Understanding Demand Forecasting Methods for Manufacturers

Demand forecasting methods manufacturers actually need in 2026 (and what your next ERP must support) Replacing legacy ERP often forces a hard truth: forecasting “in spreadsheets” doesn’t scale.

Demand forecasting methods manufacturers actually need in 2026 (and what your next ERP must support)

Replacing legacy ERP often forces a hard truth: forecasting “in spreadsheets” doesn’t scale once you have hundreds or thousands of SKUs, multiple plants, and volatile lead times. At the same time, 2025–2026 has made it clear that AI maturity is uneven. A Gartner survey (published June 2025) found only 23% of supply chain organizations reported having a formal AI strategy—even among leaders who had deployed AI in the prior 12 months. That’s why a practical buyer’s guide for SMEs should focus on methods that work reliably, and the ERP capabilities that make them repeatable.

This guide breaks down four forecasting method families manufacturers actually need—baseline time-series, intermittent/spare-parts methods, hierarchical reconciliation, and probabilistic forecasting—then maps them to ERP requirements (model orchestration, ML extensibility, governance/FVA for overrides, and KPI-safe accuracy metrics like WMAPE). The theme: tooling helps, but ROI is mostly determined by process and data foundations, a point echoed in BCG’s “Supply Chain Planning 2026: Why AI Alone Isn’t Enough”.

Method #1: Baseline time-series forecasting (your default for most A-items)

Baseline time-series models are the workhorses for steady or seasonal demand. Think exponential smoothing/ETS, ARIMA-style approaches, seasonal decomposition, and simple causal regression with calendar effects. These models are not “old-fashioned”—they’re often the best value for SMEs because they’re robust, explainable, and fast to re-fit.

What it’s good for

• Stable finished goods with weekly/monthly history
• Seasonal patterns (holidays, month-end effects)
• Products with modest promo impact where you can tag events

ERP capabilities to insist on

• Model orchestration: The ERP should run different models by segment (A/B/C, make-to-stock vs make-to-order) and choose the best based on backtests. Modern suites are explicitly moving this way—Microsoft’s demand planning supports multiple forecasting models, custom Azure ML models, and AI-based model selection (Dynamics 365 documentation, updated July 2026).
• Calendar + event management: Built-in calendars, holiday profiles, and event flags so the model can learn known demand shifts rather than treating them as “noise.”
• Automation with guardrails: Scheduled re-fitting (monthly/weekly) plus exception alerts for unusual demand changes.

Method #2: Intermittent (zero-heavy) demand forecasting for spare parts

If you manage MRO, service parts, or long-tail SKUs, intermittent demand is the norm: many periods of zero demand punctuated by occasional spikes. Standard MAPE and many “smooth” time-series models struggle here. Peer-reviewed 2026 research in an automotive spare parts case study reiterates that Croston-style methods remain effective and compares common variants (Scientific Reports, 2026).

Which method should we use for slow-moving spare parts (Croston vs SBA vs TSB vs ML)?

• Croston: Separates demand size from inter-arrival time and forecasts them independently. Good baseline for intermittent items, but can be biased.
• SBA (Syntetos–Boylan Approximation): A bias-corrected Croston variant. Often a strong default when you want “Croston, but safer.”
• TSB (Teunter–Syntetos–Babai): Explicitly models the probability of demand occurrence and adapts better when parts become obsolete or demand “dies off.” Useful for supersessions and end-of-life behaviors.
• ML for intermittent: Can work, but only when you have the right features (usage drivers, install base, machine hours, fleet size, service contracts) and clean item relationships. Recent research also warns that for intermittent demand, feature engineering can matter more than model complexity (Scientific Reports, 2026).

Practical recommendation for SMEs: Start with SBA/TSB for the long tail, reserve ML for a targeted subset where you can reliably feed driver signals. Your ERP should support multiple methods concurrently (segmentation is non-negotiable).

ERP capabilities to insist on

• Intermittent-demand methods out of the box: Croston/SBA/TSB (or equivalent) should be selectable per item-location.
• Master data for supersessions: Part-to-part substitution chains, phase-in/phase-out dates, and service BOM visibility so forecasts don’t “reset” when engineering changes part numbers.
• Inventory policy integration: Forecasts should flow into reorder points/safety stock logic; intermittent items usually need service-level-based policies, not just averages.

Method #3: Hierarchical forecasting and reconciliation (so totals match)

Manufacturers rarely plan at one level. You forecast by SKU and location, but finance wants family/region totals, and production wants plant-level load. Without reconciliation, you get mismatches: SKU forecasts don’t sum to category totals, and planners stop trusting the system. 2025–2026 research increasingly focuses on hierarchical, interpretable methods that keep forecasts coherent while explaining uncertainty (research on hierarchical methods, 2025–2026).

How do we reconcile forecasts across SKU, family, plant, and region?

• Top-down: Forecast at aggregate and allocate down using proportions. Simple, but can miss SKU-level dynamics.
• Bottom-up: Forecast at SKU-location and sum up. Often noisy for long tails.
• Reconciled (recommended): Forecast at multiple levels, then reconcile so constraints hold (totals match) while preserving accuracy.

ERP capabilities to insist on

• True hierarchy support: Item, product family, brand, plant, region, channel—maintained as master data, not a spreadsheet workaround.
• Reconciliation engine: Ability to run bottom-up plus constrained reconciliation (at minimum) with audit trails.
• Workflow + collaboration: Forecast review by level (category manager vs site planner) with role-based permissions.

Method #4: Probabilistic forecasting (planning with uncertainty, not one number)

Single-number forecasts hide risk. Probabilistic forecasting provides prediction intervals (for example P50/P90) so you can size safety stock and capacity using uncertainty. This matters when lead times swing, demand is lumpy, or supplier reliability varies. Probabilistic approaches are increasingly common in industrial forecasting research, often combined with hierarchy so you can quantify uncertainty consistently across levels (hierarchical probabilistic research, 2025–2026).

ERP capabilities to insist on

• Quantiles/intervals stored and usable: The system should store more than “forecast qty”—it should support intervals feeding safety stock and service levels.
• Scenario management: What-if comparisons (price change, promo, lost customer, supplier delay) with traceability.
• Explainability & logging: Why did uncertainty widen? What variables changed?

Accuracy metrics that won’t mislead you (MAPE vs WMAPE vs bias)

Forecast KPIs are where many teams accidentally optimize the wrong thing. A 2026 SAP guidance document recommends relative error measures such as MAPE/WMAPE for cross-material comparisons and cautions that MAPE behaves poorly with sporadic demand (SAP KPI guidance (2026)).

What to track in manufacturing

• WMAPE (Weighted MAPE): More stable across SKUs because high-volume items get appropriate weight. A good default “executive metric.”
• Bias (signed error): Systematic over-forecasting inflates inventory; under-forecasting drives stockouts and expedite costs.
• Service impact metrics: Fill rate, stockout rate, and expedite frequency—because accuracy alone doesn’t equal outcomes.

Which metrics break on intermittent demand?

• MAPE: Can explode or become undefined when actual demand is zero (common in spare parts). Use WMAPE, MAE, or intermittent-friendly metrics, and always segment reporting for intermittent SKUs.

Overrides, governance, and Forecast Value Added (FVA): stop guessing

Human overrides are sometimes essential (customer intel, known one-offs, contract wins/losses). But they should be audited. A 2025 study on Forecast Value Added supports measuring when judgmental adjustments help or hurt accuracy and bias, rather than assuming consensus always improves results (Forecast Value Added research (University of Bath, 2025)).

When do overrides improve the forecast?

• When they add information not in the data (new contract, competitor outage, plant shutdown)
• When they correct known data issues (mis-booked orders, one-time project demand)

ERP capabilities to insist on

• Governed overrides: Reason codes, approvals, and expiry dates (so a “temporary” override doesn’t become permanent).
• FVA reporting: Track baseline vs overridden forecast accuracy and bias by planner/team/category.

New products and engineered-to-order (ETO): forecasting without history

For new SKUs, statistical history is thin or nonexistent. For ETO, demand may be project-based rather than repeatable. In both cases, your ERP forecasting process should treat “forecast” as a structured estimate tied to commercial signals.

What works in practice

• Analog models: Map new items to similar products (same family, price band, channel) and borrow their curves.
• Order pipeline + stages: Convert quotes/opportunities into probabilistic demand with win rates and expected dates.
• Rules-based planning: Min/max or service-level targets until enough history accumulates, then transition to time-series.

External signals, promotions, lead times: how to incorporate them without a data science team

You don’t need a PhD team to use causal signals, but you do need an ERP that can manage the plumbing: clean event data, consistent calendars, and repeatable model runs. The reality is that many vendors now bundle “AI experiences” into planning—Microsoft’s 2025 wave 2 plan describes generative insights for demand planning (planned March 2026) to help users analyze history and forecasts across dimensions (Microsoft release plan). Treat this as augmentation, not replacement.

Practical role of generative AI vs statistical/ML models

• Generative AI: Explanations, anomaly investigation, narrative summaries, “what changed?” across SKU/location/customer, and faster root-cause exploration.
• Statistical/ML models: The actual forecast numbers, uncertainty bands, and repeatable backtests.

Retraining, drift detection, and the unglamorous prerequisites that drive ROI

Model refresh cadence depends on volatility: many SMEs do well with monthly re-fitting plus weekly exception monitoring. More important than cadence is detecting regime changes (new customer, lost distributor, substitution, policy changes). Build alerts around demand shifts, forecast error spikes, and bias trends.

Master-data prerequisites before migrating from legacy ERP forecasting

• Item/location master: Consistent units of measure, pack sizes, lead times, MOQ/lot sizes
• Hierarchies: Product family/category, site/region/channel structures used for reconciliation
• Calendars: Working days, holidays, fiscal periods, promo/event tagging
• Supersessions/substitutions: Especially for spare parts and engineering changes
• History hygiene: Returns, one-offs, and stockouts flagged so models don’t learn the wrong lesson

Can SMEs do this inside ERP without buying an expensive planning suite?

Yes—if your ERP supports segmented methods, orchestration, governance, and metrics. The direction of “ERP-linked demand planning” is clearly toward orchestrating multiple models and allowing extensibility (including custom ML) rather than betting on one algorithm (see the Dynamics 365 approach). But the biggest gains still come from fixing data and process first, then layering automation. Industry summaries continue to cite meaningful upside when done correctly—Oracle (Dec 2025) references McKinsey findings that AI-powered forecasting can reduce errors by 20%–50% and reduce product unavailability by up to 65% (Oracle’s summary).

What to look for when evaluating ERP forecasting (2026 checklist)

• Methods coverage: Time-series + intermittent methods + hierarchy + (ideally) probabilistic outputs
• Model orchestration: Segmentation, backtesting, auto-selection, scheduled runs
• ML extensibility: Ability to plug in custom models later without re-platforming
• Governance: Override workflow, reason codes, approvals, and FVA auditing
• KPI safety: WMAPE, bias, and intermittent-friendly reporting (not MAPE-only dashboards)
• Integration to execution: Forecasts feed MRP, purchasing, safety stock, and capacity plans

Next steps: modern forecasting without locking yourself into expensive legacy stacks

If you’re modernizing from a legacy ERP, start by documenting your SKU segments (steady, seasonal, intermittent, new/ETO), your hierarchy needs, and the KPIs you’ll standardize on (WMAPE + bias as a minimum). Then choose software that can orchestrate multiple forecasting methods with strong governance—because AI features are improving fast, but adoption is still immature and foundations determine ROI.

LikwidERP is a free, open-source, self-hostable ERP that brings inventory, sales, manufacturing, and finance into one system—so forecasting improvements can actually flow into execution. Explore LikwidERP features, read our buyer’s guide to choosing a manufacturing ERP, or contact us to discuss replacing your legacy ERP with a modern stack built for practical planning.