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 networks need accurate, governable forecasts despite distributed commercial data, non-identical client distributions, and shifting market regimes. FEF-NCL combines federated negative-correlation experts with drift-aware aggregation; on synthetic data, it achieved 12.4% WMAPE versus SCAFFOLD’s 13.9%.

  • Problem

    The paper asks how supply-chain networks can produce accurate, governable forecasts when commercial data are distributed, client distributions differ, and market regimes change.

    Source: §I · p. 1

  • Method

    FEF-NCL trains specialized forecasting experts with a negative-correlation objective and uses drift-aware, reliability-weighted aggregation across client nodes.

    Source: §I · p. 2, B. Model Analysis · p. 4

  • Results

    12.4% WMAPE for FEF-NCL versus 13.9% for SCAFFOLD on the synthetic dataset; delay-risk macro-F1 reached 0.801.

    Source: §VII · p. 7

  • Takeaways & Limitations

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

    Source: §VII · p. 7

    All reported data and results are synthetic, so the experiment does not validate performance on a live supply-chain network.

    Source: §VI · p. 6

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 introduction frames supply-chain forecasting as a volatility and governance problem: distributed clients face changing market regimes, while centralizing commercial data creates exposure. FEF-NCL addresses this with specialized federated experts and drift-aware reliability weighting, evaluated on a synthetic dataset that does not establish real-world accuracy.

I. INTRODUCTION

  • Federated forecasting must handle non-identical client data and changing market regimes, because standard aggregation can dilute rare-shock knowledge while centralized access risks exposing commercial data.Local-only models may also miss rare regimes when shocks are sparse at individual nodes.Source: §I · p. 1
  • FEF-NCL trains temporal experts for distinct demand patterns, penalizes correlated errors, weights updates by client drift and reliability, and explains which market, supplier, and order features drive forecasts.The architecture exchanges model updates rather than raw client records and is presented as a controlled model, not a complete production system.Source: §I · pp. 1–2
  • The evaluation uses 124,800 weekly synthetic observations from ten regional clients spanning 2021–2024, so its findings show behavior under controlled assumptions rather than real-world accuracy.The dataset includes 60 product families, 40 suppliers, five commodity groups, and calm, rising-price, and price-shock regimes.Source: §I · p. 2
  • The paper contributes a federated multi-expert system for demand, delay-risk, and market-regime estimation, an adapted negative-correlation objective with drift-aware aggregation, and a reproducible synthetic evaluation.The evaluation covers performance, convergence, volatility-stratified error, and explanatory feature attribution.Source: §I · p. 2

II. THEORETICAL BACKGROUND

The theoretical background combines federated learning, forecasting, and ensemble-learning foundations to motivate specialized, reliability-aware forecasts under market volatility. It also frames drift monitoring, explainability, and governance as operational requirements for supply-chain deployment.

Federated learning

  • Cross-silo federated learning fits supply chains because a small number of data-rich legal entities need to collaborate without centralizing commercially sensitive records,.Federated Averaging coordinates local stochastic-gradient steps with server aggregation; broader FL design also involves client selection, secure aggregation, privacy, heterogeneity, and operational evaluation.Source: §II · p. 2

Heterogeneity

  • Regional demand, supplier exposure, commodity mix, freight, and promotions create client heterogeneity that FedProx and SCAFFOLD address through training stabilization and drift correction.A commodity-shock update may appear unusual while carrying knowledge useful for future shocks; practical systems also need orchestration, fault tolerance, compression, and eligibility checks.Source: §II · p. 2

Robust aggregation

  • Robust aggregation addresses corrupted, noisy, or unrepresentative updates,, while weak validation can enable manipulation; reliability scoring also targets erroneous records and delayed feeds.Supply-chain risks additionally include unmodeled promotions and opportunistic reporting, giving reliability scoring both forecasting and governance purposes.Source: §II · p. 2

Concept drift

  • Market volatility is concept drift: changing feature-target relationships can make training conditions mismatch deployment, while local drift may be masked by a global average.Price changes, logistics disruptions, and supply constraints can alter substitution, feasible lead times, or the meaning of sales; preserving minority-regime signals matters.Source: §II · p. 2

Forecasting theory

  • Forecasting foundations emphasize baseline evaluation and rolling-origin validation; M4/M5 show no method family dominates all series, and deep-model gains depend on data and evaluation -.Ensembles, machine-learning features, hierarchy-aware design, multi-horizon attention, efficient long-sequence modeling, and seasonal-trend decomposition are among the approaches discussed.Source: §II · p. 2

Ensemble learning

  • Ensembles can reduce error when predictors have diverse failures, and negative-correlation learning penalizes shared residual patterns to support regime-specific experts.Potential experts may capture seasonal replenishment, freight and commodity spikes, supplier delays, or promotion effects; specialization must remain compatible with federated privacy boundaries.Source: §II · pp. 2–3

Resilience and governance

  • Supply-chain resilience spans interconnected services, motivating lifecycle governance and reviewable forecasts through drift scores, feature attributions, and audit records -.The cited governance guidance covers access control, risk identification, transparency, and accountable use; the explanation layer draws on additive attribution concepts related to SHAP.Source: §II · p. 3

III. RELATED WORKS · A. Federated supply-chain learning

Prior supply-chain studies show that cross-node learning can support commercial decisions across data boundaries, while FEF-NCL targets client-specific market regimes through ensemble diversity.

A. Federated supply-chain learning

  • Applied studies of dynamic pricing, federated demand forecasting, and aftermarket maintenance - motivate FEF-NCL’s focus on diverse client and expert specialization.Those studies do not fully address specialization when participants experience different market regimes; FEF-NCL makes diversity part of its optimization objective.Source: A. Federated supply-chain learning · p. 3

B. Market volatility and price-shock forecasting · C. Federated optimization under heterogeneity · D. Robustness, drift, and adversarially informed validation

These sections situate supply-chain forecasting amid volatile market conditions and heterogeneous clients, and connect robust aggregation to changing data and potentially unreliable updates. The paper uses a synthetic dataset with commodity and freight shocks and adds ensemble diversity and drift-weighted aggregation to established federated methods.

B. Market volatility and price-shock forecasting

C. Federated optimization under heterogeneity

  • FedAvg established local training and server aggregation, while heterogeneity and privacy remain challenges; FedProx and SCAFFOLD address non-identically distributed clients,.Source: C. Federated optimization under… · p. 3
  • Production FL adds orchestration and client-eligibility concerns; FEF-NCL uses established methods as baselines and adds ensemble diversity and drift-weighted aggregation.Source: C. Federated optimization under… · p. 3

D. Robustness, drift, and adversarially informed validation

  • Byzantine aggregation motivates update validation,, adversarial-learning surveys warn of fragility to perturbations, and concept drift can make ordinary validation stale.Source: D. Robustness, drift, and… · p. 3
  • In supply-chain forecasting, late files, supplier-process changes, inventory censoring, and sudden price changes can warrant lower or context-specific aggregation weights, beyond malicious behavior.Source: D. Robustness, drift, and… · p. 3

E. Ensemble specialization and explainable forecasting

The section situates FEF-NCL among work on diverse ensembles and neural forecasting, then describes its client-level experts and planner-facing feature attributions.

  • Ensemble research emphasizes predictor diversity, while negative-correlation learning jointly trains diverse predictors.Source: E. Ensemble specialization and… · p. 3
  • Temporal Fusion Transformers, Informer, and FEDformer exploit long temporal context, static covariates, and decomposed structure -, while forecasting surveys stress interpretability and evaluation design.Source: E. Ensemble specialization and… · p. 3
  • FEF-NCL uses multiple experts per client and reports planner-understandable attribution shares for commodity index, supplier on-time-in-full performance, lead time, and freight pressure.Source: E. Ensemble specialization and… · p. 3

IV. MATERIALS AND METHODS · A. Dataset Analysis

The methods evaluate FEF-NCL on one explicitly synthetic, non-identically distributed supply-chain dataset spanning 2021–2024. Its temporal split, market-shock features, and client-local processing define the evaluation conditions.

A. Dataset Analysis

  • The synthetic dataset contains 124,800 weekly observations across ten regional clients, 60 product families, 40 suppliers, five commodity groups, and three market regimes from Jan. 4, 2021 to Dec. 23, 2024.It combines smooth seasonality with abrupt shocks to evaluate federated learning under non-identically distributed client conditions.Source: A. Dataset Analysis · p. 3
  • The dataset targets four-week-ahead demand, three-class delay risk, and three-class market regime; 18,221 rows (14.6%) carry a binary shock flag.Delay-risk counts are 79,872 low, 29,952 medium, and 14,976 high; regime counts are 74,880 stable, 31,200 rising-price, and 18,720 price-shock rows.Source: A. Dataset Analysis · p. 4
  • Input fields span demand, prices, commodity and freight pressure, foreign exchange, lead time, supplier reliability, promotions, backlog, and market volatility.Reported ranges include demand 0–2,800 units, list price USD 7.50–940.00, commodity index 82–162, freight index 68–225, FX index 0.82–1.21, lead time 3–46 days, on-time-in-full 0.54–0.99, promotion depth 0–0.35, backlog 0–5,500 units, and volatility score 0.02–0.91.Source: A. Dataset Analysis · p. 4
  • A temporal split assigns weeks 1–146 (87,600 rows) to training, 147–177 (18,600) to validation, and 178–208 (18,600) to testing.The final-year test period includes a strong volatility episode, so test performance reflects regime shift rather than simple interpolation.Source: A. Dataset Analysis · p. 4
  • Each regional client retains its own table, with deliberately imbalanced seasonality, supplier mix, and shock exposure; no row-level data move to the server.Source: A. Dataset Analysis · p. 4
  • Feature engineering adds demand lags, rolling statistics, price-shock interactions, supplier-reliability windows, and inventory-censoring features, with continuous variables standardized locally.Demand lags use 1, 4, 8, 13, and 26 weeks; rolling means and standard deviations use 4-, 8-, and 13-week windows. The server receives model updates rather than client means or raw feature vectors.Source: A. Dataset Analysis · p. 4

B. Model Analysis

FEF-NCL frames supply-chain forecasting as multi-task sequence learning with specialized client experts, reliability-weighted aggregation, and drift controls. Its evaluation uses rolling-origin selection and a locked test window, with demand, classification, robustness, and governance measures.

Model formulation

  • The model jointly predicts four-week demand, delay risk, and market regime, combining task losses with a negative-correlation penalty that encourages expert specialization.A proximal term limits client movement from the global model, while a drift penalty downweights unstable updates.Source: B. Model Analysis · p. 4

Model architecture and aggregation

  • FEF-NCL uses four local experts per client, a shared temporal encoder, two classification heads, and a gate conditioned on volatility, commodity momentum, supplier reliability, and forecast horizon.Updates pass validation-loss, gradient-norm, and drift-score thresholds; accepted updates are weighted by validation quality, data volume, and regime coverage.Source: B. Model Analysis · p. 5

Baselines

  • Baselines include seasonal naive, local-only Temporal Fusion Transformer, centralized gradient-boosted trees, FedAvg, FedProx, and SCAFFOLD; the centralized reference violates the assumed data-boundary constraint.The federated recurrent baselines use a gated recurrent unit, with SCAFFOLD adding control variates.Source: B. Model Analysis · p. 5

Validation and evaluation

  • Evaluation selects hyperparameters by rolling-origin validation, then uses a fixed validation window and locked test window, measuring demand, classification, volatility-quintile, shock-episode, and client-holdout performance.Demand metrics are WMAPE, RMSE, MAE, and pinball loss at the 0.10, 0.50, and 0.90 quantiles; classification uses macro-F1 for delay risk and market regime. Governance outputs include model version, client participation, drift score, aggregation decision, and feature-attribution summaries.Source: B. Model Analysis · p. 5

V. EXPERIMENTAL ANALYSIS

On one synthetic dataset, FEF-NCL outperforms federated baselines overall and shows larger gains under high volatility. These controlled results are not field measurements, and client holdout and communication tests qualify the performance gains.

Experimental setup

  • Experiments average five random seeds on a single synthetic dataset, using 80 rounds, 70% client participation, three local epochs, and a four-week demand horizon.Hyperparameters are selected on validation data and fixed for testing; the negative-correlation coefficient is 0.10, selected for best validation WMAPE without reducing delay-risk macro-F1.Source: §V · p. 5

Main test results

  • 12.4% WMAPE for FEF-NCL beats the best federated baseline, SCAFFOLD at 13.9%, with delay-risk macro-F1 of 0.801.FEF-NCL also records 190.8 RMSE and 0.763 regime macro-F1; its absolute improvement over SCAFFOLD is 1.5 percentage points of WMAPE and 0.046 macro-F1 for delay risk.Source: Table III · p. 5, §V · p. 5

Volatility-stratified evaluation

  • In the highest-volatility quintile, FEF-NCL records 21.3% WMAPE versus 23.4% for SCAFFOLD and 24.5% for FedAvg.On 2,808 price-shock test rows, FEF-NCL improves WMAPE by 2.1 percentage points relative to SCAFFOLD and raises high-delay recall from 0.718 to 0.781; in the calmest quintile, their WMAPE is 11.2% and 12.3%, respectively.Source: §V · pp. 5–6, Figure 4 · p. 6

Ablation analysis

  • Removing the negative-correlation penalty raises WMAPE from 12.4% to 13.1%, while removing drift-weighted aggregation raises it to 12.9%.The former also lowers regime macro-F1 from 0.763 to 0.741, and using one expert per client yields 13.3% WMAPE; the analysis attributes gains to interacting diversity, drift weighting, and multi-task supervision.Source: §V · p. 6

Feature-attribution review

  • Commodity index has the largest mean absolute attribution share at 0.24, 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, and are intended to support review of plausible drivers rather than represent universal supply-chain importance.Source: §V · p. 6, Figure 5 · p. 6

Client holdout and communication overhead

  • With the two most volatile clients held out from training, FEF-NCL WMAPE rises to 14.7%, while SCAFFOLD rises to 16.4%.FEF-NCL transmits 1.38 times FedProx's parameters and has 1.31 times its mean round time; the simulation describes this overhead as acceptable for weekly planning but potentially excessive for near-real-time replenishment.Source: §V · p. 6

VI. DISCUSSION

FEF-NCL’s discussion attributes its controlled forecasting gains to preserving client-specific error patterns under uneven market volatility, not to universal superiority over other models. The authors frame it as a candidate for high-value, cross-silo planning while emphasizing governance needs and substantial validation and deployment limitations.

VI. DISCUSSION

  • FEF-NCL can improve controlled forecasting under uneven client volatility by preserving specialized local error profiles and weighting updates by drift and regime coverage.This may matter when rare shocks disproportionately affect stockout, expedited-freight, or supplier-penalty costs.Source: §VI · p. 6
  • The synthetic comparison with centralized gradient boosting is not evidence of general superiority; the narrower federated-baseline comparison shows improvements in WMAPE, delay-risk macro-F1, convergence, and high-volatility robustness.The synthetic sequential and regime-shift structure favors temporal models with shock specialization, and a centralized temporal ensemble could perform differently.Source: §VI · p. 6
  • Reviewable logging and attribution can help planners examine forecast timing, contributing client updates, drift-gate activation, and influential features, but does not guarantee fairness, causality, or regulatory compliance.Such evidence supports discussion with procurement, operations, finance, and risk teams.Source: §VI · p. 6
  • The evidence is limited because results are synthetic, privacy analysis is incomplete, and real shocks may be ambiguous or absent from the modeled feature set.Gradients and model updates can still leak information without secure aggregation, differential privacy, or contractual safeguards.Source: §VI · pp. 6–7
  • FEF-NCL is a candidate for cross-silo planning when data cannot be centralized, shocks are economically material, nodes can coordinate training, and federated rounds fit decision timelines.Communication and model-management overhead, limited client compute, minute-level latency needs, and descriptively assessed explanation quality may constrain deployment; under suitable conditions, specialization can retain local regime knowledge in a shared model.Source: §VI · p. 7

VII. CONCLUSION AND FUTURE WORKS

On a single synthetic dataset, FEF-NCL outperformed the tested federated baselines, but the paper presents it as a starting point rather than evidence of general superiority. Future work centers on real-data evaluation, privacy and governance, operational measures, and deployment readiness.

VII. CONCLUSION AND FUTURE WORKS

  • FEF-NCL combines federated temporal forecasting, multi-task risk estimation, negative-correlation specialization, drift-aware aggregation, and explanation review, evaluated on one synthetic dataset.The dataset contains 124,800 weekly observations from ten regional clients, 60 product families, 40 suppliers, five commodity groups, and a 2021–2024 volatility profile; no real-world validation is claimed.Source: §VII · p. 7
  • 12.4% WMAPE beat SCAFFOLD (13.9%), FedProx (14.2%), and FedAvg (14.8%); delay-risk macro-F1 reached 0.801 and market-regime macro-F1 0.763 on the synthetic dataset.Error increased for every model in the highest-volatility quintile, but rose less sharply for the negative-correlation ensemble; ablations suggest the gains depend on expert diversity, drift-weighted aggregation, and multi-task learning together.Source: §VII · p. 7
  • Federated forecasting may benefit from explicit ensemble diversity under non-identically distributed client data and shifting regimes, while standard aggregation remains a strong, simpler baseline.Negative-correlation learning adds complexity but can preserve specialized knowledge about rare or local regimes; the framework is a starting point requiring real-data, adversarial, privacy, and planner-centered evaluation, not a claim of general superiority.Source: §VII · p. 7
  • Future work should evaluate real or semi-real data, strengthen privacy, study adaptive expert counts, add economic outcomes, and assess planner understanding and trust.Proposed work includes blinded live backtests and privacy-preserving partnerships; secure aggregation, differential privacy, and leakage testing; and measures such as service level, holding cost, expedited freight, lost sales, and supplier penalties.Source: §VII · p. 7
  • Further challenges include hierarchical forecast reconciliation, strategic behavior affecting model outputs, and production deployment tooling.Federated reconciliation must respect sensitive aggregate volumes; robust aggregation and audit trails may help address incentives, while versioning, client onboarding, drift response, and rollback need testing.Source: §VII · p. 7

Related papers

Loading 2608.21399v1…