Shifting of Time Series data

[Elisa1234567890]

Hello,

i want to use FLAML to do a time series forecasting task. My dataset is structured as follows: timestamp, demand (output), several different exogenous features (input). One row in the dataset corresponds to the corresponding observations observed at the indicated timestamp. The exogenous features are past covariates that are only known up to the start of the forecast horizon (see Autogluon, past covariates).

As I predict the demand for a timestamp, the exogenous features are only available up to the previous timestamp. Since I am forced to include input variables in the FLAML.predict(), I am wondering whether i need to shift the time series manually beforehand such, that a row constists of the timestamp, the corresponding demand at this timestamp and the exogenous features shifted by the forecast horizon.

[prdai]

1. Shifted Model: Exogenous feature are shifted, an advantage is that it provides inputs that directly align with forecast on the other hand it reduces the dataset size.
2. Unshifted Model: Here, we jus use the exogenous features as they are originally presented, it retains the full dataset but the features may not align with the forecasting which may make it hard to align features with the forecasting data.

The above is the simply aspect of Shifted Model's and Unshifted model's in terms of time series data, so let me know if there are any other questions :)

[thinkall]

For FLAML’s time-series forecasting tasks, AutoML.predict() accepts either:

1. A pandas dataframe where the first column is the forecast timestamps, and any remaining columns are treated as exogenous variables; or
2. An integer “number of steps” only when the trained estimator is ARIMA or SARIMAX.

This means: if you train a model that uses exogenous variables, then at prediction time FLAML expects you to provide exogenous values aligned with the timestamps you are forecasting. If your exogenous features are “past covariates” (only known up to the start of the horizon), you generally cannot provide true x(t) for future timestamps t because you don’t have them.

So you should not shift data to make unknown future covariates look known. Instead, convert those exogenous variables into features that are available at forecast creation time (typically lagged versions such as x(t-1), x(t-2), etc.). Then, during prediction, you provide the forecast timestamps plus the lagged feature values computed from the last observed history.

If you want to avoid supplying exogenous inputs entirely, restrict to ARIMA/SARIMAX-style estimators and use the “steps” form (when applicable).