Scholarlink

<p>In this study, we propose an ensemble hybrid model called seasonal trend decomposition based on Loess and long short-term memory (STL-LSTM) for forecasting non-stationary, nonlinear, and seasonal agricultural price series. The model integrates STL, a decomposition technique, with a recurrent neural network-based forecasting method, long short-term memory (LSTM). The STL technique decomposes the original price series into its seasonal, trend, and remainder components, while the LSTM network forecasts each component individually. The predictions from all components are then aggregated to produce the final ensemble forecast. The developed model effectively captures the temporal patterns of complex time series through the analysis of the simplified decomposed components. Empirical results from monthly potato price data of two major markets in India reveal the superior forecasting ability of the proposed model. The STL-LSTM achieves up to 12% lower root mean square error (RMSE) and 9% lower mean absolute percentage error (MAPE) compared to benchmark models which includes both individual and hybrid ones. Additionally, the Diebold-Mariano (DM) test confirms the statistical significance of the model's improved performance. The results highlight the STL-LSTM hybrid model as a promising tool for agricultural price forecasting and provide a foundation for further advancements in this domain.</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

STL-LSTM Hybrid Model for Forecasting Seasonal Agricultural Price Series

  • 期刊论文

  • 发表日期：

    2025年3月20日

查看全文
  • Ronit Jaiswal,
  • Girish K. Jha,
  • Rajeev Ranjan Kumar,
  • Kapil Choudhary

摘要

In this study, we propose an ensemble hybrid model called seasonal trend decomposition based on Loess and long short-term memory (STL-LSTM) for forecasting non-stationary, nonlinear, and seasonal agricultural price series. The model integrates STL, a decomposition technique, with a recurrent neural network-based forecasting method, long short-term memory (LSTM). The STL technique decomposes the original price series into its seasonal, trend, and remainder components, while the LSTM network forecasts each component individually. The predictions from all components are then aggregated to produce the final ensemble forecast. The developed model effectively captures the temporal patterns of complex time series through the analysis of the simplified decomposed components. Empirical results from monthly potato price data of two major markets in India reveal the superior forecasting ability of the proposed model. The STL-LSTM achieves up to 12% lower root mean square error (RMSE) and 9% lower mean absolute percentage error (MAPE) compared to benchmark models which includes both individual and hybrid ones. Additionally, the Diebold-Mariano (DM) test confirms the statistical significance of the model's improved performance. The results highlight the STL-LSTM hybrid model as a promising tool for agricultural price forecasting and provide a foundation for further advancements in this domain.