I have a time series with weekly frequency (about 250 consecutive observations). My task is to use this series to perform 6 months ahead forecasts. I am using Machine Learning to forecast, with feature embedding and a recursive strategy.
I am actually thinking about three possibilities:
- using the original weekly data > forecast 24 steps ahead > aggregate to monthly
- aggregate the input data to monthly > forecast 6 steps ahead
- keep the weekly frequency but transform all observations into the sum of the observation itself + the 3 previous ones (each date is a week representing the sum of 4 weeks values) > forecast 24 steps ahead > select only the dates I need
In my opinion:
Option 1. has the benefit of having more observations to train the model, but the error will be high since the steps ahead are many (24)
Option 2. has less error since the steps are less, but it has less data for training the model.
Option 3. has the same benefits as 1., the error might be lower since I am aggregating very volatile values and the aggregation can smooth the series, but I am afraid with this approach I am duplicating the data and therefore it won't be a good choice at all.
What do you think? Is there a better or alternative solution? I see I could even do both 1. and 2. and then perform optimal reconciliation
, but I read here that it's not guaranteed to increase the final accuracy and it might not be worth adding the complexity.