AutoStatsBank

AutoStatsBank validates a fixed bank of local statistical forecasting experts. Use it when you want deterministic model selection among reusable classical candidates without moving to the broader AutoForecaster panel selector.

When To Use

• The forecast question is local to each series.
• You want validation-based selection across statistical experts.
• The model should stay deterministic and easy to audit.
• You need a stronger local baseline before trying lagged or neural panel models.

Use AutoStatsBank as a selection layer, not as proof that a complex model is needed. Keep naive and seasonal naive in the comparison table.

Python Example

from cartoboost.forecasting import AutoStatsBank, ForecastFrame

frame = ForecastFrame.from_pandas(
    hourly_demand,
    timestamp_col="timestamp",
    target_col="demand",
    series_id_col="series_id",
    freq="h",
)

model = AutoStatsBank(
    season_length=24,
    validation_window=12,
)
model.fit(frame)
forecast = model.predict(12)

Browser WASM Example

AutoStatsBank browser forecast

Runs autostats_bank against a bundled route-demand sample.

Ready to run in this page.

Use When

Situation	Better first choice
You need a transparent last-value or seasonal baseline.	NaiveForecaster or SeasonalNaiveForecaster
You want one selected local statistical expert.	AutoStatsBank
You need lag features shared across many panels.	CartoBoostLagForecaster
You need guarded panel selection across lag, direct, intermittent, and classical candidates.	AutoForecaster

Validation

AutoStatsBank always reserves a non-empty suffix and fits every eligible expert on the earlier prefix. An explicit validation_window that leaves no training history is rejected rather than capped. Seasonal and window experts enter the roster only when the prefix satisfies their full configured history requirements; explicitly configured experts that fail to fit or omit a validation prediction produce an error instead of a partial score table.

Evaluate the selected bank under the same rolling-origin split as the individual local forecasters. Report the selected candidate, internal validation window, external test horizon, RMSE, MAE, WAPE, train time, prediction time, and the simple baselines it had to beat.

Limitations

• The bank only chooses among its configured statistical experts.
• Local experts need enough history for their seasonal and lag requirements.
• Internal validation reduces the history available for fitting on short series.
• Revalidate when panel frequency or demand regime changes.