RelFMTabularForecaster

relationalai.semantics.reasoners.predictive

RelFMTabularForecaster(
    *,
    database: str,
    schema: str,
    context: b.Concept,
    label_column: str,
    time_column: str,
    max_context_length: int,
    item_id_column: Optional[str] = None,
    validation: Optional[b.Concept] = None,
    eval_metric: Optional[str] = None,
    test_batch_size: Optional[int] = None,
    stream_logs: bool = True,
    dataset_alias: Optional[str] = None,
    device: Literal["cpu", "cuda"] = "cpu",
    n_estimators: int = 8,
    random_state: Optional[int] = 42,
    sample_size: Optional[int] = None,
    sampling_strategy: str = "stratified",
    quantile_levels: Optional[List[float]] = None,
    norm_methods: Optional[Union[str, List[str]]] = None
)

Forecasting flat-table RelFM — see RelFMTabular.

forecast(domain=...) requires domain to carry pre-populated future (item_id, time) skeleton rows (no label column), same convention as RelFMTabularPredictor.predictions(domain=...). Output properties on prediction_concept: predicted_value (point estimate) plus, when quantile_levels is set, one quantile_<level> property per level (e.g. quantile_10/quantile_50/quantile_90 for [0.1, 0.5, 0.9]) — plain attribute access, no getattr needed.

The forecast lands exactly on domain’s own (item_id, time) rows — each series may ask for its own dates, at its own spacing. An item with no history in context, or a duplicate (item_id, time) row in domain, is refused before the job runs. A domain date that falls inside context’s own date range raises a leakage warning in get_status()["warnings"], since the model has already seen the true value for that date.

Parameters

• database

  (str) - Snowflake database to save predictions in.

• schema

  (str) - Snowflake schema to save predictions in.

• context

  (Concept) - History/panel Concept — naturally identified by (item_id, time) together (a composite identity, unlike RelFMTabularPredictor’s single natural key), carrying label_column among its own properties.

• label_column

  (str) - Name of the value property on context/validation.

• time_column

  (str) - Datetime property name on context/validation/the forecast() domain.

• max_context_length

  (int) - Max history length (in time steps) looked back over.

• item_id_column

  (str, default: None) - Series-partition property name for panel/multi-series forecasting. Omit for a single-series forecast.

• validation

  (Concept, default: None) - Optional labeled validation split.

• eval_metric

  (str, default: None) - One of "mae", "rmse", or "mape".

• test_batch_size

  (int | None, default: None)

• stream_logs

  (int | None, default: None)

• dataset_alias

  (int | None, default: None)

• device

  (int | None, default: None)

• n_estimators

  (int | None, default: None)

• random_state

  (int | None, default: 42) - See RelFMTabular.

• sample_size

  (int | None, default: 42) - See RelFMTabular.

• sampling_strategy

  (int | None, default: 42) - See RelFMTabular.

• quantile_levels

  (int | None, default: 42) - See RelFMTabular.

• norm_methods

  (int | None, default: 42) - See RelFMTabular.

Examples

relfm_forecast = RelFMTabularForecaster(
    database="MYDB",
    schema="MYSCHEMA",
    context=SalesHistory,
    label_column="sales",
    time_column="date",
    item_id_column="storeid",
    max_context_length=60,
)
FutureWindow.forecast = relfm_forecast.forecast(domain=FutureWindow)

Methods

.forecast()

RelFMTabularForecaster.forecast(domain: b.Concept) -> b.Relationship

Forecast future values for every (item_id, time) row of a domain Concept.

Context is context alone — if validation was also passed to the constructor, it is not automatically folded in as additional context (the SDK notebook this is modeled on builds two separate TabularTasks for score-context vs. predict-context, since context_table is fixed at task-definition time). Call score against validation separately if needed.

Parameters:

• domain

  (Concept) - Future (item_id, time) rows to forecast for — no label_column, but with time_column (and item_id_column, if this is a multi-series/panel forecast) present. See the class docstring: the actual dates forecast are determined by where context’s history ends, not by domain’s date values.

Returns:

• Relationship - A forecast relationship: {domain} -> prediction.

Inheritance Hierarchy