📄 earth_science

Using decadal forecasts of modes of variability to refine the first few years of climate projections

This paper introduces a method to seamlessly integrate decadal forecasts with long-term climate projections by weighting them based on near-term evolution after removing forced signals, thereby improving the accuracy of global temperature and specific regional climate predictions for the first few years while providing guidance on appropriate lead times for decision-makers.

Original authors: David Sexton, Jason Lowe, Carol McSweeney

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: David Sexton, Jason Lowe, Carol McSweeney

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Climate planners face a difficult choice when looking ahead. For decisions made decades from now, they rely on long-term projections that simulate how the planet will respond to human activity, such as burning fossil fuels. These simulations are essential for building resilient infrastructure, but they treat the future as a range of possibilities without accounting for where the climate system stands right now. For decisions made in the immediate future, such as planning for next winter's floods, forecasters use initialized models that start with the current state of the ocean and atmosphere. These short-term forecasts are often more accurate because they capture the natural rhythms of the climate, but they lose their predictive power after a few years. The challenge has been to bridge these two worlds: to create a single, coherent picture that uses the precision of short-term forecasts to sharpen the long-term projections, without letting the short-term data distort the big picture.

A team of researchers at the Met Office in the UK has developed a new method to solve this problem. They created a "merged product" that combines long-term climate projections with short-term decadal forecasts. Instead of simply picking the best short-term forecast and discarding the rest, their approach weighs the long-term projections based on how well their natural variations match the patterns seen in the short-term forecasts. This allows the long-term data to benefit from the current state of the climate system, but only for as long as that state remains predictable. The result is a seamless transition from the near future, where the climate's current mood matters, to the distant future, where the long-term trend takes over.

The researchers tested their method on four specific climate indicators: the temperature of the summer subpolar North Atlantic, the annual global mean temperature, the El Niño-Southern Oscillation (ENSO) in the Pacific, and the winter North Atlantic Oscillation. For the summer North Atlantic temperatures, the merged product showed a clear advantage. In the first year after a forecast starts, the new method captured the observed temperature changes almost as well as the short-term forecasts themselves. This skill gradually faded over time, but the method remained useful for up to five years, a significant improvement over the unadjusted long-term projections. The researchers found that for global temperatures, the benefit of this merging was most visible in the very first year, helping to correct the magnitude of the warming trend and capturing short-term cooling periods that the long-term models missed.

For the El Niño phenomenon, which drives weather patterns across the globe, the method worked well for the first three years. The short-term forecasts are highly skilled at predicting El Niño events because they start with the actual temperature of the Pacific Ocean. The merged product successfully used this early skill to refine the long-term projections, keeping the predictions accurate for a few years before the natural variability of the climate system made the forecasts too uncertain to be useful. However, the method faced challenges with the winter North Atlantic Oscillation, a pattern of atmospheric pressure that affects European weather. In this case, the underlying climate models struggled to accurately simulate the natural variability of the system. Because the models were already struggling to get the basics right, the merging process could not fully fix the errors, and the final product remained less reliable than the researchers hoped.

A key feature of this new approach is that it does not force the long-term projections to follow the short-term forecasts forever. The researchers designed the system so that the influence of the current climate state naturally fades as the forecast lead time increases. This prevents the long-term projections from becoming overly constrained by a single snapshot of the present, which could lead to overconfident and potentially wrong conclusions about the distant future. Instead, the method acts like a filter that is strongest at the beginning and gradually relaxes, allowing the long-term trends to re-emerge as the primary driver of the prediction. This ensures that the merged product remains useful for decision-makers who need to plan for both the next few years and the next few decades.

The study also highlighted the importance of knowing when to trust which tool. For variables like global temperature and El Niño, the merged product offers a reliable guide for the first two to three years, after which planners should rely more on the long-term projections. For the North Atlantic summer temperatures, the window of added value extends to about five years. However, for the winter North Atlantic Oscillation, the researchers caution that the current models are not yet good enough to support this kind of merging, and using the method there could be misleading. The work suggests that while the technology to blend these different types of climate information is ready, its success depends heavily on the quality of the underlying models. As climate models continue to improve, this method could become a standard tool for providing a clear, continuous view of our changing climate, from the next winter to the next century.

Technical Summary: Using Decadal Forecasts of Modes of Variability to Refine the First Few Years of Climate Projections

Problem Statement
Climate adaptation planners face a challenge in integrating two distinct sources of climate information: long-term projections and initialised decadal forecasts. Long-term projections (e.g., CMIP6) capture the forced response to anthropogenic and natural forcings but do not assimilate the current state of the climate system, missing near-term predictability derived from internal variability. Conversely, initialised decadal forecasts offer greater near-term accuracy by constraining internal variability based on observed oceanic and atmospheric states. However, these two sources differ in complexity, forcing scenarios, and decision-making paradigms (precision-based vs. resilience-based). Existing methods for merging them, such as "shadowing" (selecting a fixed subset of projections that match a forecast), often suffer from issues like over-constraining projections beyond the time scale of predictability, under-dispersion due to small sample sizes, and the potential for double-counting information if projections have already been weighted by historical trends.

Methodology
The authors propose a "merged product" that weights long-term projections based on their near-term evolution relative to initialised decadal forecasts, without directly using the forecast values as the final output. The methodology involves the following steps:

Conceptual Model: The approach treats the projection data as a sum of a true signal (long and short time scales) and internal variability. The forecast data is decomposed into drift, initialisation shocks, a forced signal (linearly related to uninitialised runs), a residual predictable signal (contribution from initialisation), and internal variability.
Signal Removal: To ensure independence from previously applied constraints (such as trend weighting), the method removes the forced signal from both the forecasts and the long-term projections.
- For projections, the forced signal is estimated using a 32-year low-pass filter on ensemble means (or single realisations) to isolate long-term anthropogenic signals, followed by the removal of short-term forced signals (e.g., volcanic, solar) by averaging deviations.
- For forecasts, the forced signal is estimated by regressing out the average of a corresponding uninitialised projection ensemble (or the most correlated uninitialised ensemble) from the drift-corrected forecast anomalies.
Likelihood Weighting: The method calculates likelihood weights for each projection realisation at specific lead times ( $T+1$ to $T+9$ ). These weights are derived from a multivariate Gaussian distribution comparing the residual variations in the projections against the ensemble mean and spread of the "contribution from initialisation" derived from the decadal hindcasts.
Smoothing: To ensure a seamless transition and avoid rapid jumps in weights, the weights are smoothed across lead times using a maximum-up-to-lead-time approach, retaining information about initial predictability while allowing the influence of initialisation to wane as predictability decreases.
Evaluation: The performance of the merged product is evaluated against reanalysis (ERA5) using three metrics:
- Anomaly Correlation Coefficient (ACC): Measures skill (precision).
- Spread-over-Error (SOE): Measures reliability (dispersion).
- Fraction within 80% Confidence Interval (FRAC80): Measures reliability.
- Effective Ensemble Size: Quantifies the independent information retained after weighting.

Key Contributions

Seamless Integration: The method provides a coherent framework for merging initialised forecasts with long-term projections, bridging the gap between near-term precision and long-term resilience.
Independence from Forced Signal: By removing the forced signal before weighting, the method avoids double-counting information and ensures the weights are independent of any prior trend-based constraints applied to the projections.
Dynamic Weighting: Unlike static shadowing methods that select a fixed subset, this approach applies likelihood-based weights that evolve with lead time, naturally relaxing constraints as predictability decreases.
Effective Ensemble Size: The method eliminates the need for subjective selection of the number of projections to retain, instead deriving an effective ensemble size based on the calculated weights.

Results
The method was tested on four large-scale indices: Summer Subpolar North Atlantic (SPNA) temperatures, Annual Global Mean Temperature (AGMT), December-February ENSO, and January-March North Atlantic Oscillation (NAO).

SPNA Temperatures: The merged product adds value over raw projections for lead times up to 5 years. However, the merged product became under-dispersive (observations fell within the 80% interval only 60% of the time), primarily because the underlying hindcasts underestimated warming trends and exhibited wide uncertainty.
AGMT: The merged product showed high skill across all lead times, driven by the strong forced signal. The initialisation weighting provided value primarily at $T+1$ , capturing variability around the trend (e.g., post-1998 cooling) better than unweighted projections.
ENSO: The merged product captured the high skill of hindcasts at $T+1$ (correlation ~0.87), but this skill dropped off rapidly, with no added value from initialisation by $T+4$ . The merged product became under-dispersive because the underlying CMIP6 ensemble was already under-dispersive regarding ENSO variability.
NAO: Due to low signal-to-noise ratios and biases in model variability, the merged product showed only modest skill at $T+1$ and remained under-dispersive. The method could not overcome fundamental model deficiencies in representing NAO variability.

Significance and Claims
The paper claims that the merged product offers a practical tool for decision-makers who require consistent climate information across time scales. It allows planners to identify specific "windows of opportunity" (start dates) where initialised forecasts add value to long-term projections. The authors emphasize that the method is most effective for variables where the forced signal is well-represented and where decadal forecasts demonstrate genuine skill in predicting modes of variability. Specifically, the merged product adds value over raw projections for lead times of 2 years for global temperature, 3 years for ENSO, and 5 years for North Atlantic subpolar gyre temperatures.

The authors explicitly state that the merged product inherits limitations from its parent datasets. For variables with poor model representation of variability or signal (like the NAO), the merged product does not resolve these fundamental issues. They caution that while the method improves the transition between time scales, operational use requires careful consideration of data availability (specifically hindcast data for newer models) and the reliability (dispersion) of the resulting product. The paper concludes that the merged product is not yet operationally ready for all variables but provides a robust framework for assessing when and where initialised information can refine long-term projections.

Technical Summary: Using Decadal Forecasts of Modes of Variability to Refine the First Few Years of Climate Projections

More like this