I am trying to implement a forecasting model, and have followed this Medium guide. I have changed the code minimally, to get it working with the most recent version of skforecast (e.g. changing ForecasterAutoreg to ForecasterRecursive). When I use the dataset that they use, everything works, including the end with the exogenous features.

I want to use this implementation to train with my own data, and specifically using exogenous features. When using the code from the guide, but changing the dataframe to have my own data (with the exact same date-range), everything still work, except when I use exogenous features. Then I get the following error: TypeError: unsupported operand type(s) for +: 'Timestamp' and 'NoneType'.

This is regarding this codeblock

forecaster_exog = ForecasterAutoreg(
    regressor = DecisionTreeRegressor(random_state = 123),
    lags = 30
)
# Model Fit
forecaster_exog.fit(y = df_exog.loc[train_start:train_end, 'y'],
               exog = df_exog.loc[train_start:train_end, ['exog_1', 'exog_2']]
)
# Model Predict
predicted_test_exog = forecaster_exog.predict(steps = len(df.loc[test_start:test_end]),
                                              exog = df_exog.loc[test_start:test_end, ['exog_1', 'exog_2']])

# Visualize
fig, ax = plt.subplots(figsize=(7, 3))
df.loc[test_start:test_end].plot(ax=ax, label = "Test")
predicted_test_exog.plot(ax=ax, label = 'Predicted DT Exog')
ax.legend()

I have double checked that there are no dates missing in my data, both in the training and the test set. This is the case for the actual dates, the y value, and exog_1 and exog_2. The index is of type pandas.core.indexes.datetimes.DatetimeIndex.

Does anyone know a feature of the data that might be the issue? Or an edit to the code as a workaround?

The full traceback of the error is below:

File <command-7029261884050460>, line 12
      9 # Model Predict
     10 # Set the dataset frequency to be (D)aily data
     11 df = df.asfreq('D', method = 'bfill') 
---> 12 predicted_test_exog = forecaster_exog.predict(steps = len(df.loc[test_start:test_end]) if not df.loc[test_start:test_end].empty else 0,
     13                                               exog = df_exog.loc[test_start:test_end, ['exog_1', 'exog_2']])
     15 # Visualize
     16 fig, ax = plt.subplots(figsize=(7, 3))
File /local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.12/site-packages/skforecast/recursive/_forecaster_recursive.py:1437, in ForecasterRecursive.predict(self, steps, last_window, exog, check_inputs)
   1398 def predict(
   1399     self,
   1400     steps: Union[int, str, pd.Timestamp],
   (...)
   1403     check_inputs: bool = True
   1404 ) -> pd.Series:
   1405     """
   1406     Predict n steps ahead. It is an recursive process in which, each prediction,
   1407     is used as a predictor for the next step.
   (...)
   1433     
   1434     """
   1436     last_window_values, exog_values, prediction_index, steps = (
-> 1437         self._create_predict_inputs(
   1438             steps=steps,
   1439             last_window=last_window,
   1440             exog=exog,
   1441             check_inputs=check_inputs,
   1442         )
   1443     )
   1445     with warnings.catch_warnings():
   1446         warnings.filterwarnings(
   1447             "ignore", 
   1448             message="X does not have valid feature names", 
   1449             category=UserWarning
   1450         )
File /local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.12/site-packages/skforecast/recursive/_forecaster_recursive.py:1149, in ForecasterRecursive._create_predict_inputs(self, steps, last_window, exog, predict_boot, use_in_sample_residuals, use_binned_residuals, check_inputs)
   1142     steps = date_to_index_position(
   1143                 index        = last_window.index,
   1144                 date_input   = steps,
   1145                 date_literal = 'steps'
   1146             )
   1148 if check_inputs:
-> 1149     check_predict_input(
   1150         forecaster_name  = type(self).__name__,
   1151         steps            = steps,
   1152         is_fitted        = self.is_fitted,
   1153         exog_in_         = self.exog_in_,
   1154         index_type_      = self.index_type_,
   1155         index_freq_      = self.index_freq_,
   1156         window_size      = self.window_size,
   1157         last_window      = last_window,
   1158         exog             = exog,
   1159         exog_type_in_    = self.exog_type_in_,
   1160         exog_names_in_   = self.exog_names_in_,
   1161         interval         = None
   1162     )
   1164     if predict_boot and not use_in_sample_residuals:
   1165         if not use_binned_residuals and self.out_sample_residuals_ is None:
File /local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.12/site-packages/skforecast/utils/utils.py:1033, in check_predict_input(forecaster_name, steps, is_fitted, exog_in_, index_type_, index_freq_, window_size, last_window, last_window_exog, exog, exog_type_in_, exog_names_in_, interval, alpha, max_steps, levels, levels_forecaster, series_names_in_, encoding)
   1027             raise TypeError(
   1028                 (f"Expected frequency of type {index_freq_} for {exog_name}. "
   1029                  f"Got {exog_index.freqstr}.")
   1030             )
   1032 # Check exog starts one step ahead of last_window end.
-> 1033 expected_index = expand_index(last_window.index, 1)[0]
   1034 if expected_index != exog_to_check.index[0]:
   1035     if forecaster_name in ['ForecasterRecursiveMultiSeries']:
File /local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.12/site-packages/skforecast/utils/utils.py:1585, in expand_index(index, steps)
   1581 if isinstance(index, pd.Index):
   1583     if isinstance(index, pd.DatetimeIndex):
   1584         new_index = pd.date_range(
-> 1585                         start   = index[-1] + index.freq,
   1586                         periods = steps,
   1587                         freq    = index.freq
   1588                     )
   1589     elif isinstance(index, pd.RangeIndex):
   1590         new_index = pd.RangeIndex(
   1591                         start = index[-1] + 1,
   1592                         stop  = index[-1] + 1 + steps
   1593                     )