Structural predictability differences in long-term regional crop yield modeling with climate and area signals

Area of Science:

• Agricultural Science
• Climate Science
• Data Science

Background:

• Crop yield modeling is crucial for food security.
• The role of climate signals versus temporal persistence in yield prediction needs further investigation.
• Understanding these factors is key for accurate agricultural forecasting.

Purpose of the Study:

• To assess the marginal predictive value of climate signals compared to temporal persistence in annual regional crop yield modeling.
• To investigate crop-specific differences in long-term yield predictability.
• To evaluate the contribution of climate and planting area data to forecasting models.

Main Methods:

• Utilized anonymized multi-region panel data from 2000-2023.
• Developed a pooled one-step-ahead machine learning forecasting framework.
• Compared machine learning model performance against a naïve persistence baseline.

Main Results:

• Cotton yield prediction improved significantly when planting area was included, reaching an R² of 0.72, similar to the persistence baseline (0.71).
• Wheat yield predictability was dominated by persistence (R² = 0.99), driven by stable regional differences rather than strong within-region temporal patterns.
• Climate signals alone offered limited predictive skill compared to persistence in the annual, regional setting.

Conclusions:

• Annual regional crop yield predictability is highly crop-dependent and influenced by the stability of the production system.
• Temporal persistence can explain a substantial portion of predictable yield variation in stable agricultural systems.
• Future agricultural forecasting research must incorporate persistence-aware evaluation and robust benchmark designs.