Scholarlink

<p>Accurate forecasting of agricultural commodity prices is essential for stabilizing market expectations and supporting risk-sensitive decisions in production planning and policy intervention. This study proposes Drift-Aware Conformal Prediction (DACP), a plug-and-play calibration framework that converts diverse forecasting backbones into interval predictors with stable reliability under temporal nonstationarity. Unlike conventional deep learning pipelines that emphasize pointwise losses and report only deterministic forecasts, DACP targets decision-ready uncertainty by constructing predictive intervals with controlled coverage through rolling regime-aligned calibration, adaptive coverage control, and shock-aware reset. The proposed framework is model agnostic and can be attached to representative backbones including CNN, LSTM, N-BEATS, Autoformer, and Informer without modifying their architectures. Experiments on weekly potato prices from 25 Chinese provinces between 2012 and 2018 show that DACP consistently improves risk reliability, recovering near nominal coverage under drift while maintaining competitive sharpness, and it yields modest but systematic gains in point forecasting stability across short and longer horizons. These results indicate that conformal calibration with drift awareness provides a principled route to trustworthy uncertainty quantification in agricultural markets, enabling forecasts that are better aligned with operational needs in market regulation and agricultural management.</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

Drift-Aware Conformal Prediction for Reliable Potato Price Forecasting under Nonstationary Markets

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Jianming Ouyang,
  • Chaowei Tang

摘要

Accurate forecasting of agricultural commodity prices is essential for stabilizing market expectations and supporting risk-sensitive decisions in production planning and policy intervention. This study proposes Drift-Aware Conformal Prediction (DACP), a plug-and-play calibration framework that converts diverse forecasting backbones into interval predictors with stable reliability under temporal nonstationarity. Unlike conventional deep learning pipelines that emphasize pointwise losses and report only deterministic forecasts, DACP targets decision-ready uncertainty by constructing predictive intervals with controlled coverage through rolling regime-aligned calibration, adaptive coverage control, and shock-aware reset. The proposed framework is model agnostic and can be attached to representative backbones including CNN, LSTM, N-BEATS, Autoformer, and Informer without modifying their architectures. Experiments on weekly potato prices from 25 Chinese provinces between 2012 and 2018 show that DACP consistently improves risk reliability, recovering near nominal coverage under drift while maintaining competitive sharpness, and it yields modest but systematic gains in point forecasting stability across short and longer horizons. These results indicate that conformal calibration with drift awareness provides a principled route to trustworthy uncertainty quantification in agricultural markets, enabling forecasts that are better aligned with operational needs in market regulation and agricultural management.