TSB Forecasting

TsbForecaster implements the Teunter-Syntetos-Babai intermittent-demand method. It smooths demand size and demand occurrence probability separately, which makes it useful when the chance of a non-zero event changes over time.

Interactive Example

TSB sparse-demand forecast

Runs tsb against a bundled taxi route-demand sample.

Ready to run in this page.

Python Example

from cartoboost.forecasting import TsbForecaster

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

model = TsbForecaster(alpha_demand=0.2, alpha_probability=0.1)
model.fit(demand)
forecast = model.predict(6)

Use When

Use TSB when zeros are real no-demand periods and the event probability itself is part of the signal, not just the spacing between past non-zero events.

Fit Contract

Requirement	Detail
Target values	Finite and non-negative.
Zero values	Treated as true no-demand periods.
Non-zero demand	At least one non-zero observation is required.
Demand smoothing	alpha_demand must be greater than 0 and at most 1.
Probability smoothing	alpha_probability must be greater than 0 and at most 1.

Validation

TSB can react differently from Croston and SBA when recent demand occurrence changes. Report the zero fraction and compare all three methods on the same rolling-origin split.

Tune both smoothing parameters using training-side rolling origins only. Report errors by horizon and separately inspect periods after demand occurrence changes.

Limitations

• TSB does not directly model seasonality, covariates, or related series.
• Two smoothing parameters increase selection risk on short histories.
• Zero values caused by missing data or supply constraints need preprocessing and a different interpretation.