PaperlayerView exact revision on arXiv

Source-linked AI summary

Federated Ensemble Forecasting Under Supply-Chain Market Volatility

Shunmukha Sagar Puppala

arXiv:2608.21399v1cs.LG

TL;DR

Supply-chain forecasting must handle distributed data, non-identical regional demand, and shifting market regimes without centralizing commercial records. The paper proposes FEF-NCL, which trains specialized experts with negative-correlation learning and drift- and reliability-aware aggregation. On one synthetic dataset, it reports improved forecasting and delay-risk performance, while emphasizing that real-world validation and stronger deployment safeguards remain necessary.

  • Problem

    The paper addresses how to produce accurate and governable supply-chain forecasts when data are distributed, client distributions are non-identical, and market regimes change during training.

  • Method

    FEF-NCL combines multi-expert temporal forecasting, negative-correlation specialization, drift and reliability scoring, weighted aggregation, multi-task learning, and explanation review across decentralized clients.

  • Results

    12.4% WMAPE was achieved by FEF-NCL versus 13.9% for SCAFFOLD, with delay-risk macro-F1 of 0.801 and market-regime macro-F1 of 0.763 on the synthetic dataset.

  • Takeaways & Limitations

    Federated forecasting can benefit from explicit ensemble diversity when client data are non-identically distributed and market regimes shift over time.

  • Takeaways & Limitations

    All reported data and results are synthetic, so the experiment does not validate performance on a live supply-chain network; privacy analysis is also incomplete.

Abstract

from arXiv · show

Supply chain forecasting systems increasingly operate under market shocks, non-identically distributed regional demand, and limited willingness to centralize commercial data. This work proposes Federated Ensemble Forecasting with Negative-Correlation Learning (FEF NCL), a distributed method that trains specialized forecasting experts across client nodes while discouraging redundant model errors. The framework combines temporal feature encoders, client level drift scoring, reliability-weighted aggregation, and an explain ability layer that exposes the market and supplier variables most responsible for each forecast. A single synthetic dataset is used to evaluate the design. It contains 124,800 weekly SKU region observations from ten regional client nodes, 60 product families, 40 suppliers, five commodity groups, and a 2021-2024 volatility profile with explicit price-shock regimes. Because the dataset is synthetic, the reported results should be interpreted as controlled evidence of internal consistency rather than real-world validation. Across the synthetic test split, FEF NCL reduces weighted mean absolute percentage error from 13.9% for the best federated baseline to 12.4%, improves delay-risk macro-F1 from 0.755 to 0.801, and lowers the high volatility quintile error by 2.1 percentage points relative to SCAFFOLD. The analysis suggests that negative-correlation specialization is useful when clients face different supplier, freight, and commodity conditions, although deployment would require stronger privacy analysis, live drift monitoring, and operational calibration. Index Terms federated learning, ensemble learning, negative correlation learning, supply chain forecasting, market volatility, data drift, demand planning, risk governance

I. INTRODUCTION

The paper frames supply-chain forecasting as a distributed, volatility-sensitive governance problem and proposes FEF-NCL to preserve specialized knowledge across heterogeneous clients. It evaluates the architecture on a controlled synthetic dataset designed to represent regional, supplier, commodity, and market-regime variation.

  • Motivation: Market volatility makes forecasting both a modeling and governance problem because variable relationships can change within weeks.The relevant signals include historical sales, open orders, lead-time variation, price movement, and macro-shock indicators.
  • Problem: Federated learning addresses distributed commercial data, but standard aggregation can dilute minority-regime knowledge when most clients remain stable.The paper motivates a volatility-aware ensemble for networks where rare shocks are operationally important.
  • Method: FEF-NCL trains multiple temporal experts per client and penalizes aligned residuals to encourage specialization across trend, price-shock, delay-risk, and residual-demand patterns.The architecture also jointly estimates demand, delay risk, and market regime without moving raw client data.
  • Method: Drift and reliability scores reduce weight for unstable or low-quality updates while retaining useful minority-regime evidence.An explanation layer records market, supplier, and order features that drove each forecast.
  • Evaluation: 124,800 weekly observations span ten regional client nodes, 60 product families, 40 suppliers, and five commodity groups from January 2021 through December 2024.The synthetic data include calm, rising-price, and price-shock regimes plus demand, price, lead-time, supplier, freight, foreign-exchange, backlog, promotion, and volatility features.
  • Contributions: The contributions are a federated multi-expert architecture, a drift-aware negative-correlation objective, and a reproducible synthetic evaluation.The evaluation reports performance, convergence, volatility-stratified error, and explanatory feature attribution from the same dataset.

II. THEORETICAL BACKGROUND

The theoretical background connects federated learning, heterogeneity, concept drift, forecasting evaluation, ensemble diversity, and supply-chain resilience. Together, these foundations motivate specialized models and careful validation under changing regional and market conditions.

  • Federated learning: Federated learning coordinates local optimization and server aggregation while keeping client data decentralized.The background presents federated learning as a design space involving client selection, secure aggregation, and privacy protection.
  • Heterogeneity: Statistical and systems heterogeneity arise from differences in regional seasonality, supplier exposure, commodity mix, freight routes, and promotion practices.FedProx and SCAFFOLD are introduced as methods addressing non-identical clients through proximal regularization and control variates.
  • Robustness: Robust aggregation research formalizes defenses against corrupted, noisy, unrepresentative, adversarial, or faulty distributed updates.The literature also identifies a tradeoff between robustness and estimation efficiency.
  • Concept drift: Market volatility is treated as concept drift because changing feature-target relationships create mismatch between training and deployment conditions.Examples include price-driven substitution, logistics-driven lead-time changes, and supply constraints that censor observed sales.
  • Forecasting foundations: Forecasting theory emphasizes baseline evaluation, rolling-origin validation, and error metrics, while forecasting competitions show that no single method family dominates all series.The cited competitions also support ensembles, machine-learning features, and hierarchy-aware design.
  • Ensembles: Negative-correlation learning operationalizes ensemble diversity by penalizing members that produce similar residual patterns.This supports regime-specific experts for stable replenishment, freight and commodity spikes, supplier delays, and residual demand.
  • Operational lens: Supply-chain resilience research frames forecasting as a network-level concern because disruptions can affect procurement, supplier commitments, and substitutions across interconnected services.A regional forecast is therefore connected to wider network decisions rather than functioning only as a local statistic.

A. Federated supply-chain learning

Prior work establishes federated learning as useful for connected supply-chain decisions under data boundaries, while volatility-aware forecasting and federated optimization address changing conditions and heterogeneous clients. FEF-NCL builds on these strands by focusing on ensemble specialization.

  • Federated supply-chain learning: Edge-cloud pricing, secure federated forecasting, and predictive-maintenance studies motivate cross-node learning across commercially connected but data-separated parties.These studies do not fully address ensemble specialization under differing client conditions.
  • Volatility-driven forecasting: Commodity price fluctuations can alter final product pricing through lagged channels that differ by product family, supplier contract, and region.Prescriptive analytics and digital twins connect forecasts to actions but require estimates stable under external shocks.
  • Optimization: FedAvg provides the common local-training and server-aggregation template, while FedProx and SCAFFOLD address non-identically distributed clients.Federated system design also adds orchestration and client-eligibility concerns.

D. Robustness, drift, and adversarially informed validation

The paper links robust distributed learning and concept-drift literature to validation under perturbed updates and changing data-generating relationships. It also situates ensemble diversity and interpretability as complementary foundations for volatility-sensitive forecasting.

  • Robustness, drift, and adversarially informed validation: Byzantine aggregation and adversarial-learning research motivate validating distributed updates because faulty or perturbed training inputs can compromise optimization.Concept-drift literature further explains why ordinary validation can become stale as data-generating relationships change.
  • Ensembles and interpretability: Ensemble and negative-correlation learning support diverse predictors, while neural forecasting architectures exploit long temporal context, static covariates, and decomposed structure.The related literature also connects these modeling choices with interpretability concerns.

A. Dataset Analysis

The study evaluates FEF-NCL on a single synthetic, non-identically distributed supply-chain dataset spanning regional clients, products, suppliers, commodities, and volatility regimes. Its temporal protocol separates training, validation, and test periods while preserving client-local data boundaries.

  • Dataset scope: 124,800 weekly observations cover ten regional clients, 60 product families, 40 suppliers, and five commodity groups from 2021 through 2024.The dataset includes smooth seasonality, abrupt shocks, and three market-regime labels.
  • Feature construction: Features include lagged demand, rolling volatility, price and freight changes, supplier reliability, lead time, backlog, promotion depth, and interaction terms.Demand lags use 1, 4, 8, 13, and 26 weeks; rolling windows use 4, 8, and 13 weeks.
  • Evaluation protocol: The temporal split assigns weeks 1–146 to training, weeks 147–177 to validation, and weeks 178–208 to testing.The corresponding row counts are 87,600, 18,600, and 18,600, respectively.
  • Evaluation protocol: The test window contains a strong final-year volatility episode, making evaluation reflect regime shift rather than simple interpolation.Regional seasonality, supplier mix, and shock exposure are deliberately imbalanced across clients.
  • Data governance: No row-level data move from regional clients to the server in the proposed pipeline.Each regional client receives its own local table, supporting the federated data-boundary design.

B. Model Analysis

FEF-NCL formulates forecasting as multi-task sequence learning with multiple specialized experts, negative-correlation regularization, and drift-aware aggregation. The evaluation compares it with local, centralized, and federated baselines using forecasting, classification, robustness, and governance measures.

  • Model formulation: FEF-NCL predicts four-week demand, delay-risk class, and market-regime class from previous-week feature sequences.Its objective combines Huber demand loss, cross-entropy classification losses, and an ensemble diversity penalty.
  • Model formulation: The negative-correlation term discourages experts from aligning their residuals with the ensemble residual, encouraging specialization across trend, shock, delay, and residual regimes.The ensemble forecast is a gated weighted sum of expert forecasts.
  • Aggregation: Server aggregation weights accepted updates by client validation quality, data volume, and regime coverage.Client drift and reliability scores support the drift-aware aggregation procedure.
  • Architecture and baselines: The proposed model uses four local experts per client, a shared temporal encoder, and two classification heads.The baseline set includes seasonal naive, local-only Temporal Fusion Transformer, centralized gradient-boosted trees, FedAvg, FedProx, and SCAFFOLD.
  • Evaluation: Evaluation reports WMAPE, RMSE, MAE, pinball loss, macro-F1 for delay risk and market regime, volatility-quintile error, shock robustness, and client holdout performance.Governance outputs include model version and client participation information.

V. EXPERIMENTAL ANALYSIS

Across the synthetic test evaluation, FEF-NCL outperforms federated baselines on forecasting and delay-risk metrics, while ablations indicate that its gains reflect interacting components. Attribution analysis and client-holdout testing further examine governance and robustness trade-offs.

  • Main results: 12.4% WMAPE and 0.801 delay-risk macro-F1 are FEF-NCL’s main synthetic test results, versus 13.9% WMAPE for SCAFFOLD.The absolute improvement over SCAFFOLD is 1.5 percentage points of WMAPE and 0.046 macro-F1 for delay risk.
  • Volatility-stratified results: 2.1 percentage points lower high-volatility error is reported for FEF-NCL relative to SCAFFOLD.In the highest-volatility quintile, FEF-NCL records 21.3% WMAPE, compared with 23.4% for SCAFFOLD and 24.5% for FedAvg.
  • Ablation analysis: 13.1% WMAPE follows removal of the negative-correlation penalty, compared with 12.4% for the full method.The same ablation lowers regime macro-F1 from 0.763 to 0.741.
  • Ablation analysis: 12.9% WMAPE results when drift-weighted aggregation is removed, while high-volatility error increases by 1.2 percentage points.Using only one expert per client yields 13.3% WMAPE, indicating that the observed gain is not attributed to one architectural change alone.
  • Governance and robustness: 0.24 attribution share belongs to the commodity index, followed by lead time at 0.19 and supplier on-time-in-full ratio at 0.16.The shares are derived from the synthetic data-generating process and fitted model for governance review.

VI. DISCUSSION

FEF-NCL uses ensemble specialization and drift-aware aggregation to preserve useful client differences under uneven market volatility. Its evidence is controlled and synthetic, with governance support that remains basic and deployment boundaries that require caution.

  • Discussion: FEF-NCL preserves client-specific error profiles through specialized experts and drift- and regime-aware update weighting.The design uses heterogeneity rather than compressing all client differences into one global update.
  • Discussion: The method shows useful improvements over federated baselines in WMAPE, delay-risk macro-F1, convergence, and high-volatility robustness.The comparison is narrower than a general claim that federated neural ensembles outperform centralized learning.
  • Discussion: Governance logging and attribution create reviewable evidence linking forecasts to client updates, drift gates, and recognizable market, supplier, and order features.The layer supports discussion among procurement, operations, finance, and risk teams but does not guarantee fairness, causality, or regulatory compliance.
  • Discussion: The evaluation is internally consistent but synthetic, so it does not validate performance on a live supply-chain network.The privacy analysis is also incomplete because gradients and model updates can leak information without additional safeguards.
  • Discussion: Real shocks may be ambiguous or unrepresented, while the four-expert ensemble adds communication and model-management overhead.These constraints may matter for clients with limited compute or applications requiring minute-level decision latency.
  • Discussion: FEF-NCL is best viewed as a candidate architecture for high-value, cross-silo planning when data cannot be centralized and market shocks are economically material.Federated rounds must also complete before decisions are due, and participating nodes must be willing to coordinate training.

VII. CONCLUSION AND FUTURE WORKS

The paper introduces FEF-NCL for volatility-sensitive forecasting across distributed supply-chain clients and evaluates it on one synthetic dataset. It reports better performance than tested federated baselines, while framing the method as a starting point requiring real-data, privacy, and planner-centered validation.

  • Conclusion: FEF-NCL combines temporal forecasting, multi-task risk estimation, negative-correlation specialization, drift-aware aggregation, and explanation review.The controlled evaluation uses 124,800 weekly observations from ten regional client nodes across a 2021-2024 volatility profile.
  • Conclusion: 12.4% WMAPE was achieved by FEF-NCL, compared with 13.9% for SCAFFOLD, 14.2% for FedProx, and 14.8% for FedAvg.It also improved delay-risk macro-F1 to 0.801 and market-regime macro-F1 to 0.763 on the synthetic dataset.
  • Conclusion: The improvement depended on expert diversity, drift-weighted aggregation, and multi-task learning rather than any single component.This pattern comes from the reported ablation results.
  • Conclusion: Federated forecasting can benefit from explicit ensemble diversity when client data are non-identically distributed and market regimes shift over time.Negative-correlation learning adds complexity while providing a mechanism for preserving specialized knowledge about rare or local regimes.
  • Future Work: Future work should test real or semi-real data, strengthen privacy protections, study adaptive expert creation, add economic metrics, and evaluate planner understanding and trust.These directions address the gap between controlled synthetic evidence and deployment needs.
  • Future Work: Production planning also requires hierarchical reconciliation, incentive-aware forecasting, model versioning, client onboarding, drift response, and rollback procedures.The paper identifies these as additional research and deployment needs.
  • Future Work: The framework is a technically grounded starting point rather than a claim of general superiority.The next step is evaluation with real data, adversarial review, privacy testing, and planner-centered assessment.
Loading 2608.21399v1…