src/description.md

## Data Format
- A table (**CSV, XLSX, or Parquet**), **one column per time series**.
- For uploads, use **First row contains column names** when the first row is a header.
  Otherwise columns are auto-named `Series 0, 1, …`.
- Optionally select a **Time column** before parsing. It is removed from the forecastable
  series and shown as real dates on the time axis. Leave it blank to index by step.
- Series longer than the model's context window keep their **most recent** steps.
- Pick one **Target series** to forecast, then optionally add multiple aligned **Covariates**.
- Demo examples include future-known covariates and preselect them for comparison against a
  univariate TiRex baseline.

### Modes
- **Forecast starts at time index** controls the boundary between context and prediction.
  The model observes all selected target values before that index. Choose an earlier index
  to backtest against known future target values, or choose the final index to forecast after
  the available data.

### Covariates (in *Series*)
- Any series you **don't** pick as the target can be added under **Data covariates** — known
  drivers (promotions, price, weather, related series…) that inform the forecast but are not
  forecast themselves.
- **Known into the future** (default): covariate values are provided from the context window
  through the full forecast horizon. This is required for scheduled/known-ahead drivers.
- **History only (past)**: covariates condition on history only; targets are forecast *beyond*
  the chosen start using target history and covariate history only.
- When covariates are used, the forecast tab also runs a univariate TiRex baseline and overlays
  its median forecast for comparison.

## About TiRex-2

TiRex-2 is a zero-shot multivariate forecasting model from NXAI. It forecasts multiple target
variates out of the box and natively conditions on **past** and **known** covariates
(calendar features, holidays, promotions, scheduled interventions). It outputs **quantiles**
forecasts, so every prediction comes with a calibrated uncertainty band.

## Citation

```bibtex
@misc{podest2026tirex2generalizingtirexmultivariate,
    title = {{TiRex-2}: Generalizing TiRex to Multivariate Data and Streaming},
    author = {Patrick Podest and Marco Pichler and Elias B\\\"urger and Levente Z\\\"olyomi and Bernhard Voggenberger and Wilhelm Berghammer and Daniel Klotz and Sebastian B\\\"ock and G\\\"unter Klambauer and Sepp Hochreiter},
    year = {2026},
    eprint = {2607.01204},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG},
    url = {https://arxiv.org/abs/2607.01204},
}
```