[Home](/) › Case Studies

# Case studies

Five concrete years showing how CropIntel's walk-forward ensemble performed,
bad years called correctly, a bumper year called correctly, and one miss
written up honestly. Every figure is a plotted point on the
[track-record scatter](/track-record).

### 2012, washout, 4 of 4 high-confidence regions correct

The worst UK wheat year in a generation: a relentlessly wet summer wrecked
grain fill and drove disease. The model called all 4 high-confidence regions
below-average. Predicted −0.46, actual −0.84 t/ha, directionally right and
conservatively sized.

### 2014, bumper year, 3 of 3 correct (bullish)

Not a doom call. A near-record harvest, called correctly as
**above-average** in all 3 high-confidence regions. Predicted
+0.27, actual +1.70 t/ha. The case that proves CropIntel is a directional
forecaster, not a one-sided pessimist.

### 2019, disaster year, 8 of 9 regions correct

A wet autumn drilling window, cold tillering in the east, and a hot, dry
June. UK average yield fell to 7.0 t/ha, the worst since 2012. The model
flagged 8 of 9 regions below-average ahead of the May checkpoint. Predicted
−0.73, actual −1.12 t/ha.

### 2023, wet harvest, 6 of 6 high-confidence regions correct

An unusually wet July and August, textbook compound stress at flowering
and ripening. The model called 6 of 6 high-confidence regions below-average,
intensifying the call in close to real time as the wet July developed.
Predicted −0.33, actual −0.77 t/ha.

### 2018, where we got it wrong, 2 of 6 correct

The cleanest miss in the backtest. The model read 2018's hot, dry summer as
heat stress and called mildly bearish (−0.23); UK wheat actually had a bumper
year (+1.52) because the dry weather suppressed disease. Honest writeup of why
the compound-stress model was blind to the low-disease bonus, and how the
sentiment layer addresses the gap.

Background: [Methodology](/methodology) ·
[Full track record](/track-record) ·
[Who this is for](/for/)