Quantile Regression

Quantile regression is trained by minimizing the pinball loss (also called quantile loss), an asymmetric function that penalizes errors differently depending on whether they fall above or below the predicted quantile.

• For the 90th percentile, over-prediction is penalized 9x more than under-prediction
• For the 10th percentile, under-prediction is penalized 9x more than over-prediction
• The asymmetry parameter τ directly corresponds to the target quantile
• This loss function is non-differentiable at zero, requiring linear programming or subgradient methods for optimization

By fitting models at multiple quantiles simultaneously, practitioners can construct asymmetric prediction intervals that capture heteroscedastic uncertainty without assuming a parametric error distribution.

• A 90% prediction interval is formed by the 5th and 95th percentile forecasts
• Intervals naturally widen in regions of high variance and narrow where data is dense
• Unlike Gaussian-based intervals, these can be asymmetric around the median
• This approach is critical for inventory optimization, where the cost of understocking far exceeds overstocking

Quantile regression, particularly at the median (τ=0.5), provides robust estimates that are resistant to extreme values in the response variable, unlike mean regression which can be arbitrarily distorted by a single outlier.

• The median regression minimizes absolute errors rather than squared errors
• Outliers in the target variable have bounded influence on coefficient estimates
• This property is especially valuable in demand forecasting where promotional spikes or stockouts create extreme observations
• No need for manual outlier removal or winsorization preprocessing steps

Quantile regression naturally captures heteroscedasticity—the phenomenon where the variance of the error term changes across the predictor space—without requiring explicit variance modeling.

• Different quantile slopes reveal how predictor effects change across the distribution
• For example, a price coefficient may be more negative at lower demand quantiles than upper ones
• This reveals distributional effects that mean regression completely obscures
• In retail, this shows whether promotions compress or expand demand variability

Unlike maximum likelihood methods that assume a specific error distribution (e.g., Gaussian, Poisson), quantile regression is distribution-free, making valid inferences without parametric assumptions about the data-generating process.

• No requirement for normally distributed residuals
• Handles skewed, multimodal, or heavy-tailed conditional distributions
• Valid inference relies on asymptotic theory rather than distributional assumptions
• Particularly suited for intermittent demand patterns where zero-inflated distributions violate standard assumptions

A practical challenge in quantile regression is quantile crossing, where a lower quantile prediction exceeds a higher one (e.g., the 10th percentile forecast is greater than the 90th). Modern implementations use constrained optimization to prevent this.

• Monotonicity constraints ensure τ₁ < τ₂ implies Q(τ₁) ≤ Q(τ₂)
• Non-crossing quantile regression uses simultaneous estimation with inequality constraints
• Rearrangement post-processing can also correct crossing in fitted values
• Essential for generating coherent prediction intervals in production forecasting systems

The asymmetric objective function that makes quantile regression possible. Unlike squared error loss which targets the conditional mean, pinball loss applies a tilted absolute penalty:

• For a target quantile τ, the loss is τ × |error| for over-predictions
• For under-predictions, the loss is (1-τ) × |error|
• At τ=0.5, pinball loss reduces to median regression (symmetric absolute loss)
• The tilt ratio determines whether the model errs toward over-forecasting or under-forecasting

This asymmetry is what forces the model to learn specific conditional quantiles rather than the conditional mean, enabling the construction of prediction intervals with calibrated coverage probabilities.

A range constructed from two quantile forecasts that captures future demand with a specified probability. In retail forecasting:

• A 90% prediction interval uses the 5th and 95th percentiles as lower and upper bounds
• The interval width reflects forecast uncertainty — wider intervals indicate higher volatility
• Unlike confidence intervals (which estimate parameter uncertainty), prediction intervals account for inherent observation noise
• Quantile regression natively produces asymmetric intervals, which is critical for demand data that exhibits right-skewed distributions

Supply chain planners use these intervals to set safety stock levels: the upper quantile directly informs the inventory buffer needed to achieve a target service level.

The paradigm shift from point forecasts to full distributional outputs. Quantile regression is a foundational technique within this framework:

• A point forecast (e.g., 500 units) provides no risk information
• A probabilistic forecast outputs the entire conditional distribution P(Y|X)
• By estimating multiple quantiles (10th, 25th, 50th, 75th, 90th), quantile regression reconstructs this distribution non-parametrically
• No assumption of normality is required — the method captures arbitrary distribution shapes

This is essential for inventory optimization where the cost of understocking (lost sales) differs from overstocking (holding costs), requiring asymmetric risk assessment.

The CRPS is the gold-standard metric for evaluating probabilistic forecasts, directly measuring how well quantile regression outputs match observed outcomes:

• Compares the entire predicted cumulative distribution function to the actual observation
• A CRPS of zero indicates a perfect probabilistic forecast
• Penalizes both miscalibration (wrong spread) and sharpness (overconfident narrow intervals)
• More informative than point metrics like MAE because it rewards well-calibrated uncertainty quantification

When training quantile regression models, CRPS serves as a proper scoring rule that encourages the model to produce honest prediction intervals rather than artificially narrow ones.

Demand sensing uses real-time signals (POS data, weather, social trends) to refine short-term forecasts. Quantile regression enhances this approach:

• Traditional demand sensing produces point estimates vulnerable to sudden demand spikes
• Quantile regression on streaming data provides real-time prediction intervals that adapt to signal volatility
• During promotions or weather events, the upper quantile widens automatically, signaling increased uncertainty
• This enables dynamic safety stock adjustments rather than static buffer calculations

The combination allows supply chains to be both responsive (via demand sensing) and risk-aware (via quantile regression), reducing both stockouts and excess inventory simultaneously.

When forecasting at multiple aggregation levels (SKU → category → region), quantile estimates must maintain mathematical coherence:

• Bottom-up reconciliation: Summing lower-level quantile forecasts does not preserve the target quantile at the aggregate level due to convolution effects
• Optimal reconciliation methods (e.g., MinT) must be extended to handle quantile-specific constraints
• Without reconciliation, aggregate prediction intervals can be miscalibrated, leading to incorrect safety stock allocations across the distribution network
• Modern approaches use copula-based or game-theoretic reconciliation to ensure coherent probabilistic forecasts at every hierarchy level

This is critical for global retailers managing inventory across thousands of SKUs and multiple distribution centers.