Weather Forecasts vs Seasonal Forecasts
A standard weather forecast — the kind used to decide whether to carry an umbrella — works by solving the equations of atmospheric physics forward in time from today's observed conditions. It tells you, with reasonable confidence, what the atmosphere will do over the next few days. Beyond about ten days, the chaotic nature of the atmosphere means this type of deterministic forecast loses its predictive value almost entirely.
A seasonal forecast operates on an entirely different premise. Rather than predicting specific weather events, it asks: given current ocean temperatures, soil moisture conditions, and large-scale climate drivers, is the coming season likely to be warmer, wetter, or drier than normal? The answer is expressed not as a single value but as a probability distribution — a statement about which outcomes are more or less likely relative to the historical record.
This distinction matters enormously for how you read and apply the output. A seasonal forecast that says "there is a 60% chance of above-normal precipitation over the next three months" does not mean it will definitely rain more than usual. It means the balance of evidence — from ocean heat content, atmospheric circulation patterns, and climate model physics — tilts in that direction. The remaining 40% represents genuine uncertainty that no modelling system can eliminate.
The right mental model: Think of a seasonal forecast the way a doctor thinks about risk. A patient with elevated markers has a higher probability of a particular outcome — but probability is not destiny. The forecast shifts your prior expectations; it does not replace them with certainty.
ECMWF and the SEAS5 System
The European Centre for Medium-Range Weather Forecasts — ECMWF — operates what is widely regarded as the world's leading numerical weather and climate prediction system. Based in Reading, UK, ECMWF produces forecasts used by meteorological agencies, governments, and commercial organisations globally.
Its seasonal forecasting product is called SEAS5 — the fifth-generation Seasonal Forecasting System. SEAS5 initialises a new forecast on the first of each month, coupling a high-resolution atmospheric model with an ocean model and sea-ice component. The ocean coupling is critical: sea surface temperatures evolve slowly over months and are the primary source of predictability at seasonal timescales. When large regions of ocean are anomalously warm or cool — as during El Niño or La Niña events — they systematically shift atmospheric circulation patterns in ways that are partially predictable weeks and months in advance.
SEAS5 runs 51 separate ensemble members simultaneously. Each member starts from slightly perturbed initial conditions, representing the real-world uncertainty in our knowledge of the current atmospheric state. By running the model 51 times and examining the spread of outcomes, SEAS5 produces a probability distribution rather than a single forecast trajectory.
The Ensemble and What It Tells You
The 51 ensemble members are the core of everything. Each member is a plausible realisation of how the coming months could evolve given today's climate state. When the ensemble members cluster tightly together, the model is expressing high confidence — most plausible futures look similar. When members spread widely, the model is expressing genuine uncertainty — the atmosphere could evolve in materially different ways.
For any given location and variable, you can ask: of the 51 members, how many project above-normal conditions? How many project below-normal? The ratio is the forecast probability. If 38 of 51 members project above-normal rainfall, the probability is 38/51 — roughly 75%. If members split almost evenly, the model has little useful signal beyond climatology.
Tercile Probabilities and Anomaly Detection
The standard framework for communicating seasonal forecast skill is the tercile probability. The historical climate record for a given location, month, and variable is divided into three equal thirds — the lower tercile (the driest or coldest 33% of historical years), the middle tercile (the middle 33%), and the upper tercile (the wettest or warmest 33%). These boundaries are fixed by the historical record and represent what "normal" variability looks like.
A climatological baseline — one with no predictive signal — would assign 33% probability to each tercile. A useful seasonal forecast departs from this equal split. When SEAS5 projects a 55% probability of above-normal precipitation, it is saying that the current climate state makes that outcome meaningfully more likely than historical chance alone would suggest. The degree of departure from 33/33/33 is a direct measure of forecast confidence.
Each bar shows the fraction of ensemble members falling in that tercile. Climatological baseline = 33% each.
Scenario C above — 58% probability of below-normal precipitation — is the kind of signal most relevant to coffee supply monitoring. For Arabica, below-normal rainfall during the critical January to March flowering and fruit-set window in Brazil can meaningfully suppress the following year's yield. A SEAS5 signal of this magnitude, detected in November or December, gives procurement desks and trading operations several months of lead time to reassess supply expectations before the agronomic impact becomes visible in crop surveys.
Skill and Lead Time
Not all seasonal forecasts are equally reliable. Forecast skill — the degree to which the model's probability estimates correspond to observed outcomes over many forecasts — degrades with lead time. Month 1 forecasts (the month immediately following initialisation) consistently show higher skill than month 6 or 7 forecasts, simply because the atmosphere has had less time to evolve away from the initial conditions that drive predictability.
Skill also varies by region, variable, and season. Precipitation is generally harder to forecast at seasonal lead times than temperature, because rainfall depends on mesoscale processes that large-scale ocean drivers influence only indirectly. Regions strongly coupled to ENSO — such as northeastern Brazil, parts of Colombia, and East Africa — tend to show higher seasonal forecast skill than regions where local factors dominate.
How We Apply This to Coffee
On the Data in Geospace platform, SEAS5 ensemble output is processed at fixed farm-level coordinates across each producing country. For each ensemble member, each variable, and each forecast month, we classify the projected conditions against the same S1/S2/S3/N agroclimatic thresholds used in our historical ERA5 products. The fraction of members falling into each suitability class at each farm point becomes the probability shown in the Probability Map product.
The tercile framework translates directly into our anomaly products: a farm point where the majority of ensemble members project above-normal suitability shows as green; below-normal shows as red. This gives the spatial probability distribution a concrete agronomic interpretation — not just "wetter than normal" but "conditions that historically support Optimal Arabica growing."
Lead time awareness is built into how we present forecast data. Month 1 and month 2 forecasts carry the most actionable signal and should be weighted heavily in short-run procurement and logistics decisions. Month 5 to 7 forecasts are useful for directional awareness and scenario planning, but should be treated as indicative rather than definitive — the ensemble spread at that range is wide, and any single trajectory within it remains plausible.
The value of imperfect foresight: Even a seasonal forecast with modest skill is commercially valuable. If a forecast correctly tilts probabilities toward below-normal rainfall in a key origin 60% of the time when that outcome eventually occurs, it provides a systematic information advantage over waiting for the anomaly to appear in ground-level crop surveys weeks or months later. The goal is not certainty — it is better-calibrated expectations, earlier.