New AI Framework Closes the Gap Between Prediction and Decision in Supply Chains

Every large logistics network lives with an uncomfortable secret: the machine learning models that forecast demand and the optimization engines that decide what to stock are usually two separate worlds. The forecaster is judged on statistical accuracy, the optimizer on cost and service levels, and nobody is accountable for what happens when the first feeds the second. A new study published in Neural Computing and Applications argues that this disconnect is not a minor inconvenience but a structural flaw, one that quietly amplifies errors until warehouses run dry or capital sits frozen in unsold inventory. Researchers Yumin Sun of Henan Polytechnic and Xiangfeng Liu of Wuhan Donghu University have now proposed an end-to-end, decision-aware framework that binds prediction, simulation, and robustness verification into a single coupled system, and they have tested it for a full year on the warehousing and distribution network of JD.com in North China.

The core problem the authors target is known in the decision-focused learning literature as the error amplification effect. In the conventional two-stage pipeline, a time-series model such as AutoRegressive Integrated Moving Average, or ARIMA, produces a demand forecast, and that forecast is then handed to a linear programming solver that computes inventory and replenishment decisions. Because the forecasting model is trained to minimize prediction error rather than decision error, small deviations in the demand distribution can translate into disproportionately large losses downstream. A forecast that looks excellent on paper, measured by mean absolute percentage error, may still push the optimizer toward reorder points that trigger stockouts during demand spikes or excessive holding costs during lulls. The optimization objective, in other words, never gets a chance to tell the prediction model what actually matters.

Sun and Liu’s answer is a loss function that speaks the language of the optimizer. Instead of scoring forecasts against historical observations alone, the framework constructs a Wasserstein distance-constrained loss guided by the gradient of the optimization objective. The Wasserstein distance, drawn from optimal transport theory, measures how much probability mass must be moved to transform one distribution into another, which makes it a natural yardstick for how far a predicted demand distribution lies from the true one. By constraining this distance while simultaneously following the gradient of the decision objective, the training process learns distributions that are not merely accurate in a statistical sense but useful for the inventory control problem they are meant to solve. The coupling goes deeper still: the dual variables of the stochastic programming formulation, which encode the shadow prices of constraints such as warehouse capacity and service-level requirements, are fed back into the splitting criterion of a gradient boosting tree. In effect, the tree that predicts demand grows its branches according to how sensitive the optimal decision is to each split, so the most decision-relevant structure of the demand process is learned first.

The second pillar of the framework addresses a subtler failure mode: the mismatch between the simulation environment used to evaluate policies and the physical system in which those policies ultimately run. Discrete event simulators are indispensable for testing inventory strategies before deployment, but they are built on transition probability matrices that inevitably drift from reality. A policy that shines in simulation can collapse in production if the simulator overestimates processing speeds or underestimates demand volatility. To close this gap, the authors design a simulation bias compensator based on Kullback-Leibler divergence, a measure from information theory that quantifies how one probability distribution diverges from a reference distribution. The compensator continuously measures the KL divergence between simulated and observed system behavior and adjusts the simulator’s transition probabilities accordingly.

The calibration itself is performed through multi-fidelity Bayesian optimization, a technique that intelligently balances cheap, low-fidelity simulation runs against expensive, high-fidelity ones. Bayesian optimization builds a probabilistic surrogate of the calibration objective and uses an acquisition function to decide where to sample next, which means the simulator can be re-tuned far more frequently than would be feasible with brute-force search. The result, reported in the study, is a dramatic improvement in what the authors call Simulation-Reality Correlation: the mean value of this metric rose from 0.63 under the baseline setup to 0.89 with the compensator in place. That jump means the digital twin of the supply chain became substantially more trustworthy as a decision-making instrument, narrowing the evaluation gap that has long plagued simulation-based operations research.

The third pillar is robustness verification, built around the Conditional Value-at-Risk, or CVaR, index. Unlike simpler risk measures, CVaR captures the expected loss in the worst-case tail of a distribution, which is precisely the regime where supply chains fail catastrophically. By embedding a CVaR constraint into the decision problem, the framework explicitly limits the expected shortfall under adverse demand scenarios rather than optimizing only for the average case. This gives planners a principled dial for the trade-off between efficiency and resilience: tightening the CVaR bound produces more conservative inventory buffers, while relaxing it leans into cost efficiency. Crucially, because the robustness mechanism sits inside the same end-to-end loop as prediction and simulation, the demand model learns to produce distributions whose tail behavior is compatible with the risk constraints, rather than treating risk as an afterthought bolted onto a point forecast.

The empirical evidence comes from a one-year deployment on JD.com’s North China warehousing and distribution network, one of the most demanding e-commerce logistics environments in the world, characterized by extreme promotional spikes, rapid fulfillment promises, and enormous SKU diversity. Compared with the two-stage baseline of ARIMA forecasting plus linear programming, the proposed method raised the order fulfillment rate to 92.4 percent and cut monthly stockout costs to 809,000 yuan. These are not marginal gains; they represent the difference between a supply chain that constantly fights fires and one that anticipates them. The authors attribute the improvement directly to the suppression of error accumulation: because the prediction model is trained against the decision objective and the simulator is continuously recalibrated, the errors that would normally compound across the pipeline are caught and corrected within the loop.

Equally important is the finding that the system remains computationally efficient. End-to-end approaches in decision-focused learning have sometimes been criticized for the burden of differentiating through an optimization layer at every training step, which can make them impractical at industrial scale. By routing the optimization gradient through a Wasserstein-constrained loss and a gradient boosting tree rather than through a fully differentiable solver, and by using multi-fidelity Bayesian optimization to keep simulator calibration cheap, the framework achieves its coupling without prohibitive overhead. The authors report that the ternary collaborative iterative mechanism, linking prediction, optimization, and simulation, effectively suppresses error accumulation in the decision step while enhancing adaptability to demand fluctuations and distributional shifts, the very phenomena that made traditional forecasting brittle in the first place.

The broader significance of the work lies in the path it sketches for intelligent decision-making in complex supply chain systems. Operations research has spent decades producing powerful optimization methods, and machine learning has spent the last decade producing powerful predictive ones, but the two communities have often talked past each other. This study, appearing in a special issue on machine learning and big data analytics for IoT security and privacy, demonstrates that the integration can be both theoretically verifiable and engineering-applicable: the Wasserstein constraint, the KL-divergence compensator, and the CVaR bound each carry formal guarantees, while the one-year industrial deployment shows the machinery survives contact with real warehouses, real trucks, and real customers. For an industry where a single percentage point of fulfillment rate can be worth millions, the message is likely to travel fast.

There are, of course, open questions. The framework was validated on one company’s network in one region, and extending the approach to multi-echelon global supply chains, where lead times are longer and data sparser, remains future work. The reliance on dual variables from stochastic programming also assumes the underlying optimization model is a faithful representation of operational constraints, which demands careful modeling discipline. Still, the study offers a concrete template for what decision-aware, simulation-calibrated, risk-verified supply chain intelligence can look like in practice. As distributional shifts become the norm rather than the exception, whether driven by pandemics, geopolitical disruption, or simply the whiplash of e-commerce promotions, the ability to train forecasts that know what decisions they will serve may prove to be one of the most consequential ideas in modern operations research.

Subject of Research: Decision-aware end-to-end machine learning for supply chain inventory optimization

Article Title: Decision-aware end-to-end joint modeling for supply chain optimization: integrating prediction, simulation, and robustness verification

Article References: Sun, Y., & Liu, X. (2026). Decision-aware end-to-end joint modeling for supply chain optimization: integrating prediction, simulation, and robustness verification. Neural Computing and Applications, 38(19), Article 775. https://doi.org/10.1007/s00521-026-12497-y

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12497-y

Keywords: decision-aware learning, supply chain optimization, prediction-optimization coupling, Wasserstein distance, simulation bias compensation, Bayesian optimization, Conditional Value-at-Risk, inventory control, gradient boosting, operations research, JD.com, distributional shift