Skip to content

Forecast parameters

The Forecast policy carries what the AI library needs to predict an application's metrics. Its configuration reference for v0.1.3 lists the settings that come from the SAT. This page pairs each of them with the policy field it reads from.

The specification is not entirely finalised or fixed. Field names and types may change over the course of the project.

Policy field Library setting Conversion
metrics prediction_target Direct
resolution data_frequency Seconds to a duration string
forecast_duration horizon forecast_duration / resolution steps
lookback_duration window lookback_duration / resolution steps
history_duration parquet_train_size history_duration / resolution rows
retrain_interval retrain_frequency Lowered to a supported cadence
coverage conf_q_alpha 1 - coverage
pipeline pipeline_type Direct
monitoring STANDARD_METRICS Every metric the targets declare
monitoring poll_interval_seconds The finest collection frequency, capped at resolution

A SAT states periods in seconds. get_forecast() does the arithmetic and returns the settings above, flooring step and row counts to a minimum of one. The last two are read from the monitoring declarations of the policy's targets.