1. Define the time series and its decision.
Use one stable part identity, one unit of measure, one location or pooled scope, and equally spaced periods. Record whether the forecast supports budgeting, repair planning, replenishment, or another decision. Do not mix planned overhaul demand with random corrective demand unless that is the declared model boundary.
2. Prove that a zero could have been observed.
Sales or issues are a censored view of demand when inventory was unavailable. Use the Stockout-Censored Demand Auditor to expose those periods. Recover backorders, unfilled requests, substitutions, emergency buys, or stockout timing; never apply a universal uplift to invent lost demand.
3. Describe the demand pattern.
The Pattern Classifier reports average demand interval and squared coefficient of variation for nonzero sizes. The familiar smooth/intermittent/erratic/lumpy matrix is a descriptive convention. It is useful for challenging assumptions, but it does not prove which forecast will perform best.
4. Build more than one baseline.
The Croston, SBA & TSB Calculator keeps the historical mean beside the specialized methods. Croston smooths demand size and interval separately; SBA reduces Croston's bias; TSB updates occurrence probability every period and can decay through long zero runs. A model name is not evidence of superiority.
5. Recreate past decisions without future leakage.
Use the Rolling Backtest Comparator. At each origin, fit only data already known and forecast the next period. Review both MAE and signed bias. Repeat across comparable SKUs and decision horizons because a winner on one sparse series can be luck.
6. Keep stocking logic separate.
A demand-level forecast is not a reorder point. Lead time, service objective, outage consequence, installed base, repairability, alternative parts, MOQ, shelf life, and budget still matter. The existing MRO planning workbench handles parts of that later decision, with its own assumptions.
7. Record and evaluate judgmental changes.
Planner knowledge can add information the model cannot see, such as a scheduled outage or an approved phase-out. It can also add optimism, duplicate information, or political bias. After actuals arrive, use the FVA Auditor to compare the untouched baseline with the released forecast, especially on periods that were changed.
8. Treat obsolescence as a review, not an algorithmic verdict.
TSB's occurrence probability decays when no demand occurs, which is useful evidence for an obsolescence review. It does not know whether the part protects a rare but catastrophic failure, whether the installed base remains, or whether a superseding part exists. Confirm configuration, consequence, supplier, and engineering obligations before disposition.
Minimum governance record
- Part, site/pool, unit of measure, period frequency, and data window.
- Availability and censoring evidence.
- Planned/unplanned demand treatment and known structural changes.
- Candidate method, parameters, and untouched baseline.
- Rolling-origin window, horizon, error, and bias.
- Override value, reason, owner, timestamp, and information available then.
- Actual demand and later FVA review.
- Separate stocking-policy approval and consequence review.
Public sources
- Forecasting: Principles and Practice — Count time series explains Croston's separate nonzero-size and interval estimates, plus bias and interval limitations.
- Forecasting: Principles and Practice — Time-series cross-validation defines rolling forecast origins without future leakage.
- SAP HANA PAL Croston TSB documentation publishes the TSB size/probability update and obsolescence-risk interpretation.
- Fildes et al., Forecast value added in demand planning evaluates judgmental changes through error and bias.
- Mersereau, Demand Estimation from Censored Observations explains why sales under stockout are a censored representation of demand.
- HELP Logistics CHORD demand-planning report documents ADI/CV² demand categorization in a spare-parts context.
ReliabilityBench implements transparent educational methods and declared conventions. It does not estimate service levels, certify inventory adequacy, or authorize disposal of critical spares.