def fit(self, modelarg: ModelArguments):
model_param = modelarg.model_param
ts_train = modelarg.endog[modelarg.train_start_date:modelarg.
train_end_date]
if 'window_size' in model_param:
window_size = model_param['window_size']
else:
window_size = len(ts_train)
ts_train = ts_train[-window_size:]
ARIMA_fit_results, min_aic, min_aic_fit_order, min_aic_fit_res = \
auto_arima.iterative_ARIMA_fit(ts_train, model_param['max_p'],
model_param['max_d'],
model_param['max_q'])
print("Selected Model Order=", min_aic_fit_order)
print("AIC score for selected model=", min_aic)
est_param = {
"train_size": window_size,
"order": min_aic_fit_order,
"aic": min_aic,
"result": min_aic_fit_res
}
modelres = ModelResults(model_name="ARIMA", estimated_param=est_param)
return modelres
def fit(self, modeleval: ModelEvaluation):
model_param = modeleval.model_param
ts_train = modeleval.ts_train
ts_test = modeleval.ts_test
# create daywise data
ts_dayofweek = ts_train.groupby(ts_train.index.dayofweek)
train_end_date = ts_train.index.max()
test_end_date = ts_test.index.max()
gap_and_test_dates = pd.date_range(
train_end_date + pd.Timedelta(days=1),
test_end_date)
gap_and_test_dates_dayofweek = gap_and_test_dates.groupby(
gap_and_test_dates.dayofweek
)
test_dates_dayofweek = ts_test.index.groupby(ts_test.index.dayofweek)
ts_pred = pd.Series()
for day, ts_day in ts_dayofweek:
ARIMA_fit_results, min_aic, min_aic_fit_order, min_aic_fit_res = \
auto_arima.iterative_ARIMA_fit(ts_day, model_param['max_p'],
model_param['max_d'],
model_param['max_q'])
print("Selected Model Order=", min_aic_fit_order)
print("AIC score for selected model=", min_aic)
indx_gap_test_dayofweek = gap_and_test_dates_dayofweek[day]
indx_test_dayofweek = test_dates_dayofweek[day]
forecast_res = min_aic_fit_res.forecast(
steps=len(indx_gap_test_dayofweek))
list_pred = forecast_res[0][-len(indx_test_dayofweek):]
ts_pred_sub = pd.Series(list_pred, index=indx_test_dayofweek)
ts_pred = ts_pred.append(ts_pred_sub)
print("Model info:")
print("Day =", day)
print("Min aic order =", min_aic_fit_order)
ts_pred = ts_pred.sort_index()
if "pred_days" in model_param:
pass
ts_pred.name = 'count'
ts_pred.index.name = 'date'
return ts_pred
def fit2(self, ts, max_p, max_d, max_q):
ts_dayofweek = ts.groupby(ts.index.dayofweek)
train_end_date = ts.index.max()
test_end_date = ts.index.max()
estimated_models = {}
for day, ts_day in ts_dayofweek:
ARIMA_fit_results, min_aic, min_aic_fit_order, min_aic_fit_res = \
auto_arima.iterative_ARIMA_fit(ts_day, max_p, max_d, max_q)
print("Selected Model Order=", min_aic_fit_order)
print("AIC score for selected model=", min_aic)
# print("TS: ", ts_day)
estimated_models[day] = {}
estimated_models[day]['min_aic_fit_order'] = min_aic_fit_order
estimated_models[day]['min_aic_fit_res'] = min_aic_fit_res
return estimated_models
def fit_arima(model_param, ts_train, ts_test):
ARIMA_fit_results, min_aic, min_aic_fit_order, min_aic_fit_res = \
auto_arima.iterative_ARIMA_fit(ts_train, model_param['max_p'],
model_param['max_d'],
model_param['max_q'])
# print("Min AIC=", min_aic)
# print("Order=", min_aic_fit_order)
train_end_date = ts_train.index.max()
test_end_date = ts_test.index.max()
gap_and_test_dates = pd.date_range(train_end_date + pd.Timedelta(days=1),
test_end_date)
forecast_res = min_aic_fit_res.forecast(steps=len(gap_and_test_dates))
list_pred = forecast_res[0][-len(ts_test):]
ts_pred = pd.Series(list_pred, index=ts_test.index)
ts_pred.name = 'count'
ts_pred.index.name = 'date'
return ts_pred
def fit_and_forecast(self, modeleval: ModelEvaluation):
model_param = modeleval.model_param
test_param = modeleval.test_param
ts_train = modeleval.ts_train
ts_test = modeleval.ts_test
list_pred = []
if test_param['sliding_window']:
print("Sliding window approach")
if test_param['window_size']:
window_size = test_param['window_size']
else:
window_size = len(ts_train)
if not test_param["look_ahead_step"]:
look_ahead_step = 1
else:
look_ahead_step = test_param["look_ahead_step"]
cur_window_data_points = ts_train[-window_size:]
# cur_window_data_points = list(ts_train[-window_size:].values)
print(len(ts_test))
for i in range(0, len(ts_test), look_ahead_step):
ARIMA_fit_results, min_aic, min_aic_fit_order, min_aic_fit_res = \
auto_arima.iterative_ARIMA_fit(cur_window_data_points,
model_param['max_p'],
model_param['max_d'],
model_param['max_q'])
print("Selected Model Order=", min_aic_fit_order)
print("AIC score for selected model=", min_aic)
forecast_res = min_aic_fit_res.forecast(steps=look_ahead_step)
list_pred.extend(forecast_res[0])
cur_window_data_points = cur_window_data_points[
look_ahead_step:]
temp_val = ts_test[i:i + look_ahead_step]
cur_window_data_points = cur_window_data_points.append(temp_val)
else:
ARIMA_fit_results, min_aic, min_aic_fit_order, min_aic_fit_res = \
auto_arima.iterative_ARIMA_fit(ts_train, model_param['max_p'],
model_param['max_d'],
model_param['max_q'])
print("Selected Model Order=", min_aic_fit_order)
print("AIC score for selected model=", min_aic)
train_end_date = ts_train.index.max()
test_end_date = ts_test.index.max()
gap_and_test_dates = pd.date_range(
train_end_date + pd.Timedelta(days=1),
test_end_date)
forecast_res = min_aic_fit_res.forecast(
steps=len(gap_and_test_dates))
list_pred = forecast_res[0][-len(ts_test):]
# ts_pred = pd.Series(list_pred, index=ts_test.index)
print(len(list_pred))
print(len(ts_test))
ts_pred = pd.Series(list_pred[:ts_test.size], index=ts_test.index)
ts_pred.name = 'count'
ts_pred.index.name = 'date'
return ts_pred
def fit2(self, ts, max_p, max_d, max_q):
ARIMA_fit_results, min_aic, min_aic_fit_order, min_aic_fit_res = \
auto_arima.iterative_ARIMA_fit(ts, max_p, max_d, max_q)
print("Selected Model Order=", min_aic_fit_order)
print("AIC score for selected model=", min_aic)
return min_aic_fit_res, min_aic_fit_order
