Why Part Identity, Consumption History,
and Lead-Time Accuracy
Come Before AI Forecasting
MRO Data Quality Before Spare Parts Demand Forecasting
A spare part can sit untouched for eleven months and then be drawn three times in a single maintenance window. Demand arrives when equipment fails, when a planned maintenance campaign triggers replacement, or when an operational event creates a requirement. Between those events, consumption is zero. Forecasting methods built for steady, continuous demand handle this pattern badly, and the pattern covers a large share of any MRO inventory.
Spare-parts forecasting faces two separate blockers, and projects stall when teams treat them as one. The first is the demand pattern: intermittent demand requires forecasting methods designed for sparse, irregular series. The second is the data foundation: duplicate part records, inconsistent manufacturer part numbers, missing consumption history, and lead-time records that reflect contracts rather than actual supplier performance can make even the correct method unreliable.
The pattern problem has established methods; the foundational baseline dates to 1972. The data problem requires engineering work with its own scope: master-data alignment, consumption-history repair, lead-time capture, and reporting modernization. That sequence, data first and method second, determines which parts are forecastable at all.
Stage 1: The Pattern Problem
Intermittent demand is common across spare-parts inventories: long periods of zero consumption interrupted by irregular, non-zero events. Research cited in intermittent-demand forecasting literature places this pattern at approximately half of spare-parts SKUs.
The figure comes from research cited in the forecasting literature (Nikolopoulos 2021, cited in arXiv:2304.03092) and should be read as a research estimate rather than audited market data. The exact proportion varies by industry and installed asset base. The pattern itself is structural. In practical terms, an intermittent-demand series shows:
- zero consumption across most recorded periods;
- irregular non-zero demand events with no fixed cycle;
- demand tied to equipment failures, maintenance schedules, shutdowns, and operating conditions;
- sparse part-level histories, because each individual SKU generates few demand observations.
Standard smoothing approaches struggle with series shaped this way. A forecast that averages across the many zero periods gets pulled toward zero and understates the clusters of demand that arrive during planned work. Because the non-zero events land irregularly, a method calibrated for smoother demand produces reorder parameters that overstock slow parts and run out of the parts a maintenance campaign actually consumes.
The forecasting literature developed dedicated baselines for exactly this structure.
Croston's method, developed in 1972, remains a foundational baseline for MRO spare-parts forecasting because it separates non-zero demand size from the interval between demand events. SBA and TSB variants extend that approach for known limitations in intermittent-demand series.
Croston's original formulation (Croston 1972, Operational Research Quarterly 23:289–304; reviewed in Shenstone & Hyndman) maintains two estimates instead of one: the size of demand when it occurs, and the interval between demand events. The Syntetos–Boylan Approximation introduced a correction for bias in Croston's original estimator (Syntetos & Boylan, International Journal of Forecasting, DOI: 10.1016/j.ijforecast.2004.10.001). The Teunter–Syntetos–Babai method updates the demand-probability estimate during zero-demand periods, which helps address obsolescence and changing demand probability (Teunter, Syntetos & Babai, European Journal of Operational Research, DOI: 10.1016/j.ejor.2010.09.012). Each of the three is a baseline that requires validation against part class, demand pattern, history depth, and planning objective before it earns a place in production planning.
Method selection also depends on part classification, and the research supports sequencing that step first.
De Paula Vidal et al. (2022) proposed an MRO inventory decision-support framework that combines SKU criticality ranking, demand forecasting, and a dashboard layer in a railway-logistics setting. The sequence supports a practical rule for MRO analytics: classify the part and demand pattern before selecting the forecasting method.
The framework (De Paula Vidal et al. 2022, Computers & Industrial Engineering, DOI: 10.1016/j.cie.2022.108777) was built and evaluated in railway logistics, so its specific results stay in that context. The sequencing logic is what transfers: rank the SKU, classify the demand, then attach the method.
The pattern problem, in short, has documented methods and a fifty-year literature behind them. The open question in most MRO organizations is whether the records can support any of them.
Stage 2: The Data Problem
Intermittent-demand methods make heavy demands on the data that feeds them. Each SKU produces few demand observations, so every lost, duplicated, or misattributed record removes a meaningful share of the usable signal. Practitioner reporting from the MRO sourcing field describes the typical starting state: inconsistent descriptions, missing specifications, duplicate records, and the same component appearing under different names across plants (SPARETECH). In SPARETECH's own MRO strategy research, vendor-conducted and attributed here as such, 22% of executives identified poor data quality as a major MRO challenge (SPARETECH spare-parts visibility report).
The data problems that prevent spare-parts forecasting from working in MRO environments include split part identities across ERP instances, vendor SKU and manufacturer part number conflicts, missing consumption history during shutdowns or migrations, and lead times recorded as contracted rather than actual.
Each of the four breaks forecasting through a specific mechanism, and each has a defined engineering remedy.
The reporting foundation: Synovos
Work of this kind starts with making operational data accessible enough to examine at all.
Softwarium's work with Synovos, now part of RS Integrated Supply, illustrates one part of the data foundation behind better analytics: replacing difficult-to-analyze Excel reporting files with a connected Power BI Embedded environment that improved analysis, pattern identification, and process adjustment.
Synovos ran operational reporting through Excel files that had grown to approximately 300 MB: slow to open, download, analyze, and share. Softwarium implemented a Power BI Embedded environment that replaced those files with connected, customized dashboards, improved analysis and business insight, supported pattern identification, and helped the organization adjust business processes (Softwarium case study). The engagement sits within Softwarium's Power BI development and data intelligence services.
The engagement scope covered reporting modernization and analytics access. It shows the kind of data-access foundation an analytics team needs in place before a forecasting or decision-support project can be scoped responsibly: operational data that people can actually open, query, and compare across sites, rather than reconstruct from spreadsheets. The buyer-side implication is direct: forecasting scope cannot be set responsibly while operational data remains trapped in files that teams struggle to open, share, or compare.
What a Data Readiness Assessment Covers
A readiness assessment determines which parts are ready for forecasting, which require remediation first, and which should remain under simpler planning rules until the records improve. Five questions structure it.
1. Part identity
Can each SKU be matched to a manufacturer part number? Can the same physical part be identified across sites, ERP instances, and procurement systems? Unresolved identity blocks everything downstream, because every history and every planning parameter attaches to a record that may be one of several.
2. Consumption-history completeness
Which part histories are complete, and which periods are missing? Are the gaps tied to shutdowns, migrations, manual workarounds, or system boundaries? A gap that can be annotated is manageable. A gap the model cannot see corrupts the forecast silently.
3. Lead-time accuracy
Do the recorded lead times reflect actual supplier performance or contracted targets? How wide is the spread between promised and actual delivery, and for which suppliers and part classes is the spread widest?
4. Demand classification
Which parts show intermittent, lumpy, erratic, smooth, or seasonal demand? ADI and CV² classification applies where the history supports it, and each demand class maps to suitable forecasting baselines. That is the classification step de Paula Vidal's framework places before method selection.
5. Data governance
Who creates material master records, and under what controls? Governance determines whether the remediated state holds. Without intake rules, new duplicates, naming variants, incomplete attributes, and unsupported supplier codes re-enter the system as fast as the old ones are resolved.
Each question produces a concrete deliverable. The part-identity review yields a duplicate map and a resolution plan by site and ERP instance. The history review yields a completeness report with annotated gaps, so a modelling team knows which zero periods carry demand signal and which record a system boundary. The lead-time review yields a variance table by supplier and part class, which corrects replenishment parameters even before any forecasting begins. The demand-classification pass segments the catalogue into classes with matched baseline methods. The governance review yields intake rules and ownership assignments that keep the remediated state from decaying.
The assessment output defines the realistic forecasting scope, and the segmentation usually lands in three groups. Some SKU classes have resolvable identities, adequate history, and stable lead-time records; those can enter a forecasting pilot immediately. A second group needs identity resolution, history repair, or lead-time correction first, which is software engineering and data engineering services work with definable scope and deliverables. A third group should stay under simple planning rules, such as min-max or reorder-point logic, until the records improve, because no forecasting method outperforms its inputs.
Once the readiness picture exists, the method decision follows. The choice between rules, machine learning, and mathematical optimization is covered in the Three-Method Test for supply chain workflows.
The Operational Principle
Reliable spare-parts forecasting follows a sequence: classify the demand pattern, assess the data foundation, remediate the records, then choose the forecasting method. The methods are documented, tested, and available. Intermittent-demand baselines have existed since 1972. The blocker, in most MRO environments, is the condition of the data that feeds them.
For supply chain and MRO teams preparing to evaluate or expand a forecasting programme, the first useful conversation covers what the data currently supports and which engineering work closes the gap. Softwarium's data engineering and dedicated development teams work in exactly that scope: part-identity alignment, consumption-history repair, lead-time capture, and the reporting foundation that makes the rest visible.
Source List
- Nikolopoulos (2021), cited in 'Combining Probabilistic Forecasts of Intermittent Demand'
https://arxiv.org/abs/2304.03092 - Croston, J.D. (1972), Operational Research Quarterly 23:289–304; reviewed in Shenstone & Hyndman
https://robjhyndman.com/papers/croston.pdf - Syntetos & Boylan, International Journal of Forecasting
https://doi.org/10.1016/j.ijforecast.2004.10.001 - Teunter, Syntetos & Babai, European Journal of Operational Research
https://doi.org/10.1016/j.ejor.2010.09.012 - De Paula Vidal, G.H. et al. (2022), Computers & Industrial Engineering
https://doi.org/10.1016/j.cie.2022.108777 - SPARETECH, MRO sourcing (practitioner)
https://sparetech.io/en/blog/mro-sourcing - SPARETECH, spare-parts visibility report (vendor-conducted research)
https://sparetech.io/en/blog/spare-parts-visibility-problem - Automa.net, standardizing MRO procurement data (practitioner)
https://automa.net/blog/1375/standardize-mro-procurement-data - Sharecat, MRO spare parts data challenges & guidance (practitioner)
https://www.sharecatdataservices.com/insights/spares-spir-challenges - Softwarium case study: Power BI integration & data visualization (Synovos)
https://www.softwarium.net/case-studies/power-bi-integration-data-visualization - RS Integrated Supply: RS Group unites IESA and Synovos
https://rs-integratedsupply.com/news/rs-group-unites-iesa-and-synovos-under-rs-integrated-supply/
