Scholarlink

<p>Price volatility remains a key challenge for the citrus industry in Corrientes, Argentina, where producers operate under uncertain climatic and economic conditions. This study investigates whether incorporating climatic variables can improve the prediction of citrus prices, thereby enhancing risk management and supporting data-driven decision-making. Long Short-Term Memory (LSTM) neural networks were employed to forecast annual citrus prices for the 1970–2023 period, using climatic, productive, and macroeconomic variables. Two model configurations were compared: one including climatic variables (temperature, precipitation, humidity, and solar radiation) and another one excluding them. Sequential Split Cross-Validation (SSCV) was applied for model validation, and performance was assessed using RMSE, MAE, and NRMSE metrics. The model trained with data from the Monte Caseros department—excluding climatic variables—achieved the best predictive performance (RMSE = 30.34 ± 38.58, MAE = 21.18 ± 25.87, NRMSE = 0.082), reflecting high accuracy relative to observed price variability. In contrast, models including climatic variables exhibited higher normalized errors (NRMSE ≈ 0.12–0.15), suggesting that short-term climatic fluctuations introduced additional noise not effectively captured by the model. Statistical tests (t-test, ANOVA) confirmed that these differences were not significant (<i>p</i> &gt; 0.05). The findings indicate that, at the current temporal and spatial data resolution, macroeconomic and market factors have a greater impact than climatic variables on citrus price dynamics. The proposed LSTM-based framework offers a replicable approach for forecasting perennial crop prices and contributes to the development of intelligent, climate-aware decision-support systems in agriculture.</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

Evaluating LSTM deep learning models for agricultural price forecasting: a case study on citrus markets in Corrientes, Argentina

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Claudia R. Screpnik,
  • Eduardo Zamudio,
  • Laura I. Gimenez

摘要

Price volatility remains a key challenge for the citrus industry in Corrientes, Argentina, where producers operate under uncertain climatic and economic conditions. This study investigates whether incorporating climatic variables can improve the prediction of citrus prices, thereby enhancing risk management and supporting data-driven decision-making. Long Short-Term Memory (LSTM) neural networks were employed to forecast annual citrus prices for the 1970–2023 period, using climatic, productive, and macroeconomic variables. Two model configurations were compared: one including climatic variables (temperature, precipitation, humidity, and solar radiation) and another one excluding them. Sequential Split Cross-Validation (SSCV) was applied for model validation, and performance was assessed using RMSE, MAE, and NRMSE metrics. The model trained with data from the Monte Caseros department—excluding climatic variables—achieved the best predictive performance (RMSE = 30.34 ± 38.58, MAE = 21.18 ± 25.87, NRMSE = 0.082), reflecting high accuracy relative to observed price variability. In contrast, models including climatic variables exhibited higher normalized errors (NRMSE ≈ 0.12–0.15), suggesting that short-term climatic fluctuations introduced additional noise not effectively captured by the model. Statistical tests (t-test, ANOVA) confirmed that these differences were not significant (p > 0.05). The findings indicate that, at the current temporal and spatial data resolution, macroeconomic and market factors have a greater impact than climatic variables on citrus price dynamics. The proposed LSTM-based framework offers a replicable approach for forecasting perennial crop prices and contributes to the development of intelligent, climate-aware decision-support systems in agriculture.