Croston Forecasting

CrostonForecaster is a fixed intermittent-demand method for sparse, non-negative series where zero periods are true no-demand observations. It separately smooths non-zero demand size and the interval between non-zero events, then returns a flat forecast for the requested horizon.

Interactive Example

Croston sparse-demand forecast

Runs croston against a bundled taxi route-demand sample.

Ready to run in this page.

Python Example

from cartoboost.forecasting import CrostonForecaster

demand = [0, 0, 4, 0, 0, 7, 0, 3, 0, 0, 0, 5]

model = CrostonForecaster(alpha=0.2)
model.fit(demand)
forecast = model.predict(6)

Use When

Use Croston when demand is non-negative, most periods are genuine zeros, and you need an interpretable baseline before testing SBA, TSB, seasonal naive, or AutoForecaster. Do not use it merely because observations are missing; first restore the regular time grid and distinguish missing measurements from zero demand.

Fit Contract

Requirement	Detail
Target values	Finite and non-negative.
Zero values	Treated as real no-demand periods, not missing rows.
Non-zero demand	At least one non-zero observation is required.
Smoothing	alpha must be greater than 0 and at most 1.

Validation

Croston is transparent, but it is not bias-adjusted. Compare it against SbaForecaster and TsbForecaster on the same rolling-origin split before claiming it is the best sparse-demand model for a panel.

Report MAE or RMSE, WAPE where the aggregate denominator is non-zero, the zero fraction, and results by forecast horizon. Include seasonal naive whenever a calendar cycle is plausible.

Limitations

• Forecasts are flat across the requested horizon.
• The method does not model calendar seasonality, covariates, or cross-series effects.
• Long runs of zeros caused by stockouts or missing collection violate the demand interpretation.