Forecast

Generate multi-step time-series forecasts with SAP-RPT-TS. Unlike /api/predict, which fills [PREDICT] cells in tabular data, the forecast endpoint takes one or more numeric series and projects them forward over a fixed horizon.

Availability: This endpoint is gated behind a server feature flag. When it is disabled, the endpoint responds with 404 as though it did not exist. See API Overview for authentication details and all available endpoints.

Endpoint

POST /api/forecast

Base URL: https://rpt.cloud.sap

Authentication is the same across all endpoints — pass your personal API token in the Authorization header:

Authorization: Bearer YOUR_TOKEN

Obtain and manage tokens in Settings.

Request Body

{
  "horizon": 3,
  "data": [
    {
      "timestamp": ["2024-01-01", "2024-02-01", "2024-03-01", "2024-04-01"],
      "target": [120, 135, 128, 142],
      "past_covariates": { "temperature": [15, 18, 17, 21] },
      "future_covariates": { "temperature": [22, 24, 23] }
    }
  ],
  "options": {
    "confidence_interval": 0.9,
    "cross_learning": true
  }
}

Request Parameters

• horizon (required, integer > 0): Number of future steps to forecast. Capped at 90 — larger values are silently truncated.
• data (required): Non-empty array of series objects. Each series is forecast independently (unless cross_learning is enabled).
  • target (required): The historical values to forecast from. Either a flat array for a single target ([120, 135, 128]), or an array of arrays for multiple targets sharing the same timeline ([[120, 135], [40, 44]]). null marks a missing observation. When multiple target arrays are given, all must have the same length. Series longer than 8,192 points are silently truncated to the most recent 8,192.
  • timestamp (optional): Array of timestamps (strings or numbers), one per historical step. If present, its length must equal the target history length. Regularly-spaced timestamps are expected — a limited number of missing timestamps are filled internally, but filled timestamps may not exceed 20% of the supplied timestamps (rounded up); beyond that the request is rejected.
  • past_covariates (optional): Map of covariate name → array of historical values (numbers, strings, or null). Each array's length must match the target history length.
  • future_covariates (optional): Map of covariate name → array of values covering the forecast window. Each key must also appear in past_covariates, and each array's length must equal horizon.
• options (optional): Forecast tuning. See Options below.

Note: Unknown fields are rejected. The request schema is strict — any property not listed here causes a 400 validation error.

Options

All fields are optional.

rules[] — each rule overrides forecast values based on a covariate condition:

Response

A successful request returns 200 with the raw forecast result and the upstream latency in milliseconds:

{
  "forecast": {
    "all_null_targets": [],
    "confidence_interval_output": [
      [
        [
          [
            122.7403564453125,
            126.0583724975586,
            124.12074279785156
          ],
          [
            178.7145233154297,
            204.69876098632812,
            210.86856079101562
          ]
        ]
      ]
    ],
    "detected_interval": [
      {
        "unit": "month",
        "value": 1
      }
    ],
    "group_results": null,
    "hierarchical_result": null,
    "quantile_result": null,
    "result": [
      [
        [
          147.21084594726562,
          155.68292236328125,
          154.62445068359375
        ]
      ]
    ],
    "result_timestamp": [
      [
        "2024-05-01",
        "2024-06-01",
        "2024-07-01"
      ]
    ],
    "selected_quantile_result": null
  },
  "delay": 1800
}

Response Fields

• forecast: The forecast payload from the time-series model.
  • result: Nested arrays of forecast values, shaped [series][target][horizon].
  • result_timestamp: Projected timestamps per series (null when timestamps were not supplied).
  • detected_interval: The sampling interval inferred for each series (unit and value), or null.
  • all_null_targets: Series/target positions whose history was entirely null, reported as { series_index, target_indices }.
  • confidence_interval_output: Populated by default at 80% width (q=0.10 lower, q=0.90 upper). Set confidence_interval: null to suppress it; also null when point_forecast: "hierarchical_forecast" is used.
  • quantile_result, selected_quantile_result: null unless output_quantile: true is set.
  • hierarchical_result: Populated when point_forecast: "hierarchical_forecast" is used; contains reconciled forecasts keyed by label. result is [] in that case.
  • group_results: Always null (grouped forecast support is not implemented).
• delay: Upstream model response time in milliseconds.

Error Codes

Error responses use the same structured format as the rest of the API — see Error Codes.

When the feature flag is disabled the endpoint returns 404 with no forecast-specific body.

Constraints

The forecast endpoint has its own rate and burst limits, separate from /api/predict. For exact values see Limits & Quotas.

Exceeding the rate or burst limit returns 429 with a Retry-After header. See Errors for the rate-limit headers.