Probabilistic Forecasting for Intermittent Demand Items: Crostons Method and Beyond in Micro-Retail PoS Data
Address the special forecasting challenge of infrequently sold items where standard methods fail, proposing robust intermittent-demand models.
Key Takeaways
- Intermittent demand items — those with many zero-demand periods interspersed with sporadic nonzero demand — violate the continuous-demand assumptions of standard forecasting methods, requiring specialized approaches.
- Croston method and its bias-corrected variant SBA decompose intermittent demand into separate models for demand size and inter-demand interval, producing forecasts better suited to sparse data patterns.
- Probabilistic forecasts that output full demand distributions rather than point predictions are essential for intermittent items because the high uncertainty makes point forecasts misleading for inventory decisions.
The Intermittent Demand Challenge
Micro-retail environments commonly stock items that sell infrequently: specialty products, seasonal items, slow-moving replacement parts, or niche goods that serve a small but loyal customer segment. These intermittent demand items are characterized by a preponderance of zero-demand periods punctuated by sporadic purchases, often of variable quantity. The demand time series for such items is dominated by zeros, with occasional nonzero observations that may be separated by days, weeks, or even months. Standard time-series forecasting methods such as ARIMA, exponential smoothing, and moving averages perform poorly on intermittent series because they attempt to model a continuous process when the underlying demand is fundamentally discontinuous. Applying exponential smoothing to a series of mostly zeros with occasional nonzero values produces forecasts that are perpetually close to zero — too low during demand periods and too high (relative to actual zero demand) during non-demand periods. The resulting forecasts are unreliable for inventory management, which requires knowing both the expected demand level when demand occurs and the probability that demand will occur in a given period. Despite their slow movement, intermittent items often represent a significant fraction of the SKU count in micro-retail assortments and cannot simply be ignored. askbiz.co automatically identifies intermittent demand items based on their demand occurrence rate and applies specialized forecasting methods that account for the zero-inflated nature of their demand patterns.
Croston Method and SBA Correction
Croston (1972) proposed a decomposition approach specifically designed for intermittent demand forecasting. Rather than modeling the raw demand series directly, the method separately models two components: the demand size (the quantity demanded in periods when demand occurs) and the inter-demand interval (the number of periods between consecutive demand occurrences). Both components are updated using simple exponential smoothing, applied only when a nonzero demand observation occurs. The demand rate forecast is the ratio of the smoothed demand size to the smoothed inter-demand interval, providing an estimate of average demand per period that is not biased downward by the many zero observations. However, Syntetos and Boylan (2001) demonstrated that Croston original method is positively biased — it systematically overestimates demand — and proposed the Syntetos-Boylan Approximation (SBA), which applies a bias correction factor to the Croston forecast. The SBA has been shown to outperform both Croston original method and simple exponential smoothing across a wide range of intermittent demand patterns in empirical studies. The smoothing parameter alpha, common to both components in the simplest implementation, controls the responsiveness of the forecast to new observations: higher alpha values adapt quickly to changes but are more susceptible to noise, while lower values provide smoother, more stable forecasts. askbiz.co implements both Croston and SBA methods with automated smoothing parameter optimization, selecting the approach that minimizes scaled forecast error on a rolling validation window for each intermittent demand SKU.
Beyond Croston: Modern Intermittent Demand Models
While Croston and SBA remain widely used baselines, several modern approaches extend the framework or take fundamentally different modeling perspectives. The Teunter-Syntetos-Babai (TSB) method modifies the Croston framework by modeling the demand probability (the likelihood of nonzero demand in each period) rather than the inter-demand interval, which can adapt more quickly to changes in demand frequency. Temporal aggregation — forecasting at a coarser time granularity where demand is less intermittent (weekly rather than daily, monthly rather than weekly) and then disaggregating to the required operational granularity — can transform a highly intermittent series into a manageable one, though the disaggregation step introduces additional uncertainty. Bootstrapping methods, particularly the approach of Willemain, Smart, and Schwarz (2004), generate demand distributions by resampling from the empirical demand history, preserving the zero-inflated nature of the distribution without parametric assumptions. Zero-inflated regression models (zero-inflated Poisson or zero-inflated negative binomial) explicitly model the probability of zero demand and the conditional demand distribution when demand occurs, providing a parametric framework that accommodates excess zeros. Machine learning approaches including gradient-boosted trees trained on temporal features with a zero-inflated loss function can capture complex patterns in the demand occurrence process. askbiz.co evaluates multiple intermittent demand methods for each slow-moving SKU and selects the approach that provides the best calibrated probabilistic forecasts on out-of-sample validation data.
Probabilistic Forecasts for Inventory Decisions
For intermittent demand items, point forecasts — single-valued predictions of future demand — are particularly inadequate for inventory management because the high variance and zero-inflation of intermittent demand means that the most likely outcome (zero demand) and the expected value (average demand) are both poor guides for ordering decisions. Probabilistic forecasts that output the full distribution of possible demand outcomes over the planning horizon enable more nuanced inventory decisions. The critical quantile for inventory management is the service-level quantile: if the target service level is 95 percent, the reorder point should be set at the 95th percentile of the demand-during-lead-time distribution, ensuring that demand is covered 95 percent of the time. For intermittent items, this quantile may be zero (if the probability of any demand during the lead time is less than 5 percent) or may jump discontinuously from zero to a positive value, reflecting the discrete, zero-inflated nature of the demand distribution. Computing these quantiles analytically requires specifying the full demand distribution, which bootstrapping and zero-inflated models provide. Prediction intervals — the range within which future demand will fall with stated probability — communicate the forecast uncertainty to decision-makers, helping them understand that an intermittent item forecast of "0.3 units per day" does not mean demand of 0.3 units will occur each day but rather that demand of one or more units will occur roughly once every three to four days. askbiz.co presents intermittent demand forecasts as probability distributions with clearly marked service-level quantiles, enabling automated reorder point calculations that properly account for the discrete, uncertain nature of slow-moving item demand.
Practical Considerations for Micro-Retail
Managing intermittent demand items in micro-retail involves practical considerations beyond forecasting accuracy. The decision of whether to stock an intermittent item at all depends on the balance between the gross margin earned per unit sold and the holding cost of maintaining inventory between infrequent sales events. A simple profitability threshold — will the expected annual margin from this item exceed the annual holding cost of its minimum stock level — provides a rational basis for SKU rationalization decisions. Minimum order quantities imposed by suppliers often exceed the slow-selling rate of intermittent items, forcing retailers to choose between carrying excess inventory or abandoning the product. Joint ordering with faster-moving items from the same supplier can mitigate this by spreading fixed ordering costs across a broader order. Lead time management is particularly critical for intermittent items because a stockout on an item with low demand frequency may go unnoticed for an extended period, and reorder lead time may exceed the typical inter-demand interval, creating prolonged unavailability. Safety stock for intermittent items should be calibrated not just to cover demand during lead time but to account for the delayed detection of stockouts in a low-visibility category. askbiz.co monitors inventory levels of intermittent demand items relative to forecasted demand distributions, triggering reorder alerts when stock falls below the service-level-calibrated safety threshold and flagging items whose holding costs exceed their expected margin contribution.