Season Risk Scores & Aggregation
This article explains how individual hazard alerts are turned into the Season Risk Scores you see on the dashboard - and how scores are rolled up from individual locations to the overall region level.
At a glance
The Season Risk Score answers one question: how risky is this season's weather stress compared to history? A score of 80 means the current season is more stressed than 80% of the past 30 years, at the same point in the crop cycle.
Scores are updated daily. They reflect the full accumulation of hazard alerts from the start of the season to the current date.
Step 1 - From alerts to a hazard risk score
The first step is making alert counts interpretable. The number of drought alerts at a given location this season only becomes meaningful when you compare it to the distribution of drought alert counts across 30 years of historical data at the same location.
The platform does this using a percentile ranking approach: for each hazard at each location, the current season's weighted alert total (Critical = 2, Warning = 1) is ranked against the historical distribution of weighted totals for the same hazard, location, and day of season. The result is a Hazard-Location Risk Score between 0 and 100. A score of 70 means the current alert count is higher than 70% of historical seasons to this point in the year.
This percentile framing has several advantages. It accounts for regional climate differences automatically - a score of 70 for drought in Egypt means something very different in absolute terms from a score of 70 in Poland, but both mean the same thing relative to local history. It also makes risk comparable across crops and regions.
Step 2 - From hazard scores to location scores
Each monitored location has four hazard scores - one for heat stress, one for frost, one for drought, and one for excessive rainfall. These are combined into a single Location Risk Score.
The combination uses a statistical method called a Gumbel copula - a tool widely used in finance and insurance for aggregating correlated risks. The key motivation for this approach is that weather hazards do not occur independently: a season with both severe drought and intense heat simultaneously is not just twice as bad as a season with one of those stressors - it is qualitatively more anomalous, because the combination itself is rare. A Gumbel copula is designed to capture exactly this pattern. It is sensitive to co-occurring extremes, meaning situations where multiple stressors are elevated at the same time receive higher risk scores than a simple average of the individual scores would produce.
In practice, the copula computes a joint exceedance score - a single number reflecting how risky the current combination of hazard percentiles is, given the historical relationship between stressors at this location. This score is then ranked against the same metric computed for all historical seasons, yielding the location's overall percentile risk score.
The dependence parameter used in the copula (theta) is estimated from the location's own historical data, so it captures the actual statistical relationships between hazards for that specific crop and location. Across the crops and regions currently monitored, theta typically comes out around 1.2 - reflecting mild positive correlation between stressors, consistent with the fact that drought and heat tend to co-occur.
Step 3 - From location scores to the region score
The overall Season Risk Score shown at the top of the dashboard reflects conditions across all monitored locations in the selected region. This is calculated in two steps. First, each location's copula score is ranked against that location's own historical distribution of copula scores - producing a normalised percentile rank on a 0 to 1 scale for every location. Second, these ranks are averaged across all locations to produce the region-level score.
Normalising each location's score against its own history before averaging is important: it means a location with highly variable weather history and one with stable history contribute equally to the region score, rather than the more volatile location dominating.
A note on interpretation
The Season Risk Score at region level tells you how risky the overall combination of weather stress is across all monitored locations - relative to what has been seen historically. It is not a direct measure of expected yield loss. A score of 90 means the combination of stressors is more extreme than 90% of historical seasons; it does not directly translate to a 90% probability of production disruption. The link between weather stress and actual supply outcomes depends on many factors - crop variety, irrigation, agronomic management, and market dynamics - that the platform does not model. The score is a signal that warrants attention, not a forecast of a specific outcome.
One further nuance at the region level: the regional score is the average of the individual location-level percentile ranks, not itself a direct historical percentile of the region as a unit. A score of 80 means monitored locations are, on average, experiencing conditions that rank in the 80th percentile of their respective local histories.
Risk Trend
The Risk Trend is calculated by comparing the current Season Risk Score to the score from a selectable number of days ago (default: seven days). A positive change (worsening trend) means the season is accumulating more stress relative to history than it was at the start of the window. A negative change (improving trend) means conditions have eased. The trend is shown at both region and location level.