def run_arima(dept_id, store_id):
# create timeseries for fbprophet
ts = CreateTimeSeries(dept_id, store_id)
# ARIMA
fitted1 = SARIMAX(ts['y'].values + 1, order=(3, 1, 5), trend='c').fit()
fitted2 = SARIMAX(ts['y'].values + 1, order=(2, 1, 2), trend='c').fit()
y_pred = 0.5 * fitted1.forecast(28) + 0.5 * fitted2.forecast(28)
return np.append(np.array([dept_id, store_id]), y_pred - 1)
def initialize(self, order, seasonal_order):
self.order = order
self.seasonal_order = seasonal_order
self.model = SARIMAX(self.ts, order=self.order, seasonal_order=self.seasonal_order,
enforce_stationarity=False, enforce_invertibility=False,
simple_differencing=False)
self.results = self.model.fit(low_memory=True)
def order_select(self, order_sum_max=15):
self.order_sum_max = order_sum_max
# self.seasonal_order_sum_max = seasonal_order_sum_max
seasonal_order = (0, 0, 0, 0, 0, 0)
order_sum = sum(seasonal_order)
self.bic_value = pd.DataFrame(columns=['seasonal_order', 'bic'])
new_bic = (np.inf, None)
bic = (np.inf, None)
while order_sum < order_sum_max:
order_waiting = []
for i in range(6):
_tmp = list(seasonal_order)
_tmp[i] += 1
order_waiting.append(_tmp)
for _order in order_waiting:
model = SARIMAX(self.ts,
order=_order[:3],
seasonal_order=_order[3:] + [self.period],
enforce_stationarity=False)
results = model.fit(low_memory=True)
bic_value_ = pd.DataFrame([_order, results.bic], columns=['seasonal_order', 'bic'])
self.bic_value = pd.concat([self.bic_value, bic_value_])
new_bic = min(new_bic, (results.bic, _order))
if new_bic < bic:
bic = new_bic
new_bic = (np.inf, None)
seasonal_order = tuple(bic[1])
print(f"{bic[1]} is chosed with bic={bic[0]}")
else:
break
self.order=seasonal_order
print(f"best order is {self.order}")
def test_seasonal_arima4(self):
ts_data = self.getData5()
f_name='seasonal_arima4.pmml'
model = SARIMAX(endog = ts_data,
order = (1, 0, 1),
seasonal_order = (1, 0, 1, 12),
)
result = model.fit(disp=False)
StatsmodelsToPmml(result, f_name)
self.assertEqual(os.path.isfile(f_name),True)
def test_seasonal_arima2(self):
ts_data = self.statsmodels_data_helper.get_seasonal_data()
f_name = 'seasonal_arima2.pmml'
model = SARIMAX(endog=ts_data,
exog=None,
order=(3, 1, 1),
seasonal_order=(3, 1, 1, 12))
result = model.fit()
ArimaToPMML(result, f_name, conf_int=[80])
self.assertEqual(self.schema.is_valid(f_name), True)
def test_seasonal_arima2(self):
ts_data = self.getData5()
f_name = 'seasonal_arima2.pmml'
model = SARIMAX(
endog=ts_data,
order=(1, 0, 1),
seasonal_order=(1, 1, 1, 12),
)
result = model.fit(disp=False)
ArimaToPMML(result, f_name, conf_int=[95])
self.assertEqual(os.path.isfile(f_name), True)
def test_seasonal_arima1(self):
ts_data = self.statsmodels_data_helper.get_seasonal_data()
f_name = 'seasonal_arima1.pmml'
model = SARIMAX(endog=ts_data,
exog=None,
order=(3, 1, 1),
seasonal_order=(3, 1, 1, 12),
trend='c')
result = model.fit()
StatsmodelsToPmml(result, f_name)
self.assertEqual(self.schema.is_valid(f_name), True)
def SARIMA_f(self, df, pdq, s):
try:
sarima_mod = SARIMAX(np.array(df['Actual']), order=pdq, seasonal_order=s)
fit_sarima = sarima_mod.fit(use_boxcox=True, disp=0)
forecast = fit_sarima.forecast()[0]
Cluster, Warehouse, WF, YF = generate_attrib(df)
self.df_forecast.append({'Cluster':Cluster, 'Warehouse':Warehouse, 'Year':YF, "Week": WF, "Forecast":forecast})
return print(f'DEBUG:Forecast:{Cluster}:{Warehouse}:{YF}:{WF}:{forecast}')
except:
return print("ERROR:FORECAST-SARIMA")
def sarima_(self):
model = SARIMAX(self.df.iloc[:, 1], order=self.pdq_, seasonal_order=self.PDQ_) # 与上一句等价
print('the parameters: SARIMA{}x{}'.format(self.pdq_, self.PDQ_), '\n')
self.results = model.fit()
# joblib.dump(results, f'C:\\Users\\Administrator\\Desktop\\SARIMA模型.pkl')
self.predict_ = self.results.forecast(self.forecast_num)
fig, ax = plt.subplots(figsize=(20, 6))
ax = self.predict_.plot(ax=ax)
self.df.iloc[:, 1].plot(ax=ax)
plt.legend(['y_pred', 'y_true'])
plt.show()
def test_seasonal_arima5(self):
ts_data = self.getData5()
f_name = 'seasonal_arima5.pmml'
model = SARIMAX(
endog=ts_data,
order=(0, 0, 1),
seasonal_order=(3, 1, 1, 12),
trend='c',
)
result = model.fit(disp=False)
ArimaToPMML(result, f_name)
self.assertEqual(os.path.isfile(f_name), True)
def train_SARIMA_model(data, country=None):
sarima = SARIMAX(data,
order=ARIMA_ORDER,
seasonal_order=SARIMA_SEASONAL_ORDER)
sarima_model = sarima.fit()
if country:
pickle_file = os.path.join(
MODELS_DIRECTORY,
'sarima_' + country.replace(' ', '_') + '.pickle')
else:
pickle_file = os.path.join(MODELS_DIRECTORY, 'sarima.pickle')
with open(pickle_file, 'wb') as file:
pickle.dump(sarima_model, file)
log_train(TRAIN_LOG, 'sarima', pickle_file, data.size)
def train_SARIMA_model(data,
order,
seasonal_order,
directory_models,
country=None):
''' Train a seasonal auto-regressive, integrating, moving-average (SARIMA) model '''
sarima = SARIMAX(data, order=order, seasonal_order=seasonal_order)
sarima_model = sarima.fit()
if country:
sarima_model.save(directory_models + 'sarima_' + country + '.pickle')
else:
sarima_model.save(directory_models + 'sarima.pickle')
log_train('sarima', data.shape, {})
return sarima_model
def sarima(self):
model = SARIMAX(self.df.iloc[:, 0],
order=self.param,
seasonal_order=self.param_seasonal,
low_memory=True) #与上一句等价
print('the best parameters: SARIMA{}x{}'.format(
self.param, self.param_seasonal))
self.results = model.fit()
#joblib.dump(results, f'C:\\Users\\Administrator\\Desktop\\SARIMA模型.pkl')
self.predict_ = self.results.forecast(self.forecast_num)
fig, ax = plt.subplots(figsize=(30, 6))
ax = self.predict_.plot(ax=ax)
self.df.iloc[:, 0].plot(ax=ax)
plt.legend(['y_pred', 'y_true'])
plt.show()
def sarima(data, steps):
model = SARIMAX(endog=data.values,
order=(2, 0, 1),
seasonal_order=(0, 1, 1, 7),
enforce_invertibility=False)
sarima_fit = model.fit()
print(sarima_fit.summary())
# Rollling Forecast
# Number of days to Forecast Parameter
end = int(0.2 * len(data))
values = data[:-end]
actual_values = data[len(data) - end:]
pred_values = []
indexes = data[len(data) - end:].index
for i in range(end):
model = ARIMA((values), (2, 0, 1))
arima_fit = model.fit()
fnext = arima_fit.forecast()[0][0]
pred_values.append(fnext)
values = data[:-end + i]
pred_values = pd.Series(pred_values)
pred_values.index = indexes
#Doubt
#pred_values=pred_values.shift(-1)[:]
rmse = rms(actual_values, pred_values)
# Needs correction ??
print("RMSE VALUE", rmse)
#print(actual_values,pred_values)
print(len(pred_values))
return {
"model": "Baseline",
"index": list(indexes),
"actual": list(actual_values.values),
"predicted": list(pred_values),
"rmse": rmse
}
def params_select(self):
self.p = range(0, self.p_max)
self.q = range(0, self.q_max)
d = range(0, self.d_max)
# Generate all different combinations of p, q and q triplets
pdq = list(itertools.product(self.p, d, self.q))
# Generate all different combinations of seasonal p, q and q triplets
seasonal_pdq = [(x[0], x[1], x[2], self.period)
for x in list(itertools.product(self.p, d, self.q))]
aic_value = pd.DataFrame()
for i, param in enumerate(pdq):
for param_seasonal in seasonal_pdq:
model = SARIMAX(self.df.iloc[:, 0],
order=param,
seasonal_order=param_seasonal)
results = model.fit(low_memory=True)
print('SARIMA{}x{} - AIC:{}'.format(param, param_seasonal,
results.aic))
param_list = [[param, param_seasonal, results.aic]]
aic_value_ = pd.DataFrame(
param_list, columns=['param', 'param_seasonal', 'aic'])
aic_value = pd.concat([aic_value, aic_value_])
index_list = []
for i in range(self.p_max * self.q_max * self.p_max * self.q_max *
self.d_max * self.d_max):
index_list.append(i)
aic_value.index = index_list
min_index = aic_value[aic_value.aic == min(
aic_value['aic'])].index #找到aic值最小的行索引
a = aic_value.iloc[min_index, :]
a = a.values.tolist()
self.param = a[0][0]
self.param_seasonal = a[0][1]
# \end{align*}
# $$
#
# The parameters on deterministic terms are not directly comparable to
# `AutoReg` which evolves according to the equation
#
# $$
# (1-\phi(L)) y_t = x_t \beta + \epsilon_t.
# $$
#
# When $x_t$ contains only deterministic terms, these two representation
# are equivalent (assuming $\theta(L)=0$ so that there is no MA).
#
from statsmodels.tsa.api import SARIMAX
det_proc = DeterministicProcess(idx, period=52, fourier=2)
det_terms = det_proc.in_sample()
mod = SARIMAX(y, order=(1, 0, 0), trend="c", exog=det_terms)
res = mod.fit(disp=False)
print(res.summary())
# The forecasts are similar but differ since the parameters of the
# `SARIMAX` are estimated using MLE while `AutoReg` uses OLS.
sarimax_forecast = res.forecast(12, exog=det_proc.out_of_sample(12))
df = pd.concat([auto_reg_forecast, sarimax_forecast], axis=1)
df.columns = columns = ["AutoReg", "SARIMAX"]
df
def get_feature_SARIMA_residuals(time_series):
predict = SARIMAX(time_series, trend='n').fit().get_prediction()
return time_series - predict.predicted_mean
rho = 0.8
beta = 10
epsilon = eta.copy()
for i in range(1, eta.shape[0]):
epsilon[i] = rho * epsilon[i - 1] + eta[i]
y = beta + epsilon
y = y[200:]
from statsmodels.tsa.api import SARIMAX, AutoReg
from statsmodels.tsa.arima.model import ARIMA
# The three models are specified and estimated in the next cell. An AR(0)
# is included as a reference. The AR(0) is identical using all three
# estimators.
ar0_res = SARIMAX(y, order=(0, 0, 0), trend="c").fit()
sarimax_res = SARIMAX(y, order=(1, 0, 0), trend="c").fit()
arima_res = ARIMA(y, order=(1, 0, 0), trend="c").fit()
autoreg_res = AutoReg(y, 1, trend="c").fit()
# The table below contains the estimated parameter in the model, the
# estimated AR(1) coefficient, and the long-run mean which is either equal
# to the estimated parameters (AR(0) or `ARIMA`), or depends on the ratio of
# the intercept to 1 minus the AR(1) parameter.
intercept = [
ar0_res.params[0],
sarimax_res.params[0],
arima_res.params[0],
autoreg_res.params[0],
]
# # * Specification of seasonal and nonseasonal AR and MA components # * Inclusion of Exogenous variables # * Full maximum-likelihood estimation using the Kalman Filter # # This model is more feature rich than `AutoReg`. Unlike `SARIMAX`, # `AutoReg` estimates parameters using OLS. This is faster and the problem # is globally convex, and so there are no issues with local minima. The # closed-form estimator and its performance are the key advantages of # `AutoReg` over `SARIMAX` when comparing AR(P) models. `AutoReg` also # support seasonal dummies, which can be used with `SARIMAX` if the user # includes them as exogenous regressors. from statsmodels.tsa.api import SARIMAX sarimax_mod = SARIMAX(ind_prod, order=((1, 5, 12, 13), 0, 0), trend="c") sarimax_res = sarimax_mod.fit() print(sarimax_res.summary()) sarimax_params = sarimax_res.params.iloc[:-1].copy() sarimax_params.index = res_glob.params.index params = pd.concat([res_glob.params, sarimax_params], axis=1, sort=False) params.columns = ["AutoReg", "SARIMAX"] params # ## Custom Deterministic Processes # # The `deterministic` parameter allows a custom `DeterministicProcess` to # be used. This allows for more complex deterministic terms to be # constructed, for example one that includes seasonal components with two # periods, or, as the next example shows, one that uses a Fourier series
def fit(self, X, y=None):
# Perform top percentile ceiling
self.X = X
mode = self.mode
if '*f' in self.mode:
self.X = np.minimum(X, np.percentile(X, 75))
mode = self.mode.partition('*f')[0]
# Perform transformation if specified by *transformation
if '*ln' in self.mode:
self.X = np.log(np.array(X) + 1)
mode = self.mode.partition('*ln')[0]
elif '*bc' in self.mode:
transformer = pm.preprocessing.BoxCoxEndogTransformer()
self.X = transformer.fit_transform(y=X)
self.transformer = transformer
mode = self.mode.partition('*bc')[0]
try:
if mode == 'll':
# Local Level
model = LocalLevel(self.X)
self.res_ = model.fit(disp=False)
self.k_exog = None
elif mode == 'lla':
endog = X[2:]
exog = np.column_stack((X[1:-1], X[:-2]))
self.k_exog = exog.shape[1]
model = UnobservedComponents(endog=endog,
exog=exog,
level='local level')
self.res_ = model.fit(disp=False)
elif mode == 'lls':
self.k_exog = None
model = SARIMAX(endog=self.X,
order=(2, 0, 0),
trend='c',
measurement_error=True)
self.res_ = model.fit(disp=False)
elif mode == 'llt':
# Local Linear Trend
model = UnobservedComponents(endog=self.X,
level='local linear trend')
self.res_ = model.fit(disp=False)
elif mode == 'llc':
# Local Level Cycle
model = UnobservedComponents(endog=self.X,
level='local level',
cycle=True,
stochastic_cycle=True)
self.res_ = model.fit(disp=False)
elif mode == 'arima':
self.res_ = pm.auto_arima(self.X,
start_p=1,
start_q=1,
start_P=1,
start_Q=1,
max_p=5,
max_q=5,
max_P=5,
max_Q=5,
seasonal=True,
stepwise=True,
suppress_warnings=True,
D=10,
max_D=10,
error_action='ignore')
elif mode == 'rw1':
# For RW model
self.res_ = None
self.converged = False
except np.linalg.LinAlgError:
# Some kalman filter error ==> Use random walk
print(f'Convergence failed for {mode}')
self.converged = False
return self
try:
self.converged = self.res_.mle_retvals['converged']
except AttributeError:
if mode == 'arima':
self.converged = True # auto ARIMA from pmdarima should always converge
return self
fig, axes = plt.subplots(1, 2, figsize=(15, 6))
fig = sm.graphics.tsa.plot_acf(bike['seasonal_diff'].iloc[12:],
lags=36,
ax=axes[0])
fig = sm.graphics.tsa.plot_pacf(bike['seasonal_diff'].iloc[12:],
lags=36,
ax=axes[1])
# Create a dataframe to iterate over potential values of p and q for the ARIMA(p,d,q)(P,D,Q)m
# model in order to select the optimal model (lowest AICc value). Note that we are not iterating
# over different values of the seasonal parameters.
sarima_models = pd.DataFrame(np.zeros((3, 2), dtype=float))
for p in range(3):
for q in range(2):
fit_sarima = SARIMAX(bike['num_rides'],
order=(p, 0, q),
seasonal_order=(1, 1, 0, 12)).fit()
try:
sarima_models.iloc[p, q] = fit_sarima.aicc
except:
sarima_models.iloc[p, q] = np.nan
sarima_models
# The model with the lowest AICc had p = q = 1, so we fit an ARIMA(1,0,1)(1,1,0)12 model
fit_sarima = SARIMAX(bike['num_rides'],
order=(1, 0, 1),
seasonal_order=(1, 1, 0, 12)).fit()
fit_sarima.summary()