Schemas

Days.

Hours.

Minutes.

Months.

Milliseconds.

Nanoseconds.

Quarters (3 months).

Seconds.

Microseconds.

Weeks (7 days).

Years (12 months).

Long-format forecast dataframe: one row per (cutoff, horizon step).

A non-empty array of information cutoffs. For example ["latest"], an explicit date ["2025-01-31"], or a recurring schedule [{ "every": "3d" }].

A relative time interval, accepted in any of three forms:

• an integer — that many of the series' native frequency steps (e.g. 3);
• a duration string — either space-separated <n><unit> tokens ("2mo 3d 4h 5st"; units ns us ms s m h d w mo q y st, st = native step) or an ISO 8601 duration ("P1Y2M10D", "PT1H30M");
• an object with one field per unit ({ "d": 7, "steps": 1 }).

Components add independently.

Days.

Hours.

Minutes.

Months.

Milliseconds.

Nanoseconds.

Quarters (3 months).

Seconds.

Multiples of the host timeseries' native frequency.

Microseconds.

Weeks (7 days).

Years (12 months).

The dimensional columns — every identifier column declared by the schema (including hierarchy ancestors), keyed by source column name.

The targets

Lower bound of the range; omit for an open lower end.

Upper bound of the range; omit for an open upper end.

The typed columns of a timeseries family, keyed by column name. Exactly one column must be the time axis. For example:

Code
{
  "date": { "kind": "time", "frequency": "1d" },
  "store": "identifier",
  "sales": "target"
}

Where the family's rows come from: either an inline dataframe (a column-oriented map from column name to that column's values) or a URL string the service reads the dataframe from. For example, inline:

Code
{
  "date": ["2025-01-01", "2025-01-02"],
  "sales": [10, 12],
  "product": ["banana", "apple"],
}

Optional dimensional hierarchies, keyed by hierarchy name. Each value is a list of chains; a chain lists identifier column names from the top (coarsest) level down to the finest. For example { "geo": [["country", "region", "city"]] }.

A timeseries family: its rows and the typed columns describing them (identifiers, variates, targets, and the time axis).

How much history to feed the model before the information cutoff.

A non-empty array of information cutoffs. For example ["latest"], an explicit date ["2025-01-31"], or a recurring schedule [{ "every": "3d" }].

Gap between the information cutoff and the first prediction's target time.

Total time window forecasted, measured from the end of the lead time.

Quantile levels to forecast for every target, as numbers in [0, 1]; include "mean" to also request the mean. Omit for a default spread.

JSON selector for materializing a subset of a timeseries family.

The map is keyed by identifier column name. Each value is a pattern:

• "any" selects every value at that column.
• [value, ...] selects concrete values.
• { "from": value?, "to": value? } selects a bounded or half-bounded range.

An identifier column omitted from the map is aggregated over.

Target-time interval covered by each forecasted value (e.g. a span of 1d produces daily predictions, 3d produces 3-day buckets).