Spain Almond Production Forecast Methodology

# 1. Executive summary

The Spain Almond Production Forecast is an estimate of the national almond crop for a given marketing year, produced by Demeter. It is constructed bottom-up from the country's almond orchards and takes into account age, weather, water availability, condition and a range of other metrics. It is calibrated against the official production record published by the Ministerio de Agricultura, Pesca y Alimentación (MAPA, the Spanish Ministry of Agriculture, Fisheries and Food). It is expressed in tonnes of in-shell almonds together with a kernel equivalent. It is published with a range that reflects the method's measured historical error.

The forecast is published as a point-in-time estimate. It reflects all the information available on the date it is issued and is updated as the season progresses. Applied retrospectively to every season since 2015, the method's average error against the MAPA figure is 8.7%, measured with the full growing season observed. Further detail on historical accuracy is provided below.

# 2. Design principles

Bottom-up construction. The forecast starts from a registry of almond orchards covering the great majority of the national area, each with its own estimated planting year and production class. Small traditional orchards that cannot be mapped individually are carried as a separate stratum, sized from the official area record. The forecast does not employ a sampling methodology, grower surveys or top-down trend extrapolation. National production is the sum of the expected production of each orchard and of that stratum.

Tonnes of crop as the goal Spain combines a large traditional rainfed sector with a fast-growing irrigated one, and irrigated orchards yield around five times as much per hectare. Average yield per hectare therefore moves with the mix of orchards as well as with the season. The forecast outputs a number of tonnes, and reports yields only as a derived quantity.

Tailored to the region. The core methodology builds on Demeter's experience forecasting production in other crops and regions, but each input is validated independently for its contribution to accuracy in Spain specifically. The level is fitted to MAPA's official production series, by province and nationally. Some inputs are unique to Spain, including a national network of ground weather stations.

Point-in-time and like-for-like. The current season is compared with historical seasons over the same calendar window. A forecast issued in March compares bloom-to-date inputs in the most recent year with bloom-to-date inputs in every prior year.

Transparent on accuracy and validation approach. Every published figure carries its range. The method's track record is disclosed in this document and elsewhere. Accuracy is measured by leaving each year out in turn, re-fitting the model without it, and scoring the prediction against the actual figure.

# 3. Conceptual framework

The forecast combines three layers. The first two feed the model, while the third is a set of independent checks applied before publication.

# 3.1 The orchard base

The national orchard footprint is compiled, and validated by Demeter. It comprises some 350,000 orchards covering approximately 600,000 hectares. Each is assigned a production class, distinguishing intensive orchards from extensive ones. These labels are derived from observable characteristics as determined by Demeter, and do not rely on the administrative classification of a parcel as irrigated or rainfed. Each orchard is also labelled with its year of planting, or a range of years if a precise determination cannot be made. In Spain these labels cover a period of more than 20 years.

Orchards that cannot be mapped individually are carried as a separate stratum. These are typically small traditional rainfed plantings. The size of the stratum in each season is taken from the difference between the official area record and the mapped footprint in the previous year. Each historical season is therefore modelled on the area that stood at the time.

The classes are each assigned a default yield curve derived from a range of inputs. The combination of these yield curves provides a baseline from which deviations are estimated based on conditions in the current and prior seasons. Spanish production moves in waves lasting several years, not in alternate years. The baseline for each season therefore also reflects the previous season's outturn relative to productive capacity.

# 3.2 Growing conditions

Growing conditions are captured through the variables that make up the Demeter Almond Season Index - Spain (DASI-S). These include temperature during bloom, heat accumulation during kernel fill, frost exposure at each growth stage and rainfall over the water year. They are measured at the level of each orchard and allow the model to incorporate the influence of weather in both the current and prior season.

Bloom dates are estimated for each location and season, so that frost is assessed against the growth stage the trees had reached. The response to each variable is calibrated separately for irrigated and rainfed orchards, against more than 20 years of weather conditions in Spain.

# 3.3 Independent checks

The output of the model is validated against a range of instruments that do not feed the model directly. These are intended as a sense-check against conditions on the ground and serve primarily to move the headline figure within its error range.

Examples include a national network of ground weather stations, an ongoing health monitor on orchards in the productive base and a number of sentiment channels. Where ground stations and sentiment channels both register a damaging frost, the headline figure is reduced according to a rule fixed in advance.

Not all independent checks are relevant at all stages in the season, and their impact is most important in "outlier" years. The list varies over time and is not disclosed in its entirety.

# 4. Forecast construction

Orchard-level forecasts are summed into provinces. Provincial totals are used for internal diagnostic purposes but are not published.

Provinces are summed to a national figure, which is the source of the headline publication.

The specific class definitions, yield curve parameters, channel weights, calibration values and full list of independent checks are proprietary to Demeter and are not disclosed.

# 5. Track record

The forecasting method has been applied retrospectively to every season since 2015. Earlier seasons predate the modern planting wave, and including them was found to reduce accuracy.

Each year's estimate is produced out of sample: the model is calibrated without that year, then asked to predict it.

MAPA in-shell figures are converted to kernel using a flat crack-out ratio of 30%. Area at bearing age is MAPA's official series to 2025. The 2026 figure is measured from Demeter's orchard base.

Mean absolute error: 29.2 thousand tonnes, or 8.7%. The two largest errors fall at turning points in the production cycle. In 2018 the crop recovered sharply after two suppressed seasons. In 2022 a severe frost in early April followed a run of large crops. The estimates for 2021 and 2022 include the frost adjustment described in section 3.3. The official figures for 2024 and 2025 remain provisional.

The error above is measured with the full growing season observed. Prints issued earlier in the season carry wider uncertainty. At the start of the crop year, for example, they rest on the orchard base and the previous season's outturn alone. Uncertainty narrows as the season progresses.

# 6. Definitions

In-shell The headline in-shell figure is published on the same basis as the official MAPA total to facilitate comparison.

Kernel The in-shell tonnage multiplied by a flat crack-out ratio of 30%. This is consistent with the kernel-basis reporting of the International Nut and Dried Fruit Council (INC). Actual crack-out varies by variety.

Bearing age Refers to orchards in their third year after planting and older.

# 7. Governance, updates and scoring

The forecast is maintained by Demeter's crop forecasting team. Its model specification is versioned; the 2026/27 marketing year is served under version v1. Coefficients are re-fitted whenever MAPA publishes a new annual figure. The specification changes only between marketing years, with advance notice and with the historical series restated for comparability.

Regular prints will be published on the forecast page as they become available. Every print carries its as-of date and model version.

# 8. Further information

For methodology enquiries, please contact Demeter at [email protected].

Information and current values for all of Demeter's forecast and index products are available on the Forecasts and Indices pages.