Introduction

Intermittent Demand Forecasting is a dedicated production track for SKUs with frequent zeros. You should start with Croston-family baselines (Croston/SBA/TSB), then expand to zero-inflated count time-series models only when the data-generating mechanism demands it.

  • TL;DR: Define what “zero” means (true no-demand vs stockout/censoring vs missing), split the pipeline into an intermittent track, and validate with inventory KPIs (service level/cost), not just forecast scores.

Why it matters: In intermittent SKUs, average accuracy can look fine while stockouts/overstock explode in a small subset of items.

Prerequisites

  • Data contract fields to disambiguate zeros (stockout/availability, discontinued signals)
  • Rule-based segmentation (e.g., ADI/zero-rate/consecutive zeros)
  • Fixed baselines: Croston, SBA, TSB

Why it matters: If censored zeros are treated as true zeros, the model learns the wrong world and replenishment collapses.

Step-by-step Production Design

1) Segment intermittent SKUs first

Use simple operational rules (ADI/zero-rate/consecutive zeros) to create an “intermittent track.”

Why it matters: One-size-fits-all modeling makes root-cause analysis and governance impossible.

2) Run Croston-family baselines as mandatory benchmarks

  • Croston (1972): smooth size + inter-arrival time
  • SBA (2005): bias-reduction adjustment
  • TSB (2011): update occurrence probability each period; useful for obsolescence-like patterns

Why it matters: Without strong baselines, “fancier” models can’t be trusted or debugged.

3) Expand to zero-inflated count time series only when needed

Zero-inflated Poisson-based state-space approaches explicitly model excess zeros and uncertainty in demand patterns.

Why it matters: Intermittent demand is often better served by distribution-aware forecasting linked to inventory decisions.

Verification

  • Evaluate by slices (top SKUs, long lead-time, consecutive-zero candidates).
  • Validate final selection by inventory KPIs (service level/cost) and replenishment simulations (Part 6 concept).

Why it matters: Forecast score improvements that don’t translate to service level/cost are not production wins.

Troubleshooting

  • Forecast stuck at zero → censored zeros/stockouts not handled
  • Slow decay after demand stops → consider TSB + explicit discontinued signals
  • Accuracy up, inventory worse → wrong metrics; need distribution/lead-time demand evaluation

Conclusion

  • Build an intermittent track with clear zero semantics.
  • Benchmark Croston/SBA/TSB; then extend to zero-inflated count time series if needed.
  • Prove success with service level/cost, not only accuracy.

Summary

  • Define “zero” explicitly and enforce it in the data contract.
  • Croston/SBA/TSB are mandatory baselines for intermittent SKUs.
  • Use inventory KPIs to validate production impact.

#IntermittentDemandForecasting #AISalesForecasting #Croston #SBA #TSB #ZeroInflation #InventoryOptimization #TimeSeries

References

  • (Forecasting and Stock Control for Intermittent Demands, 1972)[https://www.jstor.org/stable/3007885]
  • (2015 Intermittent demand forecasting.pdf, 2015)[https://www.bauer.uh.edu/egardner/3301H%20Operations%20Management/ESG%20Publications/2015%20Intermittent%20demand%20forecasting.pdf]
  • (Bayesian forecasting of zero-inflated time-series of counts, 2026)[https://www.sciencedirect.com/science/article/abs/pii/S0169207025001220]
  • (개수를 세서 만든 시계열 | Forecasting, 2026-02-11)[https://otexts.com/fppkr/counts.html]
  • (Intermittent demand forecasting: a guideline for method, 2026)[https://www.ias.ac.in/public/Volumes/sadh/045/00/0051.pdf]
  • (Forecasting intermittent inventory demands: simple parametric methods, 2015)[https://www.sciencedirect.com/science/article/abs/pii/S0148296315001496]
  • (AI Sales Forecasting 4: 피처 기반 ML로 판매 예측 설계, 2026)[https://royzero.tistory.com/entry/ai-sales-forecasting-part-4-feature-based-ml]
  • (The accuracy of intermittent demand estimates, 2005)[https://www.sciencedirect.com/science/article/abs/pii/S0169207004000792]
  • (Intermittent demand: Linking forecasting to inventory obsolescence, 2011)[https://www.sciencedirect.com/science/article/abs/pii/S0377221711004437]
  • (CrostonSBA Model - Nixtla, 2026-02-11)[https://nixtlaverse.nixtla.io/statsforecast/docs/models/crostonsba.html]
  • (AI Sales Forecasting 백테스트 설계: Rolling CV·베이스라인, 2026)[https://royzero.tistory.com/entry/ai-sales-forecasting-backtesting-part-3]
  • (AI Sales Forecasting 6: 서비스레벨·안전재고·ROP 설계, 2026)[https://royzero.tistory.com/entry/ai-sales-forecasting-to-replenishment-part-6]
  • (AI Sales Forecasting 7: 운영(MLOps) 설계—모니터링·드리프트, 2026)[https://royzero.tistory.com/entry/ai-sales-forecasting-part-7-mlops-monitoring]