Are you an LLM? Read llms.txt for a summary of the docs, or llms-full.txt for the full context.
Skip to content

Configuring a Forecast Version

A version’s settings live in its Model Configuration block on the Forecast Version document. Edits in the document save automatically. Three things must be configured before you can run.

1. Select training data

In the Configure Training Data section, choose which ingestions (uploads/syncs from your sources) this version trains on. This is how you control which data a version uses — for example, a version that only looks at your Variable Costs source, or one that uses everything.

Each measure you forecast is tied to the ingestion that provides its history. When a measure comes from exactly one selected ingestion, Galdera assigns it automatically; otherwise pick it in the measure's row.

2. Choose measures and their strategy

In the measure table, select which measures this version forecasts. Each selected measure has a Strategy — four to choose from:

  • Independent (the default): the measure is forecast directly from its own history by the ML models.
  • Dependent (via Metric): the measure is not forecast directly. Instead, Galdera forecasts a metric that decomposes it and reconstructs the measure from that forecast. For example, forecasting Streaming Revenue dependently via Revenue per Viewer means the model forecasts the per-viewer ratio and rebuilds revenue as ratio × viewers — useful when the ratio is more stable or better understood than the raw measure. The picker only offers metrics that actually decompose that measure.
  • Roll forward: for stocks — balances like PP&E, inventory, or debt. The measure is not forecast by a model at all. Instead it starts from its anchor (the last actual balance) and accumulates the flows you pick as terms, each with a + or − sign (e.g. PP&E = last actual + capex − depreciation). The pane shows anchor coverage so you can confirm every slice has a starting balance.
  • Derived: the measure's forecast is computed each period from a formula over other targets, built with the same formula builder used for metrics — including references to the previous period (e.g. Additions = Capex × −1, or D&A = PP&E − previous-period PP&E). Like Roll forward, a Derived measure is never an ML target.

Roll forward and Derived exist so balance-sheet and cash-flow lines can be forecast in a way that keeps the statements consistent: a stock's change always equals its flows. See Modelling the Three Statements for a worked setup.

Two rules the configuration enforces:

  • At least one measure must remain Independent — the run needs at least one ML-forecast target, so the other strategies are unavailable on the last Independent measure.
  • No same-period loops. Definitions may reference each other across periods (that's how stocks accumulate), but a loop where every step is in the same period is rejected when you save, with the loop named so you can pick one direction.

3. Set the date range

  • Training start: how far back the model learns from history
  • Forecast start / end: the prediction window

Forecast start and end are required before running, and the forecast window must come after the training start.

Once configured, see Running a Forecast Version.

Cohorted measures in chat

A cohorted measure is loaded from a file with more than one date column: a cohort date and a report date. Configure its forecast, including its cohort settings, on the forecast version page. In both Ask and Agent mode, the assistant:

  • Does not add a cohorted measure to a version, change its forecast settings, or set cohort settings.
  • Does not describe a cohorted measure's forecast configuration; it names the measure and points to the version page instead.
  • In beta Agent mode, where available, can still change version-wide settings, such as the forecast dates, on a version that includes cohorted measures.

Configure through the assistant

In Agent mode, ask for a new named Forecast Version or a change to an existing version’s Model Configuration. For example: “In Annual Plan, change only the forecast end to December 31, 2027. Keep everything else and do not run it.” The assistant presents the resolved version and changes for Allow/Deny.

The server validates the resulting settings and references before saving. Unmentioned settings and canvas notes are preserved. Derived or roll-forward strategy changes can affect shared measure definitions; review the named effects in the confirmation. Saving configuration does not start a job. If an edit fails, the assistant must not run the old settings as a substitute.