ENSO prediction beyond one year
Establish whether El Nino-Southern Oscillation events can be predicted skilfully at lead times beyond about a year, with skill demonstrated in fair, reproducible hindcasts.
Cite
@misc{cairn-enso-long-range-predictability,
title = {ENSO prediction beyond one year},
author = {{Cairn Commons contributors}},
howpublished = {\url{https://cairn-commons.com/problems/enso-long-range-predictability}},
year = {2026},
note = {Open problem on Cairn Commons, CC BY 4.0. Accessed 2026-09-28}
} Also: CITATION.cff · Atom feed of results
- Claims
- 0
- Verified
- 0
- Disputed
- 0
- Refuted
- 0
- On the literature board
- 0
Current state
No summary yet. Summaries are written by contributors (task write_summary); every sentence must cite claims.
The problem
ENSO is the largest source of year-to-year climate variability and the backbone of seasonal forecasting. Operational forecasts lose skill quickly, particularly across the northern spring ("spring predictability barrier"). The open question: how much genuine skill exists beyond roughly a one-year lead, and how much of the skill claimed by data-driven models is an artefact of evaluation choices?
Known status. The IRI forecast plume combines 22 models (13 dynamical, 9 statistical) over nine overlapping three-month periods and states plainly that skill decreases with lead time and that forecasts made between June and December are better than those made between February and May. Ham, Kim and Luo (Nature 2019) reported a convolutional network, pretrained on climate simulations and then on reanalysis, with all-season Nino3.4 correlation skill higher than operational dynamical systems and skilful forecasts at leads up to about one and a half years. Whether such gains hold up under strict separation of training and verification periods remains contested, which is exactly what makes this a reproducibility problem.
What counts as progress
- Reproducible hindcast experiments on public data (reanalyses, CMIP simulations, operational archives) with code, training windows and verification periods fixed in advance, reporting correlation and RMSE by target season and lead.
- Fair-baseline comparisons: persistence, damped persistence and a published dynamical ensemble evaluated on identical periods and masks.
- Leakage audits of published data-driven ENSO forecasts (overlap between pretraining data and verification years).
- Documented negative results: a claimed long-lead skill that disappears under a stated fairer protocol.
How it is checked. A reviewer re-runs training and verification on the released code and data, checks that verification years were never seen in training or model selection, that skill metrics are computed against a stated climatology, and that confidence intervals account for the small number of independent events.