Federated Demand Forecasting for Retail Supply Chains

Retailers, distributors and store operators all forecast demand for the same products, but almost none of them will hand their sales history to one another. Federated learning offers a way around that stalemate: each party trains on its own data and shares only model updates. The question we care about is a practical one. Does federation actually make store-level forecasts better, and does it still help once the updates are protected with differential privacy? We study this on the public M5 dataset (30,490 SKU-store series from ten Walmart stores in three US states, 1,969 days), treating every store as a separate data silo. We compare local-only training, a pooled centralized model, FedAvg, FedAvg with local fine-tuning, and a DP-FedSGD variant that guarantees item-level (ε, δ)-differential privacy for each store's catalogue, with privacy computed exactly using Gaussian differential privacy. The answer depends mostly on how much history a store has. With two years of training windows, federation is close to a wash at the SKU level (FedAvg with fine-tuning scores 0.895 WRMSSE against 0.899 for local training) and its 4.2% reduction in store-total error is within seed noise. With only two weeks of windows, local models fall apart and federation recovers much of the loss (0.939 vs 0.971), yet a plain 28-day moving average (0.922) is still hard to beat. When one newly opened store joins nine mature ones, fine-tuning the federated model lowers that store's error by 7.5% on average over local training, and by 7.1% with ε = 2 privacy. We also report where the approach breaks: near-zero sales on Christmas Day, erratic SKUs, and private training when stores drop out of rounds. The full set of experiments runs in under half an hour on two CPU cores, and every number in this paper comes from our own runs.

Authors

Publication Details

Journal
International Journal of Science and Research (IJSR)
Published
2026-11-19
DOI
https://doi.org/10.21275/sr261111062849
Primary Topic
Privacy-Preserving Technologies in Data
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Federated Demand Forecasting for Retail Supply Chains

Rahul Khurana, Jatin Singh
International Journal of Science and Research (IJSR)
Privacy-Preserving Technologies in Data
article

Federated Demand Forecasting for Retail Supply Chains

Rahul Khurana, Jatin Singh
article en

Abstract

Retailers, distributors and store operators all forecast demand for the same products, but almost none of them will hand their sales history to one another. Federated learning offers a way around that stalemate: each party trains on its own data and shares only model updates. The question we care about is a practical one. Does federation actually make store-level forecasts better, and does it still help once the updates are protected with differential privacy? We study this on the public M5 dataset (30,490 SKU-store series from ten Walmart stores in three US states, 1,969 days), treating every store as a separate data silo. We compare local-only training, a pooled centralized model, FedAvg, FedAvg with local fine-tuning, and a DP-FedSGD variant that guarantees item-level (ε, δ)-differential privacy for each store's catalogue, with privacy computed exactly using Gaussian differential privacy. The answer depends mostly on how much history a store has. With two years of training windows, federation is close to a wash at the SKU level (FedAvg with fine-tuning scores 0.895 WRMSSE against 0.899 for local training) and its 4.2% reduction in store-total error is within seed noise. With only two weeks of windows, local models fall apart and federation recovers much of the loss (0.939 vs 0.971), yet a plain 28-day moving average (0.922) is still hard to beat. When one newly opened store joins nine mature ones, fine-tuning the federated model lowers that store's error by 7.5% on average over local training, and by 7.1% with ε = 2 privacy. We also report where the approach breaks: near-zero sales on Christmas Day, erratic SKUs, and private training when stores drop out of rounds. The full set of experiments runs in under half an hour on two CPU cores, and every number in this paper comes from our own runs.

International Journal of Science and Research (IJSR)
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon