CropIntel by the numbers

Name: CropIntel headline statistics
Creator: Alastair Campbell

Headline statistics for diligence, citation, and quick reference. Each figure is grounded in the public track record and methodology.

62.7%
Tradeable hit rate, walk-forward 2003-2025: of the 59 medium/high-conviction directional calls the model made, 37 were correct.
45.9%
Direction-correct across all 253 region-years. The gap to 62.7% is the point: the consensus filter only calls when components agree, turning a coin-flip into an edge.
92%
Bad-year detection: of the 25 below-average region-years where the model made a conviction call, it correctly called 23 below average, the strategically valuable subset.
71%
High-confidence call accuracy (12 of 17).
p<0.0001
Significance of predicted-vs-actual yield anomaly across the 231 scored region-years, pooled (r=0.307). The region-years span 21 independent seasons; the year-by-year view is on the track record.
11
DEFRA regions modelled independently (Eastern, East Midlands, Yorkshire & The Humber, etc.).
253
Region-years in the walk-forward backtest population, 2003-2025 (231 with confirmed outcomes).
4 layers
Independent data sources feeding the ensemble: weather, soil, satellite, sentiment.
2 layers
Independent scoring methods per practitioner observation, a hand-curated UK-agri lexicon and a frontier LLM, blended toward the LLM.
Daily
Refresh cadence, every morning at 07:00 UTC, full pipeline finishes by 07:25 UTC.
8 of 9
DEFRA regions called correctly in 2019, the canonical UK wheat disaster year.
5 of 5
DEFRA regions called correctly in 2023, a wet-harvest year.

How to cite

CropIntel statistics may be cited freely for research, journalism, and diligence purposes. Suggested citation:

CropIntel (2026). UK wheat yield forecasting, walk-forward backtest 2003-2025. Retrieved from https://cropintel.net/statistics