Predict precipitations in Switzerland by utilizing the forecasting horizon as a characteristic.
The normal method
After we need to predict future values for a time-series, we are sometimes excited about a number of future horizons, e.g. what’s going to occur in 1, 2, or 3 months. The normal method to foretell these totally different horizons consists in coaching a separate mannequin for every goal horizon.
Widespread alternate options
A typical different consists in coaching a single mannequin on a brief horizon, and prolong it to multi-horizons by making use of it recursively (i.e. by taking the earlier predictions as inputs to supply the next ones). Nonetheless this method might be complicated to implement in manufacturing programs, and it might result in error propagation: an error on an in depth horizon can have detrimental results for the next ones.
One other different consists in forecasting all of the horizons on the similar time with a multi-variate mannequin. Nonetheless, the sort of fashions that help multi-variate outputs is restricted, and it requires additional effort in information dealing with and mannequin upkeep.
Horizon as a characteristic
An easier method consists in concatenating the information ready for every horizon, and including a brand new “horizon” characteristic. This method has a number of benefits:
- It’s easy to know and implement, because it results in a single mannequin to coach and preserve.
- It doubtlessly improves the predictions accuracy, because the mannequin is skilled on a bigger dataset. It may possibly even be used as a “information augmentation” method: if you’re excited about just a few horizons, you possibly can nonetheless add extra ones within the coaching part to enhance mannequin estimation.
- The mannequin can be utilized to foretell horizons on which it was not skilled, which is likely to be useful when you’ve got many horizons to foretell.
This method is the alter-ego of a global model, however within the context of a number of horizons as a substitute of a number of…
