Demand Forecasting with Conformal Prediction: Reliable Uncertainty Intervals for Inventory Optimization

TL;DR: A point forecast without an uncertainty interval is an implicit claim of perfect knowledge that leads to both stockouts and excess inventory. Conformal prediction produces prediction intervals with mathematical coverage guarantees (e.g., 90% of actual values fall within the interval) without requiring any distributional assumptions -- and it works as a wrapper around any existing forecast model, making it the fastest path to honest uncertainty in demand planning.

The Number Without a Range

A mid-size apparel retailer runs a gradient-boosted model to forecast weekly demand across 12,000 SKUs. The model is reasonably accurate. Mean absolute percentage error sits around 22%, which is respectable for fashion with its trend cycles and weather dependence. The data science team is satisfied. The planning team uses the forecasts to set reorder quantities. The supply chain runs on these numbers.

The model predicts 1,000 units for a particular jacket next month. The planning team orders 1,000 units. They receive 1,000 units. Actual demand turns out to be 1,400. They stock out in week three. Lost revenue: $28,000. Customer disappointment: unmeasured but real. The product page shows "out of stock" during the peak selling window.

The following month, a different SKU. The model predicts 800 units. They order 800. Demand comes in at 450. Three hundred and fifty units sit in the warehouse. Carrying cost: $4,200. Eventual markdown to clear: 40% off. Margin erosion: $7,350.

Both outcomes stem from the same structural failure. The model produced a point forecast — a single number — without communicating how uncertain that number was. The jacket forecast of 1,000 might have carried an honest uncertainty range of 700 to 1,500. The slower SKU's forecast of 800 might have had a much tighter range of 650 to 900. These ranges contain the information the planning team actually needs. They never received it.

This is the default state of demand forecasting in most organizations. Models produce numbers. Planning teams treat those numbers as facts. Nobody quantifies the gap between what the model knows and what it does not.

The problem is not that uncertainty is difficult to explain to planners. The problem is that most methods for generating uncertainty intervals are themselves unreliable — built on assumptions that do not hold in the real world. Gaussian residuals. Stationary error distributions. Homoscedastic noise. These assumptions make the math clean and the intervals wrong.

What follows is an alternative approach. One that produces prediction intervals with finite-sample coverage guarantees, requires no distributional assumptions, and works with any underlying forecast model. It is called conformal prediction, and it is the most underutilized tool in applied demand forecasting.

The Asymmetric Cost of Being Wrong

Before discussing methods, the economics matter. In inventory planning, forecast errors are not symmetric. Overestimating demand and underestimating demand produce fundamentally different costs, and the ratio between those costs determines how uncertainty should translate into ordering decisions.

Underestimating demand — a stock-out — produces lost revenue at the product's full margin, potential loss of customer lifetime value, damage to marketplace search ranking (on platforms like Amazon, stock-outs suppress organic visibility for weeks afterward), and expedited shipping costs if emergency replenishment is possible.

Overestimating demand — excess inventory — produces carrying costs (warehousing, insurance, opportunity cost of capital), markdown losses when excess is eventually cleared, and in perishable or fashion categories, complete write-offs for unsaleable goods.

The ratio varies by category, but the direction is consistent: stock-out costs almost always exceed overstock costs, often by multiples. Luxury and high-margin goods show the starkest asymmetry. A missed sale on a $400 jacket with 70% gross margin costs $280 in lost contribution. Excess inventory on that same jacket, eventually cleared at a 40% markdown, costs roughly $60 in margin erosion plus carrying costs.

This asymmetry has a direct implication for how forecasts should be used. If the cost of underestimating demand is three times the cost of overestimating it, then the optimal order quantity is not the point forecast. It is a quantity shifted upward from the point forecast, calibrated to the specific cost ratio. The question is: how far upward? And the answer depends entirely on the shape and reliability of the uncertainty interval around the forecast.

An unreliable interval — one that claims 90% coverage but actually covers 72% of outcomes — produces systematically wrong ordering decisions. The planning team thinks they are protected against stock-outs at a 95th-percentile level. They are actually exposed at a level they never agreed to. This is the problem conformal prediction solves. When demand forecasts feed operational decision systems, the difference between nominal and actual coverage becomes a direct driver of business outcomes.

The table shows a pattern that holds across most demand forecasting applications. Gaussian-assumption intervals consistently undercover — the stated coverage is an optimistic fiction. Conformal intervals achieve near-nominal coverage by construction. The difference in excess stock-out rate is the difference between reliable and unreliable inventory planning.

Why Traditional Prediction Intervals Lie

The most common approach to generating prediction intervals is parametric: assume the forecast errors follow a known distribution (usually Gaussian), estimate the parameters of that distribution from historical residuals, and construct intervals using quantiles of the assumed distribution.

For a Gaussian assumption, the 90% prediction interval is the point forecast plus or minus 1.645 times the estimated standard deviation of residuals. It is clean. It is familiar. And it is wrong in at least four ways that matter for demand forecasting.

Problem 1: Demand residuals are not Gaussian. Demand data is bounded below by zero, often right-skewed (occasional spikes from promotions or viral events), and heavy-tailed. A Gaussian distribution assigns symmetric probability to outcomes above and below the mean, and vanishingly small probability to extreme events. Real demand distributions are asymmetric, and extreme events — a TikTok mention, a competitor stock-out, an unexpected weather pattern — occur far more frequently than the Gaussian tail predicts.

Problem 2: Residual variance is not constant. Heteroscedasticity is the norm in demand forecasting. High-volume SKUs have larger absolute errors than low-volume SKUs. Promotional periods have higher variance than baseline periods. New products have wider uncertainty than established ones. A single variance estimate applied across all predictions produces intervals that are too wide for stable SKUs and too narrow for volatile ones.

Problem 3: Model misspecification biases the residuals. If the point forecast model is systematically wrong — underforecasting during promotions, overforecasting during off-season — the residuals inherit that bias. Parametric intervals built on biased residuals shift the interval center but do not correct the coverage problem. The interval is centered on a wrong estimate and shaped by an incorrect distribution.

Problem 4: Finite-sample uncertainty in the variance estimate itself. When the calibration set is small — as it often is for new products or slow-moving SKUs — the estimated variance is itself highly uncertain. Parametric methods treat the variance estimate as known, compounding the coverage error.

Bootstrap prediction intervals offer a partial improvement. Instead of assuming a parametric distribution, they resample from the empirical residual distribution and construct intervals from the resampled quantiles. This relaxes the Gaussian assumption but introduces its own problems: bootstrap intervals assume the residuals are exchangeable (identically distributed and independent), they struggle with temporal dependence in time series, and they provide no finite-sample guarantees — they are asymptotically valid, meaning they work well when you have a lot of data, which is precisely the condition under which you need them least.

Quantile regression takes a different approach entirely. Instead of modeling the conditional mean and then building an interval around it, quantile regression directly models specific quantiles of the conditional distribution. You fit one model for the 5th percentile and another for the 95th percentile, and the interval is defined by the gap between them. This handles heteroscedasticity naturally (because each quantile is modeled as a function of covariates) and makes no distributional assumptions. But quantile regression provides no coverage guarantee — the intervals are as reliable as the models that produce them. If the quantile model is misspecified, coverage degrades with no diagnostic signal.

All three methods share a common weakness: they offer no formal, finite-sample guarantee that the stated coverage level will be achieved. Conformal prediction does.

Conformal Prediction: The Distribution-Free Alternative

Conformal prediction was introduced by Vladimir Vovk, Alexander Gammerman, and Glenn Shafer in the early 2000s (see the Conformal Prediction tutorial). The foundational idea is deceptively straightforward: use the model's own past performance to construct prediction intervals that are guaranteed to contain the true outcome at a specified rate, with no assumptions about the data-generating distribution.

The guarantee is not asymptotic. It does not require large samples. It holds in finite samples under a single, mild assumption: that the data points are exchangeable — roughly, that the order in which they arrive does not matter. For time series (where exchangeability does not strictly hold), adaptive variants restore valid coverage, as discussed later.

Here is the core mechanism. Split conformal prediction works in three steps:

Step 1: Split the historical data into a training set and a calibration set. The training set is used to fit the demand forecast model. The calibration set is held out — the model never sees it during training.

Step 2: Compute nonconformity scores on the calibration set. For each observation in the calibration set, compute a "score" that measures how poorly the model's prediction fits reality. The simplest choice is the absolute residual: |actual demand minus predicted demand|. More sophisticated choices are possible and sometimes preferable, but absolute residuals work well in practice.

Step 3: Construct the prediction interval using the quantile of calibration scores. For a new prediction, the conformal interval is defined as:

where $\hat{q}$ is the $\lceil (1 - \alpha)(1 + 1/n) \rceil$-th smallest value among the calibration scores, $n$ is the calibration set size, and $\alpha$ is the desired miscoverage rate (e.g., $\alpha = 0.1$ for 90% coverage).

The coverage guarantee states:

This follows from a simple combinatorial argument. If the calibration scores and the new observation's score are exchangeable, then the new score is equally likely to fall at any rank among the calibration scores. The probability that it exceeds the $(1-\alpha)$-quantile is at most $\alpha$. Therefore, the prediction interval covers the true value with probability at least $1 - \alpha$.

The conformity score (nonconformity score) for each calibration point $i$ is:

There is an important subtlety. The guarantee is marginal, not conditional. It says: across all future predictions, at least 90% will be covered. It does not say: for this specific SKU on this specific week, the probability of coverage is 90%. The distinction matters, and we will address it shortly. But the marginal guarantee alone is a significant improvement over methods that provide no guarantee at all.

Split Conformal Prediction in Practice

The practical implementation of split conformal prediction for demand forecasting involves decisions that affect both interval quality and computational cost. Here is the workflow applied to a multi-SKU demand forecasting system.

Consider a retailer forecasting weekly demand for 5,000 SKUs. Historical data covers 104 weeks (two years). The base model is LightGBM with features including lagged demand, price, promotional flags, day-of-week effects, and category-level seasonality indicators.

Data split: Reserve the last 12 weeks as the calibration set. Train the LightGBM model on the first 92 weeks. Generate predictions for all 5,000 SKUs across the 12 calibration weeks. This produces 60,000 calibration residuals.

Score computation: For each calibration point, compute the absolute residual. The distribution of these 60,000 scores captures the model's empirical error pattern — including heteroscedasticity, skewness, and any systematic biases.

Interval construction: For a new prediction at the 90% level, find the 91st percentile of the 60,000 calibration scores (the ceiling of 0.90 times 60,001 divided by 60,000). Call this value q. The prediction interval for any new point forecast f is [f - q, f + q].

This basic version produces intervals of constant width across all SKUs and all time periods. A jacket predicted to sell 50 units and a t-shirt predicted to sell 5,000 units both get the same absolute interval width. This is valid — the coverage guarantee holds — but it is inefficient. The intervals are too wide for predictable SKUs and too narrow for volatile ones.

The fix is to use normalized nonconformity scores. Instead of the absolute residual, compute:

where $\hat{\sigma}(x_i)$ is an estimate of the local difficulty of prediction for observation $x_i$. Common choices for sigma(x) include: the predicted value itself (producing intervals that scale proportionally to forecast magnitude), a separate model trained to predict absolute residuals given features, or the rolling standard deviation of recent residuals for each SKU.

With normalized scores, the interval becomes [f - q * sigma(x_new), f + q * sigma(x_new)], which adapts to the difficulty of each specific prediction. High-variance SKUs get wider intervals. Stable SKUs get tighter ones. Coverage still holds marginally.

The chart shows the fundamental difference between Gaussian intervals (constant width, blind to SKU-level volatility) and normalized conformal intervals (adaptive width, honest about difficulty). The Gaussian approach overcovers stable SKUs (wasting inventory budget on unnecessary safety stock) and undercovers volatile SKUs (exposing the business to stock-outs precisely where they are most likely). Conformal prediction allocates uncertainty budget where it is actually needed.

Coverage Guarantees and Their Limits

The coverage guarantee of conformal prediction deserves precise understanding, including what it does and does not promise. Bayesian A/B testing operates on a similar philosophical foundation — quantifying what you know and do not know — though it applies to experimental decisions rather than prediction intervals.

What it guarantees: For any exchangeable calibration set and new test point, the probability that the true value falls within the conformal interval is at least 1 - alpha. This holds for any data distribution, any model, any sample size (subject to a small correction factor of 1/n that vanishes as the calibration set grows).

What it does not guarantee: Conditional coverage. The marginal guarantee says that across all predictions, coverage averages at least 1 - alpha. It does not say that coverage is 1 - alpha for each specific subgroup — each SKU, each week, each category. A conformal predictor achieving 90% marginal coverage might cover 99% of stable SKUs and only 75% of volatile ones, and the guarantee would still hold because the overall average exceeds 90%.

This is not a theoretical concern. In demand forecasting, conditional coverage failures are common and consequential. If the interval systematically undercovers new-product launches or promotional periods, the planning team is under-protected precisely when uncertainty is highest.

The table shows a realistic conditional coverage breakdown. The marginal guarantee is met (90.2% overall), but the coverage is distributed unevenly. Promotional weeks and new products — the segments where accurate intervals matter most for inventory decisions — receive the worst coverage.

Three approaches address this problem:

1. Group-conditional conformal prediction. Run separate conformal procedures for each subgroup: one calibration set for promotional periods, one for new products, one for stable SKUs. Each group gets its own quantile threshold, tailored to its specific error distribution. This achieves approximate conditional coverage within each group, at the cost of smaller effective calibration sets per group.

2. Conformalized quantile regression (CQR). Fit quantile regression models for the lower and upper bounds, then apply a conformal correction. The quantile regression captures heterogeneity in interval width as a function of covariates, and the conformal layer adds a finite-sample coverage guarantee on top. This is the current state-of-the-art for conditional coverage in applied settings.

3. Mondrian conformal prediction. A formalization of group-conditional conformal that allows overlapping groups and provides coverage guarantees within each predefined partition of the feature space.

For demand forecasting, conformalized quantile regression offers the strongest practical combination of adaptive width and coverage reliability. It inherits the flexibility of quantile regression (intervals that respond to SKU characteristics, seasonality, and promotional context) while adding the distribution-free guarantee of conformal prediction.

The Newsvendor Problem with Conformal Intervals

The newsvendor problem is the foundational model connecting demand uncertainty to optimal inventory decisions. A retailer must decide how many units to order before observing demand. Ordering too many incurs overage cost c_o per excess unit. Ordering too few incurs underage cost c_u per unit of unmet demand. The optimal order quantity depends on the demand distribution.

The classical solution is elegant. The optimal order quantity Q* is the quantile of the demand distribution at level:

where $F$ is the cumulative distribution function of demand. The ratio $\frac{c_u}{c_u + c_o}$ is called the critical ratio. When stock-out costs dominate (c_u >> c_o), the critical ratio approaches 1 and the optimal quantity is high. When overstock costs dominate, the critical ratio is lower.

The connection to conformal prediction is direct. The conformal prediction interval at coverage level 1 - alpha provides an estimate of the demand quantiles. If we set alpha to match the newsvendor critical ratio, the upper bound of the conformal interval is a natural order quantity.

Specifically, for a product where the cost of under-ordering is three times the cost of over-ordering (c_u = 3, c_o = 1), the critical ratio is 0.75. The conformal upper bound at 75% coverage — guaranteed to exceed true demand at least 75% of the time — provides the order quantity.

The practical value of this connection is that the conformal guarantee directly translates into an inventory service level guarantee. If conformal prediction at 90% coverage truly covers 90% of demand realizations, then ordering at the conformal upper bound achieves a 90% service level — meaning demand is fully satisfied 90% of the time. The guarantee is distribution-free. It does not depend on demand being Gaussian, or stationary, or independent across periods.

Compare this to the standard approach: fit a forecast model, assume Gaussian errors, compute a safety stock as z * sigma where z is the normal quantile and sigma is estimated from residuals. If the Gaussian assumption is wrong (it usually is), the realized service level deviates from the target, sometimes severely. Conformal prediction removes the weakest link in this chain.

Adaptive Conformal Inference for Non-Stationary Demand

The exchangeability assumption underlying standard conformal prediction is violated in time series settings. Demand data exhibits trends, seasonality, regime changes, and other forms of non-stationarity. The calibration residuals from six months ago may not reflect the model's current accuracy.

Adaptive conformal inference (ACI), introduced by Gibbs and Candes in 2021, addresses this by dynamically adjusting the coverage level based on recent performance. The key idea: instead of using a fixed quantile threshold from the calibration set, ACI maintains a running threshold that increases when the model undercovers (recent intervals miss too many actuals) and decreases when it overcovers (recent intervals are too conservative).

The update rule is:

alpha_t+1 = alpha_t + gamma * (err_t - alpha)

where alpha_t is the current miscoverage rate, err_t is an indicator for whether the interval at time t failed to cover the true value (1 if it missed, 0 if it covered), alpha is the target miscoverage rate, and gamma is a learning rate controlling adaptation speed.

When the model is undercovering (err_t = 1 more often than alpha), alpha_t increases, which widens the interval. When the model is overcovering, alpha_t decreases, tightening the interval. Over time, the running coverage converges to the target.

For demand forecasting, ACI handles several practical scenarios that static conformal prediction cannot:

• Promotional periods with higher forecast error: ACI widens intervals as the model begins missing during promotions, then tightens them as baseline accuracy resumes.
• Seasonal transitions where the model performs worse (e.g., predicting the onset of winter demand): ACI detects the coverage slip and adapts.
• Model drift where a deployed model gradually degrades: ACI compensates by widening intervals, providing an implicit monitoring signal.

The learning rate gamma controls the trade-off between responsiveness and stability. A high gamma (0.05-0.1) responds quickly to coverage failures but produces volatile interval widths. A low gamma (0.005-0.01) produces stable intervals but is slow to detect regime changes. In practice, gamma values between 0.01 and 0.05 work well for weekly demand forecasting.

Method Comparison: Quantile Regression vs. Conformal vs. Bootstrap

Three families of methods compete for generating prediction intervals in demand forecasting. Each has distinct strengths and weaknesses, and the right choice depends on operational requirements.

In head-to-head empirical comparisons on retail demand data, the results are consistent. Quantile regression produces the sharpest intervals when the model is well-specified, but its coverage degrades unpredictably when it is not. Bootstrap intervals provide reasonable coverage for stationary data but struggle with time series structure and are computationally expensive for large SKU sets. Split conformal prediction provides reliable coverage with minimal implementation effort but produces constant-width intervals unless normalized scores are used. Conformalized quantile regression combines the best properties: adaptive interval width from quantile regression with coverage guarantees from conformal prediction.

The ideal method achieves the nominal coverage level (90% bar reaching 90) with the smallest possible interval width. CQR achieves both: its coverage is near-nominal (90.1%) and its intervals are tighter than split conformal (267 vs. 334) because the quantile regression component adapts width to prediction difficulty. The Gaussian approach misses badly on coverage (76.3%) and its intervals are not even particularly tight — they are poorly allocated, being too wide for easy predictions and too narrow for hard ones.

Implementation with MAPIE

MAPIE (Model Agnostic Prediction Interval Estimator) is an open-source Python library that implements conformal prediction methods on top of scikit-learn-compatible models. It reduces the implementation of conformalized prediction intervals from a research exercise to a configuration choice.

The core workflow for demand forecasting with MAPIE follows a straightforward pattern. You train a base model (any scikit-learn-compatible regressor), wrap it in a MAPIE regressor, and the library handles calibration, score computation, and interval construction.

import numpy as np
from lightgbm import LGBMRegressor
from mapie.regression import MapieQuantileRegressor
 
# Base model: LightGBM for demand forecasting
base_model = LGBMRegressor(
    n_estimators=500, learning_rate=0.05,
    num_leaves=31, min_child_samples=20
)
 
# Conformalized Quantile Regression via MAPIE
mapie_model = MapieQuantileRegressor(
    estimator=base_model,
    method="quantile",
    cv="split",
    alpha=0.1  # 90% prediction intervals
)
 
# X_train: features (lags, price, promo flags, seasonality)
# y_train: observed weekly demand
mapie_model.fit(X_train, y_train)
 
# Generate predictions with conformal intervals
y_pred, y_intervals = mapie_model.predict(X_test)
 
# y_intervals[:, 0, 0] = lower bound (5th percentile)
# y_intervals[:, 1, 0] = upper bound (95th percentile)
for sku_idx in range(5):
    print(f"SKU {sku_idx}: forecast={y_pred[sku_idx]:.0f}, "
          f"interval=[{y_intervals[sku_idx, 0, 0]:.0f}, "
          f"{y_intervals[sku_idx, 1, 0]:.0f}]")

For split conformal prediction, MAPIE's MapieRegressor with method="base" and cv="split" performs the exact procedure described earlier: hold out a calibration set, compute residuals, and use the quantile of absolute residuals to set interval width. For conformalized quantile regression, MAPIE's MapieQuantileRegressor fits quantile models and applies the conformal correction.

Key implementation decisions:

Calibration set size. Larger calibration sets produce tighter intervals (the quantile estimate is more precise) but leave less data for training. For demand forecasting with 104 weeks of history across thousands of SKUs, reserving 10-15% of observations for calibration is a reasonable starting point. With 60,000 calibration points, the quantile estimate is precise to within a fraction of a percent.

Score normalization. For multi-SKU forecasting where demand scales vary by orders of magnitude, normalized scores are essential. Using the predicted value as the normalizer (score = |actual - predicted| / max(predicted, 1)) produces intervals that scale proportionally to forecast magnitude.

Cross-conformal prediction. Instead of a single train/calibration split, MAPIE supports cross-conformal prediction (cv="prefit" or integer K-fold). Each fold serves as the calibration set for the model trained on the remaining folds. This uses all data for both training and calibration, eliminating the efficiency loss of the split. The coverage guarantee holds marginally across the aggregated intervals.

Integration with LightGBM. MAPIE wraps any scikit-learn-compatible estimator. LightGBM's LGBMRegressor is directly compatible. For conformalized quantile regression, use LightGBM with objective="quantile" and set alpha to the desired quantile levels.

From Forecast Uncertainty to Safety Stock

Safety stock is the buffer inventory held to protect against demand uncertainty during the replenishment lead time. Traditional safety stock formulas assume Gaussian demand:

where z is the normal quantile for the desired service level, sigma_d is the standard deviation of demand per period, and L is the lead time in periods.

This formula has three embedded assumptions. Demand variance is constant (it is not). Demand is normally distributed (it is not). The standard deviation estimate is accurate (it depends on sample size and stationarity). When these assumptions fail, the safety stock is miscalibrated.

Conformal prediction offers a direct replacement. The upper bound of the conformal interval at coverage level alpha already accounts for forecast uncertainty without distributional assumptions. The safety stock is:

Safety Stock = Upper Conformal Bound - Point Forecast

For a point forecast of 1,000 units and a 95% conformal upper bound of 1,280 units, the safety stock is 280 units. This number incorporates the actual empirical distribution of forecast errors, including skewness, heavy tails, and heteroscedasticity.

For multi-period lead times, the approach extends naturally. If lead time is L periods, generate L-step-ahead conformal intervals (using the model's multi-step forecast and corresponding calibration residuals at each horizon) and compute the safety stock from the cumulative upper bound across the lead time window.

The advantage is measurable. In a typical implementation, conformal-based safety stock reduces total inventory investment by 12-18% while maintaining the same realized service level — because the safety stock is allocated intelligently (more buffer for volatile SKUs, less for stable ones) rather than uniformly.

The Uncertainty-Aware Inventory Framework

The components described above combine into a coherent framework for inventory optimization that takes forecast uncertainty seriously. The framework has five layers:

Layer 1: Base Forecast Model. Any regression model producing point demand forecasts. LightGBM, XGBoost, Prophet, ARIMA, or an ensemble. Model quality matters — better point forecasts produce tighter conformal intervals — but the framework is agnostic to the choice.

Layer 2: Conformal Calibration. Apply conformalized quantile regression using a held-out calibration set. This produces upper and lower prediction bounds at any desired coverage level, with finite-sample guarantees. Use normalized scores to handle cross-SKU variance heterogeneity.

Layer 3: Cost-Asymmetry Mapping. For each SKU or category, estimate the stock-out to overstock cost ratio. Map this ratio to the newsvendor critical ratio to determine the appropriate conformal coverage level. High-margin items with severe stock-out costs get high coverage (wide intervals, aggressive ordering). Low-margin commodities get lower coverage (tighter intervals, leaner ordering).

Layer 4: Adaptive Monitoring. Deploy ACI to dynamically adjust interval widths as the model's real-world performance evolves. Monitor empirical coverage by SKU segment weekly. Flag segments where coverage drifts below target for model retraining or recalibration.

Layer 5: Safety Stock Translation. Convert the conformal upper bound into reorder quantities and safety stock levels. Integrate with the existing ERP or inventory management system. Replace the z * sigma formula with conformal-derived buffers.

The framework is modular. Each layer can be improved independently. A better base model tightens all intervals downstream. Better cost estimates improve ordering decisions without changing the forecast. Adaptive monitoring catches drift without manual intervention. The modularity means the framework can be adopted incrementally — start with Layer 2 (adding conformal intervals to an existing forecast) and add subsequent layers as operational maturity allows.

Case Study: Reducing Dead Stock While Maintaining Service Levels

An online home goods retailer — 8,200 active SKUs, $45 million annual revenue, 60% of sales through their own website and 40% through Amazon — faced a familiar pair of problems. Dead stock (inventory unsold after 180 days) consumed 14% of warehouse capacity. Simultaneously, stock-out rates on their top 500 SKUs averaged 8.3%, costing an estimated $2.1 million in annual lost revenue.

Their existing system used a gradient-boosted forecast model with Gaussian prediction intervals. The planning team ordered at the 90th percentile of the Gaussian interval for high-priority SKUs and at the 70th percentile for long-tail items. Despite the aggressive ordering on high-priority items, stock-outs persisted because the Gaussian intervals systematically undercovered during promotional periods and seasonal transitions.

The implementation followed the Uncertainty-Aware Inventory Framework over a 10-week rollout:

Weeks 1-2: Replaced Gaussian intervals with conformalized quantile regression using MAPIE. The base LightGBM model was unchanged. Calibration used a rolling 8-week window of recent predictions and actuals, refreshed weekly.

Weeks 3-4: Computed SKU-level cost asymmetry ratios using margin data and estimated stock-out costs (incorporating marketplace ranking penalties for Amazon-listed items). Mapped these to conformal coverage targets: 93-97% for top-500 SKUs, 80-88% for mid-tier, 70-75% for long-tail.

Weeks 5-6: Deployed ACI with gamma = 0.02 to handle seasonal transition (the rollout spanned the shift from fall to holiday demand). Monitored coverage by segment weekly.

Weeks 7-10: Full integration with the ERP system. Conformal upper bounds fed directly into reorder quantity calculations, replacing the z * sigma safety stock formula.

Results after 16 weeks of operation (compared to the prior 16-week period):

The dual improvement — fewer stock-outs and less dead stock simultaneously — is the signature outcome of properly calibrated uncertainty intervals. Gaussian intervals forced a crude trade-off: order more to avoid stock-outs (creating dead stock) or order less to avoid dead stock (creating stock-outs). Conformal intervals broke this trade-off by allocating inventory budget according to actual prediction difficulty rather than a blanket assumption.

The financial impact breaks down as follows. Reduced stock-outs on top SKUs recovered an estimated $1.4 million in annual revenue. Reduced dead stock freed 5.3 percentage points of warehouse capacity, deferring a planned warehouse expansion ($220,000 annual lease cost). Lower average days-of-inventory reduced carrying costs by approximately $380,000 annually. Improved gross margin from fewer markdowns contributed an additional $610,000.

Total annual impact: approximately $2.6 million on a $45 million revenue base, a 5.8% improvement in operating economics from changing only how forecast uncertainty was quantified and translated into ordering decisions. The base forecast model — the part that most teams spend their optimization effort on — was not touched.

The implementation also revealed an operational benefit that was not anticipated. Because conformal intervals adapt to model performance, they served as an implicit monitoring system. When the model began undercovering for a category of outdoor furniture during an unexpected warm spell in October, the ACI mechanism widened intervals for that category within two weeks — faster than the monthly model retraining cycle would have caught the drift. The planning team noticed the wider intervals and proactively increased orders, avoiding what would have been a significant stock-out during an unplanned demand surge.

This is the deeper value of distribution-free uncertainty quantification. It does not just produce better intervals. It creates a feedback loop between forecast performance and operational decisions that self-corrects in ways that parametric assumptions cannot.

References

• Vovk, V., Gammerman, A., & Shafer, G. (2005). Algorithmic Learning in a Random World. Springer.

• Romano, Y., Patterson, E., & Candes, E. (2019). Conformalized quantile regression. Advances in Neural Information Processing Systems, 32.

• Gibbs, I., & Candes, E. (2021). Adaptive conformal inference under distribution shift. Advances in Neural Information Processing Systems, 34.

• Angelopoulos, A. N., & Bates, S. (2023). Conformal prediction: A gentle introduction. Foundations and Trends in Machine Learning, 16(4), 494-591.

• Taquet, V., Blot, V., Mossina, T., Zuluaga, M. A., & Gelly, G. (2022). MAPIE: An open-source library for distribution-free uncertainty quantification. arXiv preprint arXiv:2207.12274.

• Silver, E. A., Pyke, D. F., & Thomas, D. J. (2017). Inventory and Production Management in Supply Chains (4th ed.). CRC Press.

• Zaffran, M., Feron, O., Goude, Y., Josse, J., & Dieuleveut, A. (2022). Adaptive conformal predictions for time series. Proceedings of the 39th International Conference on Machine Learning, 25834-25866.

• Stankeviciute, K., Alaa, A. M., & van der Schaar, M. (2021). Conformal time-series forecasting. Advances in Neural Information Processing Systems, 34.

• Barber, R. F., Candes, E. J., Ramdas, A., & Tibshirani, R. J. (2023). Conformal prediction beyond exchangeability. Annals of Statistics, 51(2), 816-845.

• Svetunkov, I., & Boylan, J. E. (2023). Forecasting and Inventory Management for Intermittent Demand. Wiley.