How far ahead can compound droughts, whiplash and hot summers be seen?
The compounding story is about sequences: a failed rainy season after a failed one, a drought that breaks in a flood, dry soil that carries a bad winter into a bad summer. For each of the 15 brief regions we asked how much of that is predictable, how far ahead, and from what: the state of ENSO, the landscape's memory (soil water, the last months' rain, snow) or the dynamical model (ECMWF SEAS5, every start month since 1981).
Every number is a Brier skill score (BSS) against the climatological odds, from forecasts made without the year being forecast. Skill is counted only when its 90% interval excludes zero; a confirmed horizon also survives false-discovery control across all the tests on this page. The samples are small: 36 hindcast years (66 for the ENSO history), often 5 to 15 events.
The predictability matrix
How far ahead each compound outcome is predictable, region by region
Each cell is one target in one region. The number is the horizon: the longest lead, in months before the target window opens, at which a forecast beat the climatological odds with a 90% interval above zero. The letter names the source that reaches furthest. Hover or tap a cell for the detail; the curves below follow the cell you pick.
Skill against lead, by source
Which source carries the skill
Cells each source makes skilful
Reliability: do the probabilities mean what they say?
Forecasts at leads 0–2 months from every region and target, binned by probability; the dot is how often the event happened. On the diagonal is perfectly reliable. Pairs are not independent (neighbouring leads share years).
Ten models instead of one
The C3S multi-system check: September starts, 1993–2016
Is a compound event more predictable than its parts?
Two failed seasons in a row, and whiplash: joint odds against the product of the parts
The 2017–2025 check
Forecasts for years none of the models saw
Every model here was fitted only on years whose window ended before 2017, then asked about 2017–2025 (SEAS5: its real forecasts, 30 members). Brier skill pooled over the 15 regions; n is the number of forecasts, and the regions share the same few ENSO years, so this is a check, not a second verdict.
The ENSO sequence itself: El Niño, then La Niña
How often La Niña followed
Could it be seen coming?
Brier skill for "La Niña within 12 months of the El Niño peak", El Niño winters only. History: the ONI at the issue date and its 3-month change, leave-one-event-out. SEAS5: the share of members reaching La Niña before the deadline, within the seven months each start reaches.