Researchers at NYU Abu Dhabi’s Mubadala Arabian Center for Climate and Environmental Sciences (ACCESS) have developed a simple data- driven algorithm called the Random Analogue Predictor (RAP) that can forecast Arctic sea ice extent up to nine months ahead.

Published around October 8, 2026 (with the peer- reviewed paper in Scientific Reports), the work is led by Francesco Paparella (inaugural director of Mubadala ACCESS and senior author) and collaborators including Faiq Raees.

How RAP works

Unlike complex physics- based climate models that simulate interactions among the atmosphere, ocean, and sea ice, RAP relies solely on the historical record of Arctic sea ice extent. It identifies past periods with conditions similar to the present, then examines how those analogous situations evolved to generate an ensemble of possible future forecasts. The spread of the ensemble provides an uncertainty estimate for each prediction.

Performance

In tests, RAP achieved skill comparable to models used by the Sea Ice Prediction Network (SIPN). For September sea ice extent (a key seasonal target), its forecast error matched that of 34 seasonal forecasting models. Reported root- mean- square error (RMSE) stayed no larger than about $0.6\times {10}^6$ km² overall, with lower values (under $0.4\times {10}^6$km²) for winter and spring forecasts even at long lead times.

Significance

Arctic sea ice reflects solar energy and influences atmospheric and oceanic patterns beyond the polar region, so improved seasonal forecasts help anticipate broader climate impacts. The researchers position RAP as a transparent, interpretable, low-cost benchmark for both traditional physics- based models and emerging AI approaches: if a more complex system cannot beat this simple analogue method, it raises questions about the added value of that complexity. Paparella has also noted potential usefulness for the UAE’s polar research activities.

In short, RAP shows that a deliberately straightforward historical- pattern approach can compete with sophisticated models while quantifying its own uncertainty, offering a useful baseline for future Arctic sea-ice prediction research.

_____________________________________________________________________________________

The value of model complexity is precisely the central question raised by the RAP paper.

RAP’s core argument

Raees and Paparella deliberately built an extremely simple, transparent method that:

• Uses only the historical Arctic sea-ice extent (SIE) time series (no atmosphere, no ocean, no spatial fields, no physics equations, no trainable parameters).
• Still achieves RMSE ≤ 0.6 × 10⁶ km² and skill comparable to the 34 models in the Sea Ice Prediction Network for September SIE forecasts.

They then make a strong methodological claim:

If a much more complex model (whether physics-based or AI- based) cannot outperform such a simple approach, that tells us something important about how much additional predictive information that complexity is actually providing.

In other words, complexity is only valuable to the extent that it demonstrably improves forecast skill beyond what a naïve but well- designed baseline can already achieve.

Failure to beat RAP is presented as a serious shortcoming for any model that claims to capture realistic sea- ice dynamics.

Why this matters

1. Complexity is costly
   Full dynamical models require enormous computational resources, careful tuning of many physical parameterizations, and high- quality initial conditions. Large AI models require massive training data, compute, and often lack interpretability. RAP needs none of these.
2. Added complexity can be illusory
   The paper notes that even sophisticated dynamical and machine- learning systems frequently show year- to- year errors that fluctuate widely (sometimes > 1 × 10⁶ km²). Matching or beating a simple analogue method is therefore a meaningful test of whether the extra machinery is extracting genuine additional predictability or merely fitting noise and historical correlations already captured by delay embedding.
3. A useful “null model” for progress
   Just as persistence was once the standard benchmark that forced numerical weather prediction to prove its worth, RAP is proposed as the modern seasonal- scale equivalent. Beating it is not a major achievement; failing to beat it is a red flag.
4. Interpretability and uncertainty
   Because RAP is stochastic and produces ensembles, it naturally supplies its own uncertainty estimate. Many complex models struggle to provide well- calibrated uncertainty, so simplicity here also brings practical advantages.

Broader takeaway

Model complexity has value only when it delivers measurable, reliable improvement over simpler alternatives on the specific task that matters (here, seasonal pan- Arctic SIE). RAP forces the community to quantify that improvement rather than assume that more physics or more parameters automatically equal better forecasts.

In short: the paper does not claim that complex models are useless.

It claims that their complexity must earn its keep by clearly outperforming a transparent, physics- free, data- only baseline. That is a stringent and useful standard.

_____________________________________________________________________________________

Random analog prediction provides a benchmark for seasonal Arctic sea ice extent forecasting

Abstract (key points)

The authors introduce a Random Analogue Predictor (RAP) algorithm for seasonal forecasting of Arctic sea ice extent (SIE). It is a stochastic variant of the classical method of analogues applied to a scalar time series. RAP uses only the historical SIE record (from NSIDC) to generate ensemble forecasts.

When the observations are compared with the most representative member of the ensemble (selected via band- depth, a centrality measure for functional data), the method shows:

• Negligible bias
• Root- mean- square error (RMSE) no larger than $0.6\times {10}^6$0.6×106 km²

Hindcasts of the September average SIE achieve skill comparable to the models participating in the Sea Ice Prediction Network (SIPN).

The authors argue that RAP’s simplicity, interpretability, independence from physical hypotheses or parameterizations, and built-in uncertainty estimates make it a necessary baseline: any physics- based or AI- based model that fails to outperform RAP should be viewed as having a serious shortcoming if it claims to realistically describe sea- ice dynamics.

Core idea and method

• RAP is deliberately minimalistic. It relies on delay- coordinate embedding of the single scalar SIE time series (plus a date- of- year component) to reconstruct an effective state space.
• Analogues are chosen stochastically, weighted by distance in the embedding space.
• Because of the random selection, repeated runs produce an ensemble whose spread quantifies uncertainty.
• No spatial fields, no atmospheric and oceanic variables, no trainable parameters, and no physical assumptions about ice dynamics are used.

Context and motivation

Seasonal Arctic SIE forecasting sits between short-range weather prediction (“first- kind” forecasts that try to match the actual trajectory) and long- term climate projections (“second- kind” forecasts that match statistics). RAP is intentionally a second- kind method that still retains useful skill for quantities such as the September minimum.

The authors note that even sophisticated dynamical and machine- learning models often show RMSE values around 0.45 × 10⁶ km² (or higher and more variable) for June- initialized September forecasts, performance that RAP matches with far less complexity. They position RAP as the modern analogue of the old “persistence” benchmark that once helped drive progress in numerical weather prediction.

Authors & affiliations

• Faiq Raees: Courant Institute of Mathematical Sciences (NYU), Division of Science & Center for Stability, Instability, and Turbulence (NYU Abu Dhabi)
• Francesco Paparella (corresponding author): Division of Science & Mubadala Arabian Center for Climate and Environmental Sciences (ACCESS), NYU Abu Dhabi

Both authors contributed equally. Funding came from Tamkeen awards to SITE (CG0002) and ACCESS (CG0009).

Availability

Code and data related to the method are available (earlier versions appeared on Zenodo and as a Research Square preprint). The paper itself is open access.

In short, the work presents a transparent, ultra- simple, data-only ensemble forecasting method that performs competitively with far more elaborate systems and is proposed as a required benchmark for future Arctic sea- ice prediction research.

Journal: Scientific Reports (2026)

DOI: 10.1038/s41598-026-72959-0

Provided: New York University

Published online: 5 October 2026 (received 15 Oct 2025; accepted 21 Sep 2026)

Authors: Faiq Raees &
Francesco Paparella

Abstract

We develop a random analog predictor (RAP) algorithm for the forecasting of Arctic sea ice extent (SIE) on seasonal timescales. This is a stochastic variant for scalar time series of the method of the analogues that only uses the historical SIE record to produce ensemble forecasts. When comparing the observations with the most representative forecast of the ensemble (as identified through the band–depth, a centrality measure for functional data) the algorithm shows negligible bias and RMSE no larger than km. Hindcasts of the September average SIE show a level of skill comparable to that of the models of the Sea Ice Prediction Network. We argue that simplicity, interpretability, independence on physical hypotheses, and the ability to attach an uncertainty estimate to its own forecasts, should make RAP a necessary baseline for physics–based and AI–based models alike: failure to outperform RAP should be considered as a serious shortcoming for models aiming at realistically describing sea ice dynamics.