Croston's Method

Croston's method separates the forecasting problem into two distinct components, updated only when demand occurs:

• Demand Size (z_t): The non-zero quantity demanded in a period, smoothed using simple exponential smoothing
• Inter-Arrival Interval (p_t): The number of periods since the last demand occurrence, also exponentially smoothed

The final forecast is the ratio z_t / p_t, representing demand per period. This decomposition prevents zero-demand periods from artificially deflating the forecast, which is the critical flaw of standard exponential smoothing on intermittent series.

Classical forecasting methods like Simple Exponential Smoothing (SES) produce systematically biased low forecasts when applied to intermittent demand because they update the level estimate in every period—including zeros—causing the forecast to drift downward during demand droughts.

Croston's method eliminates this bias by:

• Updating only on demand-positive periods, preserving the true demand rate
• Maintaining separate smoothing parameters for size and interval, allowing each component to adapt at its own rate
• Producing an unbiased estimate of mean demand per period under the assumption of independent demand sizes and intervals

Croston's method typically uses a single smoothing constant α (between 0 and 1) for both the demand size and interval estimates, though variants allow separate parameters:

• Low α (e.g., 0.05–0.1): Heavy smoothing, appropriate for very slow-moving items with stable demand patterns
• High α (e.g., 0.3–0.5): More responsive to recent changes, suitable when demand characteristics shift
• α = 0: The forecast never updates; it remains the initial estimate
• α = 1: The forecast equals the most recent observation, ignoring all history

Parameter selection is typically done via grid search minimizing MSE on a holdout set, though care must be taken because intermittent series have sparse error observations.

Research by Syntetos and Boylan (2001) demonstrated that Croston's original formulation is not fully unbiased—it exhibits a slight positive bias because the expected value of a ratio does not equal the ratio of expected values.

The Syntetos-Boylan Approximation (SBA) corrects this by applying a deflating factor:

Forecast = (1 - α/2) × (z_t / p_t)

This adjustment is now considered best practice for intermittent demand forecasting and is implemented in most production forecasting libraries. The bias is most pronounced when α is large and demand is extremely sporadic.

Croston's method operates under specific assumptions that define its domain of applicability:

• Independence: Demand sizes and inter-arrival intervals are assumed to be independent of each other and identically distributed over time
• No trend or seasonality: The method assumes a stationary demand process; it cannot capture growth trends or seasonal patterns
• Non-negative demand: All demand occurrences must be positive; the method does not handle returns or negative adjustments
• Point forecast only: Standard Croston provides a mean forecast but no prediction intervals; extensions like the Croston with bootstrapping variant are required for uncertainty quantification

For items exhibiting trend or seasonality, extensions such as TSB (Teunter-Syntetos-Babai) or iETS (intermittent ETS) models should be considered.

Croston's method is rarely used in isolation—it feeds directly into inventory control policies for spare parts and slow-moving SKUs:

• Reorder Point Calculation: The demand per period forecast (z_t / p_t) is multiplied by lead time to estimate lead time demand, which determines safety stock levels
• Periodic Review Systems: In (R, S) policies, Croston forecasts drive the order-up-to level S for each review interval
• Service Level Targeting: The forecast mean is combined with an assumed demand distribution (often negative binomial or compound Poisson) to set stock levels that achieve target fill rates

Proper integration requires understanding that Croston's output is a rate, not a period-by-period prediction, and inventory models must account for the lumpy nature of demand arrivals.

The core problem Croston's Method solves. A demand pattern characterized by sporadic occurrences interspersed with many periods of zero demand, common in spare parts, capital goods, and long-tail retail. Standard models like exponential smoothing fail here because they are biased by the high frequency of zeros, leading to forecasts that drift toward zero. Key characteristics include infrequent transactions and variable demand sizes when a transaction does occur.

The systematic tendency of a model to consistently over-predict or under-predict actual demand. Croston's Method specifically addresses the downward bias introduced by traditional methods on intermittent series. Bias is measured as the Mean Forecast Error (MFE). A positive MFE indicates persistent under-forecasting, while a negative MFE signals over-forecasting. Eliminating bias is critical for setting accurate safety stock levels.

An additional quantity of inventory held in reserve to mitigate the risk of stockouts caused by variability in both supply and demand during lead time. For intermittent demand items, safety stock calculations rely heavily on the forecast error distribution from Croston's Method. The goal is to achieve a target service level—the probability of not stocking out during a replenishment cycle—without tying up excessive working capital in slow-moving inventory.

A direct refinement of Croston's original method that corrects a subtle but significant positive bias in the original estimator. Croston's Method was found to over-forecast slightly because the expected value of a ratio is not the ratio of expected values. SBA applies a simple deflating factor of (1 - α/2) to the demand size estimate, making it nearly unbiased. It is the de facto standard for intermittent demand forecasting in practice.

A range of values within which a future observation is expected to fall with a specified probability, quantifying forecast uncertainty. For Croston's Method, these intervals are not symmetric and must account for the compound distribution of demand occurrence and demand size. Techniques like bootstrapping from the empirical distribution of non-zero demands and inter-demand intervals are used to construct realistic intervals for inventory optimization.

The only robust way to evaluate a Croston's model. This technique simulates production by sequentially retraining the model on an expanding window of historical data and testing on the subsequent held-out period. Standard k-fold cross-validation is invalid for time series because it shuffles data and breaks temporal dependencies. Walk-forward validation preserves the order of observations and provides a realistic measure of out-of-sample accuracy.