What is the difference between yield anomaly and absolute yield?

Yield anomaly is the deviation from a year's expected trend yield, in tonnes per hectare. CropIntel forecasts anomaly because trend changes, varietal improvement, agronomic practices, over decades make absolute-yield prediction fragile.

What does 'walk-forward' mean?

Each historical year's prediction uses a model trained only on data available before that year. No in-sample fit, no look-ahead. The 62.3% hit rate is walk-forward across 253 region-years.

What's the difference between milling wheat and feed wheat?

Milling wheat (e.g. Group 1/2 varieties) goes to bread and food, commands a £10-30/t premium over feed wheat. Feed wheat goes to livestock and ethanol. Wet harvests push milling-spec into the feed market via low Hagberg / specific weight.

Why are T1, T2 and T3 spray timings significant?

They mark the rough crop calendar stage, T1 in late April (leaf-3 emerged), T2 mid-May (flag leaf, the most economically important spray), T3 ear emergence. Mentions in farmer chatter signal disease and progress.

Glossary

UK arable terminology, model methodology terms, and the specifics behind every claim CropIntel publishes. Anchor-linked for direct sharing (e.g. /glossary#compound-stress).

Compound stress score: A multi-stage aggregation across the wheat growth cycle, summing how badly each stage's weather deviated from the 1999-present baseline. Positive = worse than baseline. Used as the headline yield-risk feature in CropIntel's ensemble model. Aggregating across stages distinguishes one-off bad weeks (often recoverable) from compound years where multiple stages all deteriorated (rarely recoverable).
Walk-forward backtest: A model validation method where each historical year's prediction uses a model trained only on data available before that year, no in-sample fit, no look-ahead. CropIntel's 62.3% hit rate is walk-forward validated across 253 region-years (1999-2025). Walk-forward is the gold standard for time-series forecasting because it mirrors how the model would behave in real-time deployment.
Yield anomaly: Difference between a year's actual yield and the trend-line yield expected for that year, expressed in tonnes per hectare. Positive anomaly = above trend; negative = below trend. CropIntel predicts yield anomaly rather than absolute yield, because trend changes (variety improvements, agronomic practices) over decades make absolute-yield prediction fragile.
Ensemble model: A multi-component model with consensus filtering. Several complementary components, each with a different inductive bias, vote on the directional call, and CropIntel emits a tradeable call only when they agree. Different inductive biases make consensus more meaningful than any single model in isolation.
DEFRA region: One of the 11 official Department for Environment, Food and Rural Affairs regions used to report UK arable statistics: East Midlands, Eastern, North East, North West & Merseyside, Northern Ireland, Scotland, South East & London, South West, Wales, West Midlands, Yorkshire & The Humber. CropIntel reports per-region calls and uses regional weather and yield data.
UK-agri lexicon: CropIntel's hand-curated dictionary of UK arable language, organised across the major signal categories (disease, weather, progress, market mood, etc.). Each term carries a polarity. The lexicon scores forum posts deterministically; a frontier LLM scores them again with context. The lexicon alone reproduces a substantial portion of the final blended score, providing a deterministic floor and explainability for diligence.
Score normalisation: A small-sample-stable normalisation used in CropIntel's lexicon scorer. Returns a value in (-1, +1) that doesn't saturate easily and behaves sensibly when only a few signals are present, a single strong post can't swing the score on a quiet day.
Out-of-season filter: Posts with no agronomic-lexicon terms at all are excluded from CropIntel's daily sentiment aggregate rather than counted as zero. Without this filter, a day with 50 political tweets and 5 farming posts would dilute the farming signal; with it, only the 5 in-season posts contribute. The filter is a key reason the daily aggregate stays informative.
Compound flowering stress: Compound stress concentrated specifically at the flowering stage of UK wheat (typically late May to mid-June). One of the strongest historical predictors of below-average yield, because grain set is largely determined during flowering and is irrecoverable later. CropIntel's 2023 wet-harvest call fired heavily on flowering-stage rainfall anomaly.
Disease pressure: Aggregate level of fungal and viral disease threat across UK wheat: septoria, yellow rust, brown rust, fusarium, BYDV, eyespot, take-all. Different pathogens favour different conditions, septoria likes wet leaves; rust likes warm humidity; BYDV needs aphid vectors. UK wheat yield outcomes are sensitive to disease pressure independently of weather, which is why CropIntel's sentiment layer (which captures real-time farmer reports of disease) is methodologically necessary.
Low-disease bonus: The yield uplift UK wheat experiences in years where disease pressure is unusually low, typically because dry weather suppresses pathogen lifecycles. 2018 was the canonical low-disease year, UK wheat had a bumper harvest despite hot, dry conditions that the compound-stress model classified as adverse. See 2018 case study.
Septoria: The most economically significant fungal disease of UK wheat. Caused primarily by Zymoseptoria tritici. Spreads via rain-splash, so wet weather amplifies it. Major target of T1 (April) and T2 (May) fungicide applications. A wet, mild spring drives heavy septoria pressure; a dry summer cuts it sharply.
Yellow rust: Puccinia striiformis, a fungal disease of wheat that produces yellow-orange pustules in stripes along leaves. Favours cool, humid conditions. Variety-resistance breakdowns mean newly-vulnerable varieties can spread rapidly in a single season. Significant UK pressure in 2017 and 2024.
Brown rust: Puccinia triticina, a fungal disease of wheat that produces orange-brown pustules randomly across leaves. Prefers warmer conditions than yellow rust. Less consistently damaging in the UK but can flare in late springs.
Fusarium head blight: A fungal disease that infects wheat ears, particularly during wet flowering periods. Reduces yield and produces mycotoxins (DON, ZON) that fail food-grade quality standards even when the wheat looks acceptable. 2023 was a high-fusarium year due to the wet July.
BYDV (Barley Yellow Dwarf Virus): An aphid-vectored virus affecting wheat, barley, and oats. Causes stunted growth and yellowing. UK BYDV pressure depends on autumn aphid activity; mild autumns produce more BYDV. The end of neonicotinoid seed treatments (Cruiser SB withdrawal, 2019) increased UK BYDV exposure for arable growers.
T1 / T2 / T3 spray timings: UK wheat fungicide application stages. T1 = leaf-3 fully emerged (typically late April), targeting septoria and rust on early leaves. T2 = flag leaf fully emerged (typically mid-May), the most economically important spray. T3 = ear emergence (typically late May to early June), targeting fusarium and final-stage disease. Mentions of T1/T2/T3 in farmer chatter signal the rough crop calendar stage.
Drilling: UK term for sowing seed (the equivalent of US 'planting'). Winter wheat is typically drilled September through November; spring wheat in February to early April. Drilling conditions (soil moisture, seedbed quality, weather windows) are a major determinant of crop establishment and ultimately yield.
Tillering: The stage where a wheat plant produces additional shoots from its base. Typically October through February in the UK. Cold tillering periods produce fewer tillers and so fewer ears at harvest, suppressing yield potential.
Stem extension: Wheat growth stage from late February to mid-May, when the main stem rapidly elongates. Spring nitrogen applications time to this stage. Cold or droughty stem extension caps yield potential before flowering.
Flowering (anthesis): The stage where wheat sets grain, typically late May to mid-June in the UK. Heat or drought stress at flowering reduces the number of viable grains per ear, a yield component that is largely irrecoverable. The single most important growth stage for CropIntel's compound-stress score.
Ripening (grain fill): Wheat growth stage from flowering through harvest, when grains accumulate dry matter. Wet ripening periods reduce yield (poor grain fill) AND quality (low Hagberg, low specific weight, mycotoxin risk). The 2023 case study turned on ripening-stage rainfall.
Hagberg falling number: A milling-quality measure for wheat, indicating alpha-amylase activity. Low Hagberg (caused by pre-harvest sprouting in wet years) means the grain can't be sold for milling and gets discounted to feed prices. Quality losses on top of yield losses compounded the 2023 wet harvest impact.
Specific weight: Density of wheat grain in kg/hl. Below ~72 kg/hl typically falls out of milling spec. Wet ripening produces shrivelled grain and low specific weight.
Milling vs feed wheat: Two main UK wheat markets. Milling wheat (e.g. Group 1 / Group 2 varieties) goes to bread and food. Feed wheat goes to livestock and ethanol. Milling commands a premium of typically £10-30/t over feed wheat, but only if quality specs (protein, Hagberg, specific weight) are met. Wet harvests push milling-spec wheat into the feed market.
ICE London Feed Wheat futures: The benchmark UK wheat futures contract, traded on the Intercontinental Exchange (ICE). Standard contract: 100 tonnes, deliverable from East Anglian ports. The most liquid expressions of UK wheat price, used by merchants and processors for hedging. CropIntel's directional calls are most actionable via this contract.
MATIF (Euronext Paris milling wheat): The European benchmark milling wheat futures contract, traded on Euronext Paris. Strongly correlated with UK feed wheat prices but with a quality premium. UK arable markets price relative to MATIF for international arbitrage.
AHDB (Agriculture and Horticulture Development Board): UK levy-funded body publishing weekly arable market data (Corn Returns prices), monthly crop condition surveys, and seasonal Outlook reports. CropIntel uses AHDB regional ex-farm prices in its time-series database. AHDB's editorial reports are slower-cadence than CropIntel's daily refresh.
DEFRA yield survey: Annual UK arable yield estimate published by DEFRA (England, with separate publications for Scotland, Wales, NI). Final yields are released approximately a year in arrears; early-bird estimates land in late August / September. Used as ground-truth in CropIntel's backtest.
Area anomaly: Year-on-year change in UK wheat planted area, often a leading indicator of yield outcome. Wet autumns force area down (drilling abandoned) and historically correlate with below-average per-hectare yield in the surviving area. The 2019 disaster year saw both.
NDVI (Normalized Difference Vegetation Index): A satellite-derived measure of vegetation greenness, derived from near-infrared and red reflectance. Higher NDVI = healthier canopy. CropIntel uses Sentinel-2 NDVI per region, but the signal is currently inconclusive without crop-type masking, the SCL mask excludes forest and water but cannot distinguish wheat from grass.
SCL (Scene Classification Layer): A per-pixel classification on Sentinel-2 data tagging cloud, water, vegetation, urban, etc. CropIntel uses SCL to mask non-cropland from NDVI calculations. Provides better signal than unmasked NDVI but still doesn't isolate wheat specifically: that requires a crop-type classification layer (UKCEH Land Cover Plus: Crops, pending licensing).
ERA5 reanalysis: European Centre for Medium-range Weather Forecasts (ECMWF) reanalysis dataset, providing global weather data 1940-present at ~30km resolution. CropIntel uses ERA5 daily temperature, rainfall, sunshine, frost-day counts via Open-Meteo's archive API. ERA5 has a ~5-day lag for finalised data.
COSMOS-UK: UK soil moisture network data, 2013-present, providing daily volumetric water content across UK sites. CropIntel uses it as a contextual variable; the pre-2013 gap means it isn't a direct model feature.
Forward track record: Calls the model has made about future harvests, separate from the historical backtest. Lives in CropIntel's signals table and is published openly on the Track Record page. Currently in flight for the 2026/27 crop year. Outcomes populate as harvests confirm. The forward track record is the most diligence-relevant artefact for buyers.
Tradeable call: A CropIntel call where the model emits medium- or high-confidence direction on a region, i.e. a call a desk would actually take. The 62.3% hit rate refers specifically to tradeable calls, not the full 253 region-year population. Years with no tradeable calls are years the model said 'no high-conviction view' for every region, a feature, not a missed run.
Consensus filter: CropIntel's ensemble emits a directional call only when its complementary modelling components all agree on direction. When they disagree, no tradeable call is emitted. The consensus filter is why the tradeable subset is smaller than the full population, and why hit rates within it are higher.
Sentiment overlay multiplier: A bounded post-hoc confidence multiplier applied to CropIntel's headline call, never wide enough to flip the directional call. Edges up when sentiment direction matches the ensemble call (confirms), edges down when sentiment contradicts. It does NOT change the directional call, it adjusts the displayed confidence label only.
Keyword spike: An anomaly detection on CropIntel's sentiment layer: when mentions of a category (disease, weather, progress, etc.) in the recent rolling window exceed the prior baseline by a meaningful margin, the dashboard surfaces an early-warning panel. Used to catch sentiment shifts before they show up in the 7-day average.
Autumn drilling stress: Compound stress at the drilling stage (Sep-Nov), typically driven by wet conditions preventing seedbed preparation and forcing drilling delays. Strong leading indicator: wet autumns historically correlate with below-average yield even before any later weather events. The 2019/20 and 2024/25 wheat years both had elevated autumn drilling stress.
Frost day: A day where the minimum temperature is below 0°C. CropIntel counts frost days per region per stage, with frost-during-tillering being the strongest yield-suppressing variant. The 2018 'Beast from the East' produced extreme frost days in March 2018 but did not damage UK wheat yields, illustrating that frost timing matters more than frost count.
Sentinel-2: The European Space Agency's pair of optical Earth-observation satellites, providing 10m-resolution multi-spectral imagery on a 5-day revisit cycle. CropIntel uses Sentinel-2 NDVI per region, sourced via Copernicus.
The Farming Forum (TFF): A UK-focused agriculture web forum at thefarmingforum.co.uk. CropIntel ingests posts from the active arable sub-forums each day as part of the sentiment signal. Author handles are anonymised on the public site.
Curated practitioner signal: An actively curated UK arable practitioner signal mapped to DEFRA regions. The list is the project's primary defensible IP, building a comparable curation takes sustained UK arable activity and is not extractable from the public site.
Acreage anomaly: Difference between a year's planted wheat area and the trend-line area expected for that year. Often a leading indicator: wet autumns force area down (drilling abandoned) and historically correlate with below-average per-hectare yield in surviving area. Sourced from DEFRA's Early Bird Survey (annual).
AHDB Early Bird Survey: Annual UK arable area-intentions survey published by AHDB in late autumn, indicating planted area vs prior year by crop. CropIntel ingests Early Bird as a structured feature alongside the model's compound-stress feature.

Related: Methodology · Track Record · Case studies