Scholarlink

<p>Potato prices evolve through intertwined local production cycles, interregional circulation, and recurrent seasonal demand, making their future movements difficult to infer from isolated historical curves. This paper develops the Regime-Aware Causal Spectral Graph Learning Network (RCSGNet), a forecasting model that treats provincial potato markets as a changing economic network rather than a collection of independent sequences. The proposed framework learns latent market states from recent price contexts, estimates directed cross-province predictive influence patterns under each state, and propagates frequency-decomposed temporal features over the resulting graph. By coupling directional graph inference with spectral filtering, RCSGNet is designed to distinguish persistent seasonal movement from short-lived market disturbance while preserving the transmission structure among regions. The learned graph is interpreted as state-dependent directional predictability among regional price series rather than as formal causal identification from an intervention design. Forecasts are obtained through a state-gated decoder that fuses local temporal evidence, graph propagated information, and horizon embeddings. Experiments on weekly potato prices from 25 Chinese provinces during 2012 to 2018 show that RCSGNet obtains lower MAE, RMSE, and MAPE than CNN, LSTM, N-BEATS, Autoformer, and Informer on both one-step and four-step horizons. Component analysis indicates that removing the market state encoder, the directed graph learner, or spectral propagation consistently weakens performance. The results suggest that regional market connectivity and frequency structure are complementary signals for agricultural price forecasting under nonstationary conditions.</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

Regime-Aware Causal Spectral Graph Learning for Regional Potato Price Forecasting

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Changqian Bai,
  • Youpeng Wang

摘要

Potato prices evolve through intertwined local production cycles, interregional circulation, and recurrent seasonal demand, making their future movements difficult to infer from isolated historical curves. This paper develops the Regime-Aware Causal Spectral Graph Learning Network (RCSGNet), a forecasting model that treats provincial potato markets as a changing economic network rather than a collection of independent sequences. The proposed framework learns latent market states from recent price contexts, estimates directed cross-province predictive influence patterns under each state, and propagates frequency-decomposed temporal features over the resulting graph. By coupling directional graph inference with spectral filtering, RCSGNet is designed to distinguish persistent seasonal movement from short-lived market disturbance while preserving the transmission structure among regions. The learned graph is interpreted as state-dependent directional predictability among regional price series rather than as formal causal identification from an intervention design. Forecasts are obtained through a state-gated decoder that fuses local temporal evidence, graph propagated information, and horizon embeddings. Experiments on weekly potato prices from 25 Chinese provinces during 2012 to 2018 show that RCSGNet obtains lower MAE, RMSE, and MAPE than CNN, LSTM, N-BEATS, Autoformer, and Informer on both one-step and four-step horizons. Component analysis indicates that removing the market state encoder, the directed graph learner, or spectral propagation consistently weakens performance. The results suggest that regional market connectivity and frequency structure are complementary signals for agricultural price forecasting under nonstationary conditions.