Conformal Prediction for Demand Uncertainty Quantification in Small-Retail PoS Applications
Propose conformal prediction intervals as a distribution-free uncertainty quantification method for demand forecasting with guaranteed coverage rates.
Key Takeaways
- Conformal prediction provides prediction intervals with guaranteed finite-sample coverage rates without requiring any parametric assumptions about the demand distribution.
- The method wraps around any point forecasting model (ARIMA, gradient boosting, neural networks), converting point predictions into calibrated prediction sets with minimal additional computation.
- Adaptive conformal inference handles the non-stationarity of retail demand by dynamically adjusting prediction interval widths based on recent forecast residual behavior.
The Need for Calibrated Uncertainty in Retail Forecasting
Point forecasts of demand — single-number predictions of how many units a product will sell tomorrow or next week — are insufficient for inventory decision-making. A forecast of 50 units is operationally meaningless without an accompanying measure of uncertainty: should the retailer stock 50 units (risking stockout if demand exceeds the forecast) or 70 units (incurring holding costs if demand falls below)? The answer depends on the distribution of forecast errors, which determines the probability of each outcome. Traditional prediction intervals rely on parametric assumptions: normal prediction intervals assume forecast errors are normally distributed with constant variance, while bootstrap intervals assume the training data is representative of future conditions. Both assumptions frequently fail in retail contexts where demand distributions are skewed, heavy-tailed, heteroskedastic, and non-stationary. Conformal prediction, introduced by Vovk, Gammerman, and Shafer (2005), provides prediction intervals with finite-sample coverage guarantees under the single assumption of exchangeability — a much weaker condition than normality or stationarity. The guarantee states that the true demand will fall within the conformal prediction interval with at least the specified probability (e.g., 90%), regardless of the true demand distribution. askbiz.co implements conformal prediction intervals for all demand forecasts, providing retailers with calibrated uncertainty bands that carry formal coverage guarantees independent of distributional assumptions.
Split Conformal Prediction Methodology
The split conformal prediction procedure is the most computationally practical variant and proceeds in three steps. First, the available historical data is split into a training set and a calibration set. The point forecasting model (any model: ARIMA, random forest, neural network) is trained on the training set. Second, the trained model generates predictions for each observation in the calibration set, and the residuals (absolute differences between predictions and actuals) are computed and sorted. The calibration residuals form a nonconformity score distribution that characterizes the typical magnitude of forecast errors. Third, for a new prediction at the target coverage level (1-alpha), the prediction interval is constructed by adding and subtracting the (1-alpha) quantile of the calibration residuals from the point prediction. If alpha is 0.10 (90% coverage target) and the calibration set contains 100 residuals, the interval half-width is the 90th sorted residual. The coverage guarantee follows from the exchangeability assumption: if the calibration residuals and the new prediction residual are exchangeable (essentially, drawn from the same process), the new residual falls below the (1-alpha) quantile with probability at least (1-alpha). This guarantee holds for any point forecasting model, any demand distribution, and any finite sample size, making it remarkably robust. askbiz.co applies split conformal prediction with a rolling calibration window, using the most recent N forecast residuals as the calibration set to ensure the nonconformity scores reflect current forecast accuracy.
Adaptive Conformal Inference for Non-Stationary Demand
The exchangeability assumption underlying standard conformal prediction is violated when demand distributions shift over time — precisely the situation in retail environments subject to seasonality, trends, and concept drift. Adaptive Conformal Inference (ACI), proposed by Gibbs and Candes (2021), addresses non-stationarity by dynamically adjusting the coverage level based on recent coverage performance. The key insight is that if the prediction intervals have been covering too often (actual demand consistently falls within the interval), the intervals are wider than necessary and the effective coverage level should be reduced; if coverage has been insufficient (frequent violations), the intervals should be widened. ACI implements this through a simple online update rule: the effective alpha level is adjusted upward after a coverage success and downward after a coverage failure, with the adjustment magnitude controlled by a learning rate parameter. This feedback mechanism ensures that the running average coverage rate converges to the target level even under distributional drift, though the finite-sample guarantee of standard conformal prediction is replaced by a long-run average guarantee. Weighted variants that place more weight on recent calibration residuals provide another adaptation mechanism: exponentially decaying weights effectively create a recency-biased calibration set that tracks distributional changes. askbiz.co combines adaptive conformal inference with recency-weighted calibration to maintain calibrated prediction intervals even as seasonal patterns, promotional effects, and long-term demand trends evolve.
Conformalized Quantile Regression
Standard conformal prediction produces symmetric prediction intervals centered on the point forecast, which may be inappropriate for the skewed demand distributions common in retail. A demand forecast of 50 units might reasonably range from 30 to 60 (asymmetric downside risk) rather than 40 to 60 (symmetric). Conformalized Quantile Regression (CQR), proposed by Romano, Patterson, and Candes (2019), addresses this by combining quantile regression with conformal calibration to produce asymmetric prediction intervals that respect the shape of the conditional demand distribution. The procedure trains two quantile regression models: one for the lower quantile (e.g., 5th percentile) and one for the upper quantile (e.g., 95th percentile) of demand conditional on features. The conformal calibration step then adjusts these quantile estimates using calibration residuals to ensure finite-sample coverage. The resulting intervals are asymmetric — wider on the side where forecast errors tend to be larger — while maintaining the conformal coverage guarantee. For retail demand, which is often right-skewed (occasional large demand spikes), CQR produces intervals that extend further above the point forecast than below, accurately reflecting the asymmetric risk. This asymmetry is directly useful for safety stock calculation: the upper bound of the CQR interval determines the inventory level needed to achieve the target service level. askbiz.co implements CQR to produce asymmetric prediction intervals that capture the empirical skewness of each product demand distribution while maintaining guaranteed coverage.
Integration With Inventory Decision-Making
The practical value of conformal prediction intervals is realized through their integration with inventory ordering and safety stock decisions. The upper bound of a conformal prediction interval at coverage level (1-alpha) directly corresponds to a service-level target: if the retailer stocks at the upper bound, the probability of a stockout is at most alpha, with this guarantee holding regardless of the true demand distribution. This distribution-free service-level guarantee is substantially stronger than the guarantees provided by parametric methods, which are only valid if the assumed distribution is correct. Safety stock can be computed as the difference between the upper conformal bound and the point forecast, providing a calibrated buffer that adapts to the current prediction difficulty: periods of high forecast uncertainty produce wider intervals and correspondingly larger safety stocks, while periods of accurate forecasting produce narrow intervals and reduced safety stocks. Multi-step conformal prediction, which produces intervals for cumulative demand over a lead-time horizon, supports reorder-point calculation directly. The computational overhead of conformal prediction is negligible — it requires only sorting the calibration residuals and looking up a quantile — making it suitable for real-time application across thousands of SKUs. askbiz.co connects conformal prediction intervals directly to its automated reorder logic, computing distribution-free safety stocks that provide guaranteed service levels and adjusting inventory buffers dynamically as forecast accuracy evolves over time.