Forecast Modes
This page explains what the column roles in the forecast workbench do: how your data is handed to the model and how the result reports it. The workbench is in beta, and the model it offers is SequenxAI Omni.
1. A single series
A table with one "Time" column and one "Target" column is one series: the model reads the target's history and forecasts the next "horizon h" points.
With a "Series ID" column, each ID is a separate series, forecast on its own without looking at the others.
2. Several target columns
Set other numeric columns to "Extra target" in the header to forecast them together with the "Target" column. Each target column counts as one series.
Metering is per column: each target column counts as one series toward the per-run series limit, and its data points are counted separately (rows × target columns). The daily free quota still counts runs (one per model), whatever the number of target columns.
- Joint multi-target forecasting: a model that supports joint forecasting puts the target columns of one entity (the whole table without a series ID) into one inference so they can inform each other.
- Each column separately: when the model does not support joint forecasting, the workbench still runs and forecasts each target column as its own series. Before the run, the workbench says "This model does not support joint forecasting, so the target columns will be forecast separately."
In the result, each model carries a badge, "Joint" or "Per column"; hover it for the reason. SequenxAI Omni does not support joint forecasting yet, so several target columns are forecast separately.
3. Covariates
Covariates are numeric columns that help explain the target but are not forecast themselves. Pick one of these roles in the header:
| Role | Meaning | Examples |
|---|---|---|
| Covariate · past only | Values only in the history, recorded alongside the target | Past temperature, machine state |
| Covariate · known future | Values for the forecast period are known in advance and filled in the future rows at the end of the table | Holidays, shift plans, published weather forecasts |
When the model does not support covariates (or known-future covariates), those columns are not given to the model; the forecast still runs and the result marks them unused with the reason. SequenxAI Omni does not use covariates yet; see Covariate Support.
4. Holdout backtest
With "Holdout backtest" on, the last h points of your data are withheld: the model forecasts them from the earlier data only, and the forecast is compared with the withheld actuals to report error metrics. With it off, the whole series is used to forecast the future, so there are no actuals and no error metrics.
A backtest only shows how the model did on the most recent stretch of this data, not on other data or other periods.
5. Limits follow each model
History length, horizon, covariates and joint forecasting follow what the selected model actually supports:
- History longer than the model reads: not an error; the most recent points are used, and the workbench says how many.
- Horizon above the model's maximum: before the run, the workbench says the model does not support a horizon that long and gives its maximum.
- Covariates and joint forecasting: handled as described above when the model does not support them, and reported in the result.
SequenxAI Omni's figures are in Models & Limits.
Separately, the platform has compute limits that are not about the model: the data points per run (rows × target columns) and the upload size are capped; the figures are in the workbench's "Usage notes and limits". When you hit one, the workbench says it is a platform limit, not a model limit.