Scholarlink

<p>Potato prices are shaped by local production cycles, interregional market co-movement, and recurrent seasonal fluctuations, making reliable forecasting difficult when only historical price records are available. This paper develops MarketFM, a price-only self-supervised foundation model for regional potato price forecasting. Instead of training a supervised predictor directly from limited labelled windows, MarketFM first learns a general representation of agricultural market dynamics from historical provincial price panels. The proposed framework tokenizes regional price histories into temporal patches, encodes both within-province evolution and cross-province co-movement through a factorised market encoder, and pretrains the encoder using masked price modeling, market-view contrastive learning, and frequency-temporal alignment. These self-supervised objectives encourage the model to recover missing price segments, learn perturbation-invariant market states, and preserve oscillatory price structures without relying on weather, logistics, inventory, policy, or other external variables. Forecasts are obtained by adapting the pretrained encoder with a lightweight horizon-aware decoder. Experiments on weekly potato prices from 25 Chinese provinces during 2012 to 2018 show that MarketFM achieves lower MAE, RMSE, and MAPE than CNN, LSTM, N-BEATS, Autoformer, and Informer on both one-step and four-step horizons. Component analysis further indicates that removing self-supervised pretraining, masked price modeling, market-view contrastive learning, or frequency-temporal alignment consistently weakens performance. The results suggest that price-only self-supervised market representation learning is an effective and practical strategy for agricultural price forecasting under limited data 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

MarketFM: A Self-Supervised Foundation Model for Regional Potato Price Forecasting

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Shengjie Gong,
  • Yanfei Peng

摘要

Potato prices are shaped by local production cycles, interregional market co-movement, and recurrent seasonal fluctuations, making reliable forecasting difficult when only historical price records are available. This paper develops MarketFM, a price-only self-supervised foundation model for regional potato price forecasting. Instead of training a supervised predictor directly from limited labelled windows, MarketFM first learns a general representation of agricultural market dynamics from historical provincial price panels. The proposed framework tokenizes regional price histories into temporal patches, encodes both within-province evolution and cross-province co-movement through a factorised market encoder, and pretrains the encoder using masked price modeling, market-view contrastive learning, and frequency-temporal alignment. These self-supervised objectives encourage the model to recover missing price segments, learn perturbation-invariant market states, and preserve oscillatory price structures without relying on weather, logistics, inventory, policy, or other external variables. Forecasts are obtained by adapting the pretrained encoder with a lightweight horizon-aware decoder. Experiments on weekly potato prices from 25 Chinese provinces during 2012 to 2018 show that MarketFM achieves lower MAE, RMSE, and MAPE than CNN, LSTM, N-BEATS, Autoformer, and Informer on both one-step and four-step horizons. Component analysis further indicates that removing self-supervised pretraining, masked price modeling, market-view contrastive learning, or frequency-temporal alignment consistently weakens performance. The results suggest that price-only self-supervised market representation learning is an effective and practical strategy for agricultural price forecasting under limited data conditions.