Blockchain Federated Learning for Sustainable Retail: Reducing Waste through Collaborative Demand Forecasting

AI summary

Overview

Research area: Federated Learning (FL) applied to Sustainable Supply Chain Management (SSCM), specifically demand forecasting for perishable goods in grocery retail, combined with blockchain and IPFS for transparency and integrity.

Technical level: Intermediate. Readers should be comfortable with basic machine learning concepts (train/test splits, loss functions, neural networks) and have some familiarity with federated learning aggregation and blockchain basics. No advanced mathematics is required.

Scope: The paper presents a preliminary empirical case study showing that a blockchain-based horizontal federated learning system can forecast retail demand and reduce perishable-product over-provisioning without retailers sharing raw data.

What This Paper Is About

Retailers of perishable goods need accurate demand forecasts to avoid over-ordering (which creates waste) and under-ordering (which creates stockouts), but individual shops often have sparse or unreliable sales data. Sharing data with competitors would improve forecasts, yet retailers are reluctant to disclose sensitive commercial information. The paper tests whether Federated Learning — where shops train a shared model collaboratively while keeping their data local — can approach the accuracy of full data sharing while preserving privacy, and adds blockchain and IPFS components so participants can verify that the training server behaved honestly.

Key Contributions

1. A three-way comparison for retail demand forecasting. The authors benchmark Standalone Learning (each store trains alone), Centralized Learning (all stores pool data on one server), and Federated Learning (stores share only model parameters) using the same neural network architecture in all three cases, plus a waste-oriented metric called over-provisioning error (OE).

2. A privacy-and-transparency architecture combining FL with blockchain. The system integrates Secure Aggregation Plus (SecAgg+) with additive masking and secret key sharing, Global Differential Privacy with gradient clipping, an Ethereum-based blockchain, and IPFS storage of model weights referenced by content identifier (CID) hashes.

3. A quantitative assessment of over-provisioning error and waste reduction. The study reports OE reductions attributable to federated learning across stores with different levels of data availability.

4. A cost comparison across six Ethereum-based blockchain platforms. The same smart contract was deployed on Ethereum, Sepolia, Optimism, Arbitrum, Polygon PoS, and Polygon zkEVM, with gas costs for smart contract deployment, transactions, and validation reported for each.

Main Findings

• Centralized beats federated, federated beats standalone. The error under Centralized data handling is lower than under FL, which in turn performs better than standalone forecasting. Exact MSE values are shown in Figure 4 but are not reported numerically in the text. FL's advantage over centralized learning is privacy preservation, described as particularly valuable when multiple companies are involved.

• FL reduces over-provisioning error by over 5% on average. On average, FL reduced OE — and consequently perishable waste — by more than 5% compared with the alternatives evaluated.

• Waste reductions exceed 40% in poorly performing stores. The reduction was most pronounced in stores where standalone forecasting performed poorly, with waste reductions exceeding 40% reported for those cases.

• FedAvg was the chosen aggregation algorithm. FedProx and FedAdam were also tested, but FedAvg proved particularly suitable given the low heterogeneity of the data.

• Differential privacy degraded performance in these experiments. The paper reports that the data was highly sensitive to noise; while local DP enhances privacy, it exponentially amplifies noise and degrades performance.

• Layer 2 blockchains are far cheaper than mainnet. Gas usage was unchanged across platforms but gas prices varied significantly. Reported overall costs include 0.070487114 ETH for Ethereum mainnet, 0.038856533 ETH for Sepolia, 1.70423E-05 ETH for Optimism, 0.00022155 ETH for Arbitrum, 1.806487934 ETH for Polygon PoS, and 0.172775233 ETH for Polygon zkEVM. Prices are based on the rate as of 30/01/2025, 17:00 CET, with 1 Gwei equal to 10^-9 ETH.

• Costs can scale substantially in more complex deployments. In deployments involving a larger number of clients — with token-based participation incentives, client reputation tracking, and potential recording of partial updates — operational costs can increase up to fivefold, with estimates ranging from approximately 8.50 × 10^-5 ETH for Optimism to 4.03 ETH for Polygon PoS.

• The system is deliberately not fully decentralized. The server remains a single point of failure. The authors state this was a deliberate choice because the setting is horizontal cross-device federated learning within entities of the same organization, so the trust level is already sufficiently high and no hierarchical server structure or expensive on-chain computation is needed.

• Context figures cited from prior literature: food production accounts for 26% of total greenhouse gas emissions, 26.7% of food is not consumed, and food waste from supply chain losses and end-consumers accounts for 6% of total emissions. The food loss index is reported as 98.3, with the 2021 food loss percentage at 13.23%, unchanged from 2016. Almost two thirds of all food waste in Europe comes from the consumer sector, including retail.

Methodology in Plain English

The researchers used a public Walmart sales dataset covering 45 stores selected for basic analysis out of 10,585 stores globally, with weekly sales from 2010 to 2012 and contextual variables including holiday flags, temperature, fuel price, Consumer Price Index (CPI), and unemployment rate.

They simulated three learning arrangements with an identical model in each: each store training only on its own history, all stores pooling their data on one central server, and stores training collaboratively through federated learning. Numerical values were standardized using z-score, and categorical variables (weekday, month, year) were extracted from the date field. Stores used either a 70-30 or a 30-70 train-test ratio to simulate varying data availability. Forecasting was done client-side with a simple neural network using lag-based feature engineering, dropout and weight decay for regularization, and Leaky ReLU activations to maintain gradient flow. More complex architectures such as LSTM or attention-based models were deliberately excluded so that the standalone, centralized, and federated versions could be compared fairly — the authors note that an oversized network could lead clients to overfit on their own data in the federated case.

On the privacy side, the architecture aggregates masked parameters at the server using SecAgg+, with gradients clipped and local training performed via FedAvg. After aggregation, differential privacy is applied to the global model, which is then saved to IPFS, where model weights are compressed into a CID hash stored on the blockchain along with information such as timestamp and number of rounds. Clients verify the received model by recomputing its CID and comparing against the value on the blockchain, raising an alert if there is an inconsistency. To measure waste impact, the authors computed over-provisioning error, which quantifies how much forecasts exceed actual demand — a direct driver of excess stock, storage costs, and waste for perishable goods.

Why This Matters

Impact on research. The paper sits at the intersection of federated learning, sustainable supply chain management, and blockchain-based verifiability. Its literature comparison table positions the work as covering grocery retail, vertical/horizontal FL, sustainability focus, and perishable focus simultaneously — a combination the authors indicate was not covered by the prior studies they surveyed, which addressed sectors such as retail, automotive, e-commerce, agri-food, closed-loop supply chains, and energy. It also documents a negative practical result: differential privacy was too noisy to be useful in this setting, which is useful evidence for others designing privacy budgets.

Real-world applications.

• Grocery and perishable retail chains forecasting weekly store-level demand without pooling commercially sensitive sales data.
• Smaller or data-poor stores that could improve their own forecasts by participating in a shared model trained across larger partners.
• Supply chain coordination scenarios where retailers want verifiable evidence that a central training server did not tamper with model updates, using blockchain-recorded CIDs.
• Blockchain platform selection, where the reported gas and ETH cost table can inform whether to deploy on mainnet, a testnet, a side-chain, or a Layer 2 solution.

Industry relevance. The paper frames the situation as a prisoner's dilemma: all parties would gain from cooperation but hesitate because of data-sharing concerns. It argues privacy-preserving federated learning combined with security mechanisms offers a viable path, with the caveat that smart contract optimization is necessary to avoid excessive gas fees. It also notes that retail is already a domain where FL has been applied and that corporate strategy in the sector has shifted toward long-term social impact alongside short-term profit, aligning with Sustainable Development Goal 12.3 on halving per capita global food waste at retail and consumer levels.

Future Directions

1. Incentive mechanisms for data-rich participants. The authors state that encouraging participation from data-rich buyers may require mechanisms such as token-based rewards.

2. Scaling beyond the current setting. Large-scale supply chains with frequent supplier changes present challenges that the current 45-store, single-retailer-style scenario does not capture.

3. Addressing broader supply chain pathologies. Future work should explore challenges such as demand amplification (the bullwhip effect) and shortage gaming.

4. Standards and regulation. Establishing standardized communication protocols and regulatory frameworks is described as crucial for widespread adoption of FL in supply networks.

5. Smart contract optimization. The paper explicitly identifies minimizing smart contract operational costs as an open challenge for scalable and cost-effective blockchain-based FL.

Target Audience

Researchers and practitioners in federated learning, privacy-preserving machine learning, and sustainable supply chain management will find the core contribution most useful. Retail operations and demand-planning analysts will benefit from the over-provisioning error framing and the reported waste reductions. Blockchain engineers evaluating deployment platforms will find the gas and ETH cost comparison table directly actionable. The paper is also a reasonable entry point for graduate students or industry newcomers wanting a concrete applied case study that connects FL mechanics, privacy mechanisms, and sustainability metrics in one system. Acknowledgment is given to the NOUS project (Grant Agreement Number 101135927) and the NextGenerationEU Ecosister project (CUP B33D21019790006, PNRR Missione 4 Componente 2 Investimento 1.5, Spoke 4).

Authors’ abstract

Effective demand forecasting is crucial for reducing food waste. However, data privacy concerns often hinder collaboration among retailers, limiting the potential for improved predictive accuracy. In this study, we explore the application of Federated Learning (FL) in Sustainable Supply Chain Management (SSCM), with a focus on the grocery retail sector dealing with perishable goods. We develop a baseline predictive model for demand forecasting and waste assessment in an isolated retailer scenario. Subsequently, we introduce a Blockchain-based FL model, trained collaboratively across multiple retailers without direct data sharing. Our preliminary results show that FL models have performance almost equivalent to the ideal setting in which parties share data with each other, and are notably superior to models built by individual parties without sharing data, cutting waste and boosting efficiency.