Intermittent Demand

The defining feature is a time series with a high frequency of zero-demand periods. Unlike slow-moving items that sell consistently at low volumes, intermittent items have no demand at all during many observation intervals. This zero-inflation violates the assumptions of standard Gaussian models, requiring specialized discrete probability distributions like the Poisson or Negative Binomial to accurately model the demand generation process.

When demand does occur, the transaction size is often erratic and non-normal. A single order might be for 1 unit or 100 units. This high coefficient of variation makes point forecasts unreliable. Effective modeling requires separately forecasting the demand interval (time between occurrences) and the demand magnitude (size of the order), a principle central to methods like Croston's Method.

A specific subset of intermittent demand where both the demand interval and the demand magnitude exhibit high variability. Lumpy demand is characterized by infrequent transactions with a large variance in order size. This pattern is common in capital goods and high-value spare parts. Forecasting lumpy demand is exceptionally difficult, as historical averages are skewed by rare, large transactions.

Classical forecasting models rely on identifying autocorrelation in a time series. However, the long stretches of zeros in intermittent demand data destroy the serial correlation structure. The demand signal is often too weak to identify meaningful patterns like seasonality or trend using standard Autocorrelation Function (ACF) plots, forcing reliance on purely statistical smoothing methods rather than pattern-recognition algorithms.

Standard Economic Order Quantity (EOQ) models fail under intermittent demand because they assume continuous, deterministic depletion. Instead, inventory control relies on base-stock policies or (s, Q) systems. The primary goal shifts from optimizing order quantity to setting the correct reorder point (s) to achieve a target service level, often using a compound Poisson process to model the demand distribution.

Standard accuracy metrics like Mean Absolute Percentage Error (MAPE) are undefined or infinite when actual demand is zero. This necessitates alternative evaluation approaches. Periods in Stock (PIS) or Mean Absolute Scaled Error (MASE) are preferred. Ultimately, the model's value is judged by its inventory efficiency—achieving the target service level with the minimum possible safety stock investment.

The foundational statistical model built specifically for intermittent demand. Instead of a single forecast, it decomposes the problem into two separate exponential smoothing estimates: one for the demand interval (time between non-zero periods) and one for the demand size (magnitude when it occurs). This prevents the forecast from being biased downward by long sequences of zero values, a critical flaw of standard methods like simple exponential smoothing when applied to spare parts.

Not all intermittent demand is equal. The Syntetos-Boylan classification scheme segments items by the squared coefficient of variation of demand size and the average inter-demand interval. This creates four quadrants:

• Erratic: High variability in size, frequent demand.
• Lumpy: High variability in size, infrequent demand.
• Smooth: Low variability, frequent demand.
• Intermittent: Low variability, infrequent demand. Selecting the right forecasting model depends entirely on this categorization.

For intermittent items, traditional safety stock formulas based on a normal distribution of demand fail catastrophically. Instead, inventory planners must use compound distributions (e.g., Poisson distribution for demand arrivals combined with a Gamma distribution for demand size) or bootstrapping methods on historical demand intervals. The goal is to calculate a reorder point that achieves a target cycle service level without holding excessive capital in slow-moving stock.

A point forecast is nearly useless for a sporadic SKU. Probabilistic forecasting outputs a full predictive distribution, allowing a supply chain director to make a risk-aware decision. For example, the model might predict a 60% chance of zero demand next month, a 30% chance of 1-5 units, and a 10% chance of a 15-unit spike. This distribution is evaluated using proper scoring rules like the Continuous Ranked Probability Score (CRPS) rather than traditional metrics like MAPE.

A modern deep learning approach that excels at intermittent demand by modeling it as a negative binomial likelihood within an autoregressive recurrent neural network. DeepAR learns a global model across thousands of related time series, allowing it to borrow statistical strength from similar items. This is particularly effective for cold-start items where a new spare part has no history, but the model can infer its pattern from items with similar attributes like category, price, or supplier lead time.

Intermittent demand at the retail level is a primary trigger for the bullwhip effect. A single, lumpy order from an end customer can be misinterpreted upstream as a permanent shift in demand. Without recognizing the intermittent nature of the signal, each tier of the supply chain (retailer, wholesaler, manufacturer) amplifies the order variability, leading to excess inventory, costly production ramps, and subsequent obsolescence. Proper classification dampens this distortion.