Scholarlink

<p>The growing demand for sustainable, pure-cotton products in offline retail necessitates accurate demand forecasting, yet the complex interplay of weather, holidays, and discounts, and their nonlinear and heterogeneous effects, remains underexplored. This study addresses this gap by quantifying multi-factor impacts and developing a two-stage framework integrating empirical analysis with machine learning to leverage nonlinear effects of external factors. We first conduct hierarchical regression using authentic sales data from a leading cotton textile company to identify statistically significant features and regional heterogeneity, which reveals that features such as the quadratic effects of temperature are highly significant, with some product categories sensitive to specific weather ranges. Guided by empirical results, a hybrid model integrates feature engineering and regional segmentation to capture heterogeneities using ensemble learning. This integration of evidence-based variable engineering allows our model to capture complex interactions, adapt to new products and regional variability, and achieve an average accuracy (ACU) of 0.72, a 73% improvement over the moving average baseline. These actionable insights contribute to sales strategy in offline retail and demonstrate the power of making interpretable predictive decisions grounded in robust empirical analysis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Empirical and Machine Learning Forecasting for Offline Retail: Nonlinear Weather Effects and Heterogeneity

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Mingyuan Zhu,
  • Chenbin Luo,
  • Xiaoqiang Cai,
  • Shan Dai,
  • Weili Xue,
  • Lianmin Zhang

摘要

The growing demand for sustainable, pure-cotton products in offline retail necessitates accurate demand forecasting, yet the complex interplay of weather, holidays, and discounts, and their nonlinear and heterogeneous effects, remains underexplored. This study addresses this gap by quantifying multi-factor impacts and developing a two-stage framework integrating empirical analysis with machine learning to leverage nonlinear effects of external factors. We first conduct hierarchical regression using authentic sales data from a leading cotton textile company to identify statistically significant features and regional heterogeneity, which reveals that features such as the quadratic effects of temperature are highly significant, with some product categories sensitive to specific weather ranges. Guided by empirical results, a hybrid model integrates feature engineering and regional segmentation to capture heterogeneities using ensemble learning. This integration of evidence-based variable engineering allows our model to capture complex interactions, adapt to new products and regional variability, and achieve an average accuracy (ACU) of 0.72, a 73% improvement over the moving average baseline. These actionable insights contribute to sales strategy in offline retail and demonstrate the power of making interpretable predictive decisions grounded in robust empirical analysis.