Sorting-inspired data prioritization cuts supply chain costs in simulation study

A new study published in Discover Computing on 26 October 2026 proposes a hybrid framework that combines a data prioritization mechanism inspired by sorting principles with neural network forecasting inside a single cost-minimization optimization model, reporting substantial reductions in forecast error and total cost in simulation experiments. The research, led by Saeed Bahrami with colleagues including Majid Toulabi, Farnaz Javadi Gargari, Hossein Amoozad-Khalili and Adel Pourghader Chobar, was tested on synthetic supply chain scenarios rather than real-world data, a limitation the authors openly acknowledge.

The central idea of the framework is a composite priority score, denoted δij, calculated for each product-location pair from three supply chain attributes: demand, transportation cost and inventory level. Each attribute is normalized and combined through weighted averaging, with the weights summing to one, so that different prioritization logics—such as favouring low inventory, high demand or low cost—can be selected by adjusting the weights. Crucially, the score is not merely a preprocessing step. It is embedded directly into the mathematical optimization model as an upper bound on assignment variables, meaning that a product-location pair can only receive shipments if its priority score is sufficiently high. The authors describe this formal mathematical linkage between data characteristics and optimization decisions as the primary methodological novelty of the work.

The forecasting component relies on a feedforward neural network with 12 input neurons corresponding to demand values from the previous 12 time periods, two hidden layers with 64 and 32 neurons using rectified linear unit activation, and a single linear output neuron producing the forecasted demand. The network was trained with the Adam optimizer at a learning rate of 0.001, a batch size of 32, and up to 200 epochs with early stopping after a patience of 10 to prevent overfitting. Data were split into 80 percent training, 10 percent validation and 10 percent test sets, with inputs normalized by Min-Max scaling and mean squared error used as the loss function. The optimization model itself minimizes total cost, comprising transportation costs, inventory holding costs and backorder penalties, subject to inventory balance, vehicle capacity and binary assignment constraints, and was solved with Gurobi.

Importantly, the optimization model uses only forecasted demand for decision-making, not actual demand, which is revealed only afterwards for evaluation. The authors note that this maintains non-anticipativity, ensuring no future information leaks into the decision process. The framework was evaluated on 10 numerical examples simulating diverse supply chain scenarios with varying demand patterns, inventory constraints and transportation networks, and each instance was replicated 30 times with different random seeds to ensure statistical reliability. The experiments ran on an Intel Core i7 processor with 32 GB of RAM using Python 3.8 with TensorFlow/Keras.

The reported results favour the hybrid approach across all tested metrics. The framework achieved an average total cost of 12,769, a 12.1 percent reduction compared with the Single Data Science method at 14,528 and a 22.2 percent reduction compared with the Single Sorting method at 16,407. Against an ARIMA-based approach, which averaged 15,335, the reduction was 16.7 percent. The hybrid framework also delivered the highest average service level at 94.2 percent, compared with 89.6 percent for the Data Science method, 88.6 percent for ARIMA and 86.6 percent for the Single Sorting method. The corresponding backorder rate of 5.8 percent was the lowest among the methods tested, with the Single Sorting method at 13.4 percent, ARIMA at 11.4 percent and Data Science at 10.4 percent.

Forecasting accuracy improved as well. The hybrid framework reduced mean absolute error by 40.7 percent compared with the Single Sorting method, from 178.50 to 105.80, and by 23.7 percent compared with the Data Science method, from 138.70 to 105.80. Root mean square error fell by 35.4 percent, from 235.80 to 152.40, against Single Sorting and by 20.3 percent, from 191.20 to 152.40, against Data Science. The authors attribute the accuracy gains to the prioritization mechanism organizing input data before it reaches the neural network, which they argue enhances the model’s ability to capture trends, patterns and nonlinear relationships. In terms of cost structure, the hybrid framework achieved the lowest inventory holding cost at an average of 7,064, a 9.3 percent reduction compared with the Data Science method, while its transportation costs of 3,585 were slightly lower than the Data Science method’s 3,909.

Statistical validation supported the observed differences. Before analysis, the authors verified the prerequisites for analysis of variance: Shapiro-Wilk tests on residuals yielded p-values of 0.432, 0.521, 0.398 and 0.476 for the Hybrid, Sorting-Only, Data Science and ARIMA methods respectively, all above the 0.05 significance level, and Levene’s test returned a p-value of 0.612, indicating homogeneous variances. The one-way ANOVA on total costs produced an F-statistic of 28.4 with a p-value below 0.001, and a post-hoc Tukey HSD test confirmed that the hybrid framework was significantly better than all three benchmark methods at the 0.05 significance level.

Sensitivity analysis examined the framework’s robustness under varying conditions. Demand variability, expressed as the coefficient of variation, was increased from 0.2 to 0.8 in increments of 0.2, unit transportation costs were varied by plus or minus 20 and 40 percent, and maximum inventory capacity was reduced by 20 and 40 percent. Under these perturbations the hybrid framework maintained service levels above 90 percent across all scenarios, while the Single Sorting method dropped below 80 percent under high demand variability. The hybrid framework showed a maximum cost deviation of approximately 8 percent under the most extreme scenarios, compared with deviations of up to 18 percent for the Single Sorting method, and the authors report the lowest cost deviations overall at 8.2 percent under extreme scenarios, against 18.3 percent for Single Sorting, 12.4 percent for Data Science and 15.2 percent for ARIMA.

Scalability testing showed near-linear growth in computation time as problem size increased. The largest instance tested, involving 50 products and 20 locations, was solved in 920 seconds, roughly 15 minutes, which the authors describe as practical for strategic and tactical planning scenarios. For the base case, the hybrid framework’s average run time was 118 seconds, somewhat slower than ARIMA’s 95 seconds, but the authors argue that the accuracy advantage justifies the modest additional computational burden. They attribute the computational efficiency to the prioritization mechanism, which they say organizes data and reduces the effective search space for the optimization solver.

The authors are careful to note the study’s limitations. The experiments relied entirely on synthetic datasets designed to replicate realistic supply chain patterns, including seasonal demand fluctuations, trend components and random noise, and the authors state that validation on real-world data would strengthen the findings, with industrial case studies planned for future work. The framework also focuses on a cost-centric single-objective model and does not directly address sustainability metrics such as carbon emissions or broader social considerations. The authors suggest future extensions including real-time adaptive models potentially incorporating reinforcement learning, integration with blockchain and digital twins, multi-objective formulations balancing cost, emissions and social impact, and alternative forecasting architectures such as CNN-LSTM or transformer-based models. They position the framework as a decision-support tool for supply chain managers, arguing that its customizable priority weights allow organizations to tailor the approach to their own strategic priorities, whether cost reduction, service improvement or inventory optimization.

Subject of Research: A hybrid data prioritization and neural network optimization framework for cost-minimized supply chain management, evaluated on synthetic scenarios

Article Title: A hybrid data prioritization and neural network framework for cost‑optimized supply chain management

Article References: Bahrami, S., Toulabi, M., Gargari, F. J., Amoozad-Khalili, H., & Chobar, A. P. (2026). A hybrid data prioritization and neural network framework for cost‑optimized supply chain management. Discover Computing, 29(1), Article 807. https://doi.org/10.1007/s10791-026-10766-x

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10766-x

Keywords: supply chain management, neural networks, data prioritization, optimization, demand forecasting, cost minimization, service level, backorder rate, ANOVA, sensitivity analysis, scalability, synthetic data