Forecast Time Series (Chronos)

How it fits

Typical backends

Caveats

• Hard constraintThe target_column is parsed as strict numeric; cells that do not parse as a finite number are skipped, and a column with no numeric values aborts the call.
• LatencyHorizons longer than the checkpoint's native prediction length are produced autoregressively — supported but slower and progressively less sharp further out.
• Hard constraintThe Chronos checkpoint is supplied through the bound `model` file slot, not downloaded at runtime; bind a Chronos-family snapshot (`config.json` plus the safetensors weights, default `amazon/chronos-bolt-base`) to the slot at deploy, and the pipeline loads it offline.
• I/O contractWhen the target history is at least twice the longest scenario horizon the component holds out the final horizon and reports Weighted Quantile Loss as the result metric/score; shorter series skip the backtest and report score 0.
• I/O contractEach scenario's variables contribute a scalar offset added to every predicted step and its interval; variable names are not interpreted, only the magnitudes matter.