Quantile Forecasting

Unlike traditional methods that output a single number, quantile forecasting estimates the entire conditional distribution of future demand. It answers "What is the maximum demand we expect with 95% confidence?" rather than just "What is the average demand?" This is achieved by minimizing the pinball loss function during model training, which asymmetrically penalizes over-prediction and under-prediction depending on the target quantile.

The primary operational advantage is the direct mathematical link to business policy. A service level target (e.g., 95% cycle service level) maps exactly to a specific quantile forecast (the 0.95 quantile). This eliminates the need for arbitrary safety multipliers or assumptions of normality. The forecast itself becomes the reorder point:

• 50th quantile: Median demand forecast
• 95th quantile: Demand level sufficient to cover all demand in 95% of replenishment cycles
• 99th quantile: Near-stockout-proof buffer level

Quantile models are trained using the pinball loss (or quantile loss), which applies different penalties for over-forecasting versus under-forecasting. For a 95% quantile, underestimating demand (leading to a stockout) is penalized 19 times more heavily than overestimating (leading to excess inventory). This asymmetry forces the model to learn the upper boundary of the demand distribution rather than the central tendency, making it inherently conservative for buffer-sizing applications.

Quantile forecasting makes no assumptions about the underlying demand distribution. It accurately models:

• Intermittent demand with frequent zero periods
• Multi-modal distributions caused by promotions or seasonality
• Heavy-tailed distributions where extreme events are more common than a normal distribution predicts This is critical for spare parts, luxury goods, and erratic SKUs where traditional parametric methods like Gaussian safety stock calculations fail.

Advanced implementations enforce quantile monotonicity—the constraint that a higher quantile forecast must always be greater than or equal to a lower quantile forecast at every time step. Without this constraint, a model might predict a 90th quantile value lower than its 50th quantile prediction, creating logical inconsistencies. Techniques like the multi-quantile recurrent forecaster architecture output all required quantiles simultaneously with built-in crossing penalties.

By forecasting multiple quantiles simultaneously, the system constructs prediction intervals with precise coverage guarantees. A 90% prediction interval is formed by the 5th and 95th quantile forecasts, meaning the actual demand will fall within this range 90% of the time. This provides planners with a complete risk spectrum:

• Narrow intervals: High confidence, stable demand patterns
• Wide intervals: High volatility, signaling the need for larger buffers or alternative sourcing strategies

A buffer calculation method that models demand and lead time as probability distributions rather than fixed values. Unlike deterministic approaches that use single-point averages, stochastic safety stock captures the full range of variability to achieve a precise target service level. The quantile forecast directly feeds this calculation by providing the demand percentile corresponding to the desired protection level.

The statistical characterization of historical prediction deviations used to calibrate safety stock. Rather than assuming errors follow a normal distribution, modern approaches fit empirical error patterns to identify skewness and heavy tails. Key components include:

• Mean Absolute Percentage Error (MAPE): Measures average deviation magnitude
• Root Mean Squared Error (RMSE): Penalizes large errors more heavily
• Pinball Loss: The scoring rule specifically designed to evaluate quantile forecast accuracy

The desired probability of not stocking out during a replenishment cycle, expressed as a percentage. A 95% service level means stockouts are tolerated in only 5% of cycles. Quantile forecasting directly maps to this target: to achieve a 95% service level, you set safety stock at the 95th percentile of the demand forecast distribution. Higher service levels require disproportionately more inventory due to the shape of the demand distribution's tail.

A computational technique that runs thousands of randomized demand-supply scenarios to empirically determine required safety stock. Each simulation draws from the full probability distributions of demand and lead time, generating a distribution of possible inventory outcomes. The quantile forecast provides the demand distribution input, while the simulation aggregates it with supply uncertainty to compute the empirical buffer needed for any target percentile.

A phenomenon where large demand fluctuations tend to be followed by more large fluctuations, violating the assumption of independent, identically distributed observations. This clustering means that during turbulent periods, the conditional variance increases, requiring adaptive safety stock. Quantile forecasting models must account for this heteroskedasticity by dynamically widening prediction intervals when volatility regimes shift.

The degradation of a safety stock model's accuracy over time as the underlying statistical properties of demand or supply change. When concept drift occurs, the quantile forecast's calibrated percentiles no longer match observed outcomes, leading to systematic overstocking or understocking. Detection requires continuous monitoring of prediction interval coverage—the percentage of actual observations falling within forecasted quantile ranges.