_model.fit(df)
"""
Vary the periods as per the forecasting window
n_periods= 30 = 5mins
n_periods= 60 = 10mins
n_periods= 90 = 15mins
"""
future_forecast = stepwise_model.predict(n_periods=len(test))
future_forecast = pd.DataFrame(future_forecast,index=test.index,columns=["prediction"])
res=pd.concat([df,future_forecast],axis=1)
'''
Save Forecast in Pickle
'''
forecast_out = open("forecast.pickle","wb")
pickle.dump(future_forecast,forecast_out)
forecast_out.close()
trace1 = go.Scatter(x=res.index, y=res["prediction"],name="Prediction", mode='lines')
trace2 = go.Scatter(x=df.index, y=df["memory_used"],name="DF data", mode='lines')
data=[trace1,trace2]
layout = go.Layout(
title=x
)
fig = go.Figure(data=data, layout=layout)
plot(fig, filename="prediction")
print("Current values")
print(df)
print("Predicted Data Points")
print(future_forecast)
time_out=current_milli_time()
print("TIME for RNN(ms):"+str(time_out-time_in))
return future_forecast
except Exception as e:
time_out=current_milli_time()
print("TIME for RNN(ms):"+str(time_out-time_in))
print(e)
return None
def model_plot(days):
days = int(days)
pd.plotting.register_matplotlib_converters()
df = pd.read_csv('data/new_york.csv')
df['Date'] = pd.to_datetime(df['Date'])
#converting data to daily usage.
df.index = df.Date
df = df.drop('Date', axis=1)
# resample the dataframe every 1 day (D) and sum ovr each day
df = df.resample('D').sum()
df = df.tz_localize(None)
nyc_weather = pd.read_csv('data/weather/weatherNY.csv')
nyc_weather['DATE'] = pd.to_datetime(nyc_weather['DATE'])
nyc_weather = nyc_weather.set_index('DATE')
nyc_weather.drop(['NAME','STATION'],axis=1,inplace=True)
nyc_weather = nyc_weather['2015-07-01':'2020-08-10']
df = df[:'2020-08-10']
#trying 1 day increments with EXOG. MAYBE BEST CANDIDATE? with fourier terms june to june as 638 and august to august 516
day = days
real_values = []
predictions = []
df1 = df["2016":"2019"]
nyc_weather = nyc_weather["2016":"2019"]
y = df1.Consumption
exog = pd.DataFrame({'date': y.index})
exog = exog.set_index(pd.PeriodIndex(exog['date'], freq='D'))
exog['is_weekend'] = np.where(exog.index.dayofweek < 5,0,1)
#add weather data
exog['TMIN'] = nyc_weather['TMIN'].values
exog['sin1'] = np.sin(2 * np.pi * exog.index.dayofyear / 638)
exog['cos1'] = np.cos(2 * np.pi * exog.index.dayofyear / 638)
exog['sin2'] = np.sin(4 * np.pi * exog.index.dayofyear /638)
exog['cos2'] = np.cos(4 * np.pi * exog.index.dayofyear /638)
exog['sin3'] = np.sin(2 * np.pi * exog.index.dayofyear / 516)
exog['cos3'] = np.cos(2 * np.pi * exog.index.dayofyear / 516)
exog['sin4'] = np.sin(4 * np.pi * exog.index.dayofyear /516)
exog['cos4'] = np.cos(4 * np.pi * exog.index.dayofyear /516)
exog = exog.drop(columns=['date'])
num_to_update = 0
y_to_train = y.iloc[:(len(y)-100)]
exog_to_train = exog.iloc[:(len(y)-100)]
dates = []
steps = []
for i in range(5):
#first iteration train the model
if i == 0:
arima_exog_model = ARIMA(order=(3, 0, 1), seasonal_order=(2, 0, 0, 7),exogenous=exog_to_train, error_action='ignore',
initialization='approximate_diffuse', suppress_warnings=True).fit(y=y_to_train)
preds = arima_exog_model.predict_in_sample(exog_to_train)
#first prediction
y_to_test = y.iloc[(len(y)-100):(len(y)-100+day)]
y_exog_to_test = exog.iloc[(len(y)-100):(len(y)-100+day)]
y_arima_exog_forecast = arima_exog_model.predict(n_periods=day, exogenous=y_exog_to_test)
real_values.append(y_to_test.values)
predictions.append(y_arima_exog_forecast.tolist())
dates.append(y_to_test.index)
steps.append(y_to_test.index[-1])
#y_arima_exog_forecast = arima_exog_model.predict(n_periods=2, exogenous=exog_to_test)
else:
y_to_update = y.iloc[(len(y)-100+num_to_update):(len(y)-100+num_to_update)+day]
exog_to_update = exog.iloc[(len(y)-100+num_to_update):(len(y)-100+num_to_update)+day]
#to test
to_test = y.iloc[(len(y)-100+num_to_update)+day:(len(y)-100+num_to_update)+(day*2)]
exog_to_test = exog.iloc[(len(y)-100+num_to_update)+day:(len(y)-100+num_to_update)+(day*2)]
#update the model
arima_exog_model.update(y_to_update,exogenous=exog_to_update)
y_arima_exog_forecast = arima_exog_model.predict(n_periods=day, exogenous=exog_to_test)
dates.append(to_test.index)
steps.append(to_test.index[-1])
predictions.append(y_arima_exog_forecast.tolist())
real_values.append(to_test.values)
num_to_update += day
predict = [item for sublist in predictions for item in sublist]
true = [item for sublist in real_values for item in sublist]
dates = [item for sublist in dates for item in sublist]
#for viz purposes
y_to_train2 = y_to_train[-200:]
preds = preds[-200:]
y_to_train2 = y_to_train2.to_frame()
fig = go.Figure()
# Create and style traces
fig.add_trace(go.Scatter(x=y_to_train2.index, y=y_to_train2.Consumption, name='True values',
line=dict(color='firebrick', width=4,dash='dot')))
fig.add_trace(go.Scatter(x=y_to_train2.index, y=preds[-200:], name='In-sample Prediction',
line=dict(color='royalblue', width=4)))
fig.add_trace(go.Scatter(x=dates, y=predict, name='Prediction',
line=dict(color='green', width=4)))
fig.add_trace(go.Scatter(x=dates, y=true, name='True',
line=dict(color='firebrick', width=4,dash='dot')))
fig.update_layout(title='Electricity Consumption in New York',
xaxis_title='Date',
yaxis_title='Consumption',
xaxis_showgrid=True,
yaxis_showgrid=True,
#autosize=False,
#width=500,
#height=500,
paper_bgcolor=app_colors['background'],
plot_bgcolor=app_colors['background'])
return fig