`nexyron.forecast_intervals`

Generic Description

nexyron.forecast_intervals returns calibrated conformal forecast intervals for one numeric target in an abstraction dataset.

CALL nexyron.procedures()
YIELD name, description, parameters, output_columns
WHERE name = 'nexyron.forecast_intervals'
RETURN name, description, parameters, output_columns

Consumer-Level Explanation

Use this procedure when a forecast needs a planning range, not only a point estimate. The procedure uses the abstraction dataset's temporal target history, builds one-step residuals from prior observations, and wraps the next point forecast with a split-conformal interval.

Parameters

• dataset: Dataset name. Required.
• target_feature: Numeric target feature to forecast. Optional when the dataset already declares a target feature or label.
• horizon: Forecast horizon label such as next_period or 30d. Optional.
• confidence: Requested interval confidence between 0 and 1. Optional, default 0.8.
• method: Interval method. Optional. The initial supported runtime method is split_conformal.
• calibration_split: Optional split label used for conformal residual calibration.
• at: Optional as-of timestamp in epoch milliseconds for bound dataset execution.
• system_at: Optional system-time cutoff in epoch milliseconds for dual-time replay.
• prefer_materialized: When true, reuse a matching materialized dataset before rebuilding it. Optional.

Output Columns

• dataset
• target_feature
• horizon
• horizon_start
• horizon_end
• point_forecast
• lower_bound
• upper_bound
• confidence
• method
• calibration_count
• empirical_coverage
• interval_width
• source
• warnings

Conceptual Explanation

The procedure is additive to existing point-forecast and Lens forecasting behavior. It does not create a new forecasting store or bypass abstraction datasets. Lens forecasting modules use the same idea to save calibrated interval findings, forecast_chart_candidates, warnings, and coverage metadata in Lens analysis artifacts before the unified report writer receives compact Lens evidence.

Limitations

• The initial procedure expects numeric temporal target observations.
• Small calibration windows return warnings because empirical coverage is weak.
• Intervals are calibrated ranges from observed residual behavior, not guarantees.