Forecasting Lumpy and Intermittent SKU Demand: Croston's Method, TSB and MinT Reconciliation
Intermittent demand forecasting addresses products with sporadic, zero-heavy sales histories—such as spare parts, B2B industrial components, specialty SKUs and long-tail retail variants. Standard time-series and black-box machine-learning models fail on intermittent demand because they smooth across zero-sale periods or overfit historical noise, producing biased inventory buffers and forecasts that do not add up across SKU, category and national totals.
Ordeen Forecasting (£2,000/month with ERP and SAP Business One integration) automatically detects intermittent demand patterns to switch between Croston's method, Syntetos-Boylan Approximation (SBA) and Teunter-Syntetos-Babai (TSB), applies MinT (Minimum Trace) hierarchical reconciliation, and gates every AI-discovered external driver through strict out-of-sample econometric validation.
Why Standard Time-Series and Pure ML Fail on Real Supply Chains
The Zero-Demand Bias on Lumpy SKUs. If a SKU sells 50 units in Week 1, 0 units in Weeks 2–4, and 40 units in Week 5, exponential smoothing or naive ARIMA treats the zeros as a downward trend right before the next order arrives—then over-forecasts right after a spike. Croston's method decouples the problem into two separate estimates: (a) the inter-arrival interval between non-zero demand occurrences, and (b) the demand size when an order occurs. SBA corrects Croston's positive mathematical bias, and TSB updates the probability of demand occurring even during long strings of zeros (essential for obsolescence and end-of-life SKUs).
Incoherent Forecasts Across the Hierarchy. When SKU-level models are run independently from regional or product-group models, the sum of the SKU forecasts rarely equals the category or CFO revenue forecast. Ordeen applies MinT (Minimum Trace) optimal hierarchical reconciliation (Wickramasuriya, Athanasopoulos & Hyndman), using the covariance structure of the forecast errors across all levels so SKU, category, warehouse and total company forecasts sum up identically while improving accuracy at every level.
Spurious Correlations in Black-Box AI. Feeding 50 external macro or search-trend variables into a neural network guarantees overfitting on historical data. In Ordeen, AI proposes; statistics disposes: AI discovers candidate external signals (such as search trends or sector indicators), and econometric tests verify out-of-sample predictive lift, lead-lag causality, and stability after controlling for empirically tested seasonality before any variable enters the model.
Selecting the Right Estimator by Demand Pattern in Ordeen
SKU Demand Profile | Failure of Standard ML / Smoothing | Ordeen Forecasting Treatment |
|---|---|---|
Smooth / High-Velocity SKUs | Overfits spurious macro correlations; assumes fixed Q4 seasonality | Econometric time-series + statistically tested seasonality + out-of-sample validated external drivers with confidence intervals |
Intermittent / Lumpy SKUs (spare parts, B2B bulk orders) | Biased after demand spikes; under-stocks or over-stocks safety inventory | Croston's Method & Syntetos-Boylan Approximation (SBA) separating occurrence interval from order size |
Dying / Obsolescence-Risk SKUs (long runs of zeros) | Croston freezes its estimate when no demand occurs for months | Teunter-Syntetos-Babai (TSB) updating demand probability every period including zero-demand weeks |
Multi-Level Product / Region Trees | Bottom-up sums disagree with top-down finance targets | MinT (Minimum Trace) Hierarchical Reconciliation enforcing exact additivity across SKU → Group → Total |
Frequently Asked Questions
What is the difference between Croston's method, SBA and TSB in intermittent demand forecasting? Croston's method separately forecasts the time between non-zero orders and the average order size, but has a small positive mathematical bias. The Syntetos-Boylan Approximation (SBA) applies a bias-correction factor to Croston's formula. The Teunter-Syntetos-Babai (TSB) method models the probability of demand occurring in each period directly, allowing the forecast to decay toward zero during long stretches without sales (making TSB ideal for slow-moving or obsolete inventory).
What is MinT (Minimum Trace) hierarchical reconciliation? MinT reconciliation adjusts forecasts across a hierarchy (such as SKU → Product Category → Country → Total Company) by minimising the trace of the covariance matrix of the reconciled forecast errors. This guarantees that all levels add up exactly to the same total while weighting more reliable levels higher than noisy ones.
How does Ordeen prevent AI from overfitting demand forecasts? Ordeen follows "verified data first, AI last". AI is used to propose candidate external signals, but every variable must pass four statistical gates—temporal stability, lead-lag direction, significance after seasonality controls, and genuine out-of-sample cross-validation improvement—before it is included in your forecast.
Want demand forecasts you can explain to your Finance Director? Book a 20-minute demo or email im@ordeen.com.
