Scholarlink

<p>Agricultural supply-chain managers need early warning of unusual price movements, yet deterministic monthly forecasts remain difficult when commodity, input, fuel, and freight regimes shift. This study evaluates an explainable machine-learning workflow using 18 monthly U.S. Bureau of Labor Statistics Producer Price Index series. The supervised sample contains 189 forecast origins, 164 engineered predictors, 141 training months, and a 48-month chronological holdout. Model families and hyperparameters are selected exclusively by five-fold time-series cross-validation; the holdout is used once for evaluation. Elastic net regression is preselected for signed one-month-ahead Farm Products PPI change and achieves holdout MAE of 2.00%, RMSE of 2.72%, R<sup>2</sup> of 0.143, and directional accuracy of 58.33%. Its squared-loss improvement over the zero-change baseline is suggestive but not significant at 5% (Harvey-Leybourne-Newbold-adjusted Diebold-Mariano <i>p</i> = 0.062). Extra trees regression is preselected for absolute-change forecasting and yields high-volatility screening ROC-AUC of 0.652, recall of 85.71%, precision of 25.00%, and F1 of 0.387. The exact 95% recall interval is 42.13%-99.64%, and the moving-block bootstrap ROC-AUC interval is 0.375–0.881, demonstrating substantial uncertainty. The evidence indicates limited signed-forecast skill and a high-recall but low-precision screening pattern in this holdout sample. The workflow should therefore be interpreted as a preliminary human-review risk screen, not as autonomous price prediction or a deployment-grade probability model.</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

Explainable machine learning for agricultural supply chain price volatility risk screening

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月19日

查看全文
  • Raisul Islam Khan,
  • Taycir Ben Abid,
  • Mohammad Abu Jahed

摘要

Agricultural supply-chain managers need early warning of unusual price movements, yet deterministic monthly forecasts remain difficult when commodity, input, fuel, and freight regimes shift. This study evaluates an explainable machine-learning workflow using 18 monthly U.S. Bureau of Labor Statistics Producer Price Index series. The supervised sample contains 189 forecast origins, 164 engineered predictors, 141 training months, and a 48-month chronological holdout. Model families and hyperparameters are selected exclusively by five-fold time-series cross-validation; the holdout is used once for evaluation. Elastic net regression is preselected for signed one-month-ahead Farm Products PPI change and achieves holdout MAE of 2.00%, RMSE of 2.72%, R2 of 0.143, and directional accuracy of 58.33%. Its squared-loss improvement over the zero-change baseline is suggestive but not significant at 5% (Harvey-Leybourne-Newbold-adjusted Diebold-Mariano p = 0.062). Extra trees regression is preselected for absolute-change forecasting and yields high-volatility screening ROC-AUC of 0.652, recall of 85.71%, precision of 25.00%, and F1 of 0.387. The exact 95% recall interval is 42.13%-99.64%, and the moving-block bootstrap ROC-AUC interval is 0.375–0.881, demonstrating substantial uncertainty. The evidence indicates limited signed-forecast skill and a high-recall but low-precision screening pattern in this holdout sample. The workflow should therefore be interpreted as a preliminary human-review risk screen, not as autonomous price prediction or a deployment-grade probability model.