In CH9, "res.forecast" should be used instead of "res.get_predicted_mean"

[ara-git]

Ch9 does not seem to be doing a good job of predicting one period ahead.
Although the code is trying to forecast GDP one period ahead of the train data, it seems to be predicting the last value of train data.

To fulfill the objective, "res.forecast()" should be used instead of "res.get_predicted_mean()".

For example, in the definition of "recursive_forecast", instead of

predictions = res.get_prediction(exog=exog)
oos_pred = predictions.predicted_mean.iloc[-window:]
pred_SARIMAX.extend(oos_pred)

,

predictions = res.forecast(step = 1, exog=exog.iloc[i])
pred_SARIMAX.extend(predictions)

would be work.

[adamscarlat]

I think that you're right. I found the same issue and was able to fix it using the following code (which I think does the same thing as forecast with step 1 in your example). I was able to get the SARIMAX MAPE down from 0.7 to 0.25 after this fix:

  elif method == "SARIMAX":
    pred_SARIMAX = []
    for i in range(train_len, total_len, window):
      model = SARIMAX(
        endog[:i], 
        exog=exog[:i], 
        order=order,
        seasonal_order=seasonal_order,
        simple_differencing=False
      ).fit(disp=False)

      # FIX: setting start and end to the get_prediction function
      oos_predictions = model.get_prediction(
        start=i,
        end=i + window - 1,
        exog=exog[i:i+window]
      ).predicted_mean.iloc[-window:]
      pred_SARIMAX.extend(oos_predictions)

    return pred_SARIMAX