Forecast API Schema Alignment

#Forecast API Schema Alignment

The TSFM.ai forecast API now uses a standardized wire format across /v1/forecast, /v1/forecast/batch, and /v1/forecast/ensemble.

This unifies single, batch, and ensemble forecasting around the same request and response shape, with better support for multivariate inputs, missing values, and probabilistic output.

#What changed

• inputs[].target is now always 2D with shape [num_timesteps][num_channels]. Univariate series use a single-value inner array such as [[120], [135], [142]].
• Missing observations are sent inline as null.
• Regularly spaced input timestamps now use start plus parameters.frequency instead of an explicit timestamps array.
• parameters.freq was renamed to frequency, and parameters.quantiles was renamed to quantile_levels.
• Forecast responses now return outputs instead of predictions, with quantile_predictions and optional samples for probabilistic models.
• Per-item metadata can be sent with each input and is echoed back on the corresponding output item.

#Migration notes

This is a breaking change for clients using the older forecast schema. If you previously sent 1D target arrays, item_id, series, or explicit input timestamps, update your integration to the new shape before your next deployment.

The forecasting docs now include updated examples for univariate, multivariate, and probabilistic requests so migrations are mostly mechanical.