Foundation Models for Demand Forecasting via Dual-Strategy Ensembling Conference: Proceedings of the 1st Workshop on "AI for Supply Chain: Today and Future" @ 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 3, 2025; Toronto, ON, Canada. Proceedings of the 1st Workshop o

Wei Yang, Defu Cao, Yan Liu · 2025

Read Foundation Models for Demand Forecasting via Dual-Strategy Ensembling Conference: Proceedings of the 1st Workshop on "AI for Supply Chain: Today and Future" @ 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 3, 2025; Toronto, ON, Canada. Proceedings of the 1st Workshop o by Wei Yang, Defu Cao, Yan Liu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-22053.txt.

Introduction

Wei Yang, Defu Cao, and Yan Liu wrote a 2025 paper on forecasting product demand with foundation models. Shops and warehouses must guess future sales to stock well. The task is hard because sales form layers, from store to category, and tastes keep shifting.

The team proposes a frame that joins two ensemble plans. One splits training and use by those layers, so each level learns its own pattern. The second plan adds strength against shifts in the data. Big pre-trained models give the base forecasts. Blending them aims for steadier accuracy in real supply chains. The paper targets sales forecasts that hold up as conditions change.

Catalog record

CREATORWei Yang, Defu Cao, Yan Liu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-22053

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Foundation Models for Demand Forecasting via Dual-Strategy Ensembling Conference: Proceedings of the 1st Workshop on "AI for Supply Chain: Today and Future" @ 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 3, 2025; Toronto, ON, Canada. Proceedings of the 1st Workshop on "AI for Supply Chain: Today and Future" @ 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’25), August 3, 2025, Toronto, ON, Canada ISBN: 979-8-4007-1454-2/25/08 DOI: 10.1145/XXXXXX.XXXXXX CCS: Computing methodologies Time series analysis CCS: Computing methodologies Ensemble methods CCS: Computing methodologies Transfer learning CCS: Applied computing Forecasting CCS: Applied computing Supply chain management Wei Yang, Defu Cao, Yan Liu Abstract . Accurate demand forecasting is critical for supply chain optimization, yet remains difficult in practice due to hierarchical complexity, domain shifts, and evolving external factors. While recent foundation models offer strong potential for time series forecasting, they often suffer from architectural rigidity and limited robustness under distributional change. In this paper, we propose a unified ensemble framework that enhances the performance of foundation models for sales forecasting in real-world supply chains. Our method combines two complementary strategies: (1) Hierarchical Ensemble (HE) , which partitions training and inference by semantic levels (e.g., store, category, department) to capture localized patterns; and (2) Architectural Ensemble (AE) , which integrates predictions from diverse model backbones to mitigate bias and improve stability. We conduct extensive experiments on the M5 benchmark and three external sales datasets, covering both in-domain and zero-shot forecasting. Results show that our approach consistently outperforms strong baselines, improves accuracy across hierarchical levels, and provides a simple yet effective mechanism for boosting generalization in complex forecasting environments. 1. Introduction Accurately forecasting future sales is a fundamental task in modern supply chain management ( Ghalehkhondabi et al. 2017 ; Nia et al. 2021 ; Hasan et al. 2025 ; Mitra et al. 2022 ) . It drives core decisions in procurement, inventory planning, production scheduling, and logistics. Yet, despite its importance, demand forecasting remains highly challenging in practice ( Liu et al. 2024b ; Swaminathan and Venkitasubramony 2024 ) . Real-world supply chains are complex, hierarchical, and increasingly sensitive to disruptions such as economic shocks, pandemics, and geopolitical tensions ( Trapero et al. 2024 ; Kagalwala et al. 2025 ; Zhang et al. 2024 ; Verma 2024 ) . These disruptions frequently induce regime shifts in demand signals, rendering many forecasting models brittle or unreliable. This situation creates a critical tension. While accurate forecasts are essential for operational resilience, building models that are robust, adaptive, and generalizable across diverse scenarios remains an open research problem ( Khlie et al. 2024 ; Jahin et al. 2024 ; Oyewole et al. 2024 ) . Sales forecasting plays a central role in supply chain optimization ( Theodorou et al. 2022 ; Lainder and Wolfinger 2022 ; Kumar et al. 2024 ) . It requires accurate forecasts across product, store and regional levels while accounting for seasonality, promotions and external disruptions. The M5 competition has emerged as a key benchmark in this domain, stimulating diverse modeling strategies ( Hewamalage et al. 2021 ; Seaman and Bowman 2022 ; Kolassa 2022 ; Makridakis et al. 2022 ) . Top solutions fall into three paradigms: tree-based models like LightGBM ( Ke et al. 2017 ) , which rely on rich features and hierarchical ensembling; neural approaches such as DeepAR ( Salinas et al. 2020 ) , which leverage autoregressive structures and distribution-aware objectives; and hybrid or statistical methods that remain effective under sparse or intermittent demand. Recently, foundation models like Chronos ( Ansari et al. 2024 ) and TEMPO ( Cao et al. 2023a ) have gained traction by pretraining on large-scale time series to enable zero-shot generalization. While promising in bypassing task-specific tuning, these models often underperform when faced with domain shifts or hierarchical imbalances ( Gong et al. [n.d.] ; Ma et al. 2024 ; Jin et al. 2024 ) . This highlights the need for new methods that can amplify their strengths while mitigating structural brittleness, particularly in complex, real-world supply chain environments. Despite progress in time series modeling, single-model forecasters remain limited by the trade-off between bias and variance ( Liu and Zhou 2024 ; Muñoz-Zavala et al. 2024 ) . Models based on trees, neural networks or foundation architectures often rely on fixed assumptions that do not adapt well across forecasting conditions. In supply chains, these limitations are amplified. Demand varies widely across products, stores and time periods, and often shifts due to promotions, seasonality or external events ( Hasan et al. 2025 ; Mitra et al. 2022 ; Khlie et al. 2024 ) . Even foundation models trained on diverse data can struggle in zero-shot settings when the target domain differs from the pretraining distribution. These failures lead to blind spots across segments and unstable predictions across different hierarchy levels. To address these issues, prior work has explored combining models. Mixture-of-experts methods are one example ( Shi et al. 2024 ; Vallarino 2024 ; Alkilane et al. 2024 ) . They aim to select or weight models dynamically, but often depend on specialized gating functions or fixed architecture designs. As a result, they can be sensitive to noise and fail to generalize under structural shifts. In real-world supply chains, where both data distribution and domain structure vary, such methods often lack robustness ( Li et al. 2024 ; Oyewole et al. 2024 ) . This highlights the need for more flexible ensemble strategies that can integrate diverse model perspectives while remaining stable across different forecasting scenarios. To overcome the limitations of single-model forecasting in complex supply chain environments, we develop an ensemble framework that combines hierarchical structure awareness with architectural diversity. This design is guided by two key insights. First, supply chain data exhibits inherent hierarchies across stores, categories, and regions. Modeling each semantic level separately allows better alignment with localized demand patterns and improves generalization across structural variations. Second, different forecasting architectures—such as tree-based models, recurrent networks, and transformers distinct inductive biases and capture complementary aspects of temporal dynamics. By integrating these perspectives, our approach incorporates two coordinated strategies. Hierarchical Ensemble (HE) partitions the training and inference process across semantic levels to promote subgroup specialization. Architectural Ensemble (AE) combines predictions from heterogeneous backbones to mitigate model-specific variance and improve robustness. Together, these components enable foundation models to produce more accurate and stable forecasts under both in-domain and zero-shot conditions. Our contributions are summarized as follows: • We propose a unified ensemble framework that enhances the performance of foundation models for supply chain forecasting. The framework boosts performance by strategically orchestrating how models are used and their predictions combined, thereby leveraging their collective strengths without requiring modifications to their underlying architecture. • We introduce two complementary strategies: Hierarchical Ensemble (HE), which models group-specific patterns through semantic partitioning of the data, and Architectural Ensemble (AE), which integrates multiple forecasting backbones to improve robustness under distribution shifts. • We conduct extensive experiments on the M5 benchmark and external sales datasets, demonstrating that our approach consistently outperforms strong baselines in both in-domain and zero-shot forecasting scenarios. 2. Related Work 2.1. Sales Forecasting Methods Sales forecasting has long served as a core task in retail supply chains (

Read the complete text