[Paper Review] Five year ahead prediction of Sea Surface Temperature in the Tropical Atlantic: a comparison between IPCC climate models and simple statistical methods

Thomas Laepple, Stephen Jewson|ArXiv.org|Jan 15, 2007

Tropical and Extratropical Cyclones Research1 references10 citations

TL;DR

This study compares five-year ahead predictions of tropical Atlantic sea surface temperature (SST) using IPCC climate models versus a simple statistical method (8-year flat-line average). Despite climate models' complex physics, the statistical model outperforms them, especially when models lack realistic initial conditions or volcanic forcing, concluding that statistical methods are more reliable for short-term hurricane activity forecasting due to lower complexity and comparable or better skill.

ABSTRACT

There is a clear positive correlation between boreal summer tropical Atlantic sea-surface temperature and annual hurricane numbers. This motivates the idea of trying to predict the sea-surface temperature in order to be able to predict future hurricane activity. In previous work we have used simple statistical methods to make 5 year predictions of tropical Atlantic sea surface temperatures for this purpose. We now compare these statistical SST predictions with SST predictions made by an ensemble mean of IPCC climate models.

Motivation & Objective

• To evaluate whether state-of-the-art IPCC climate models can outperform simple statistical methods in predicting five-year-ahead tropical Atlantic sea surface temperatures (SST).
• To assess the predictive skill of climate models under idealized conditions—specifically, perfect knowledge of future climate forcings—versus statistical models relying on historical SST trends.
• To determine whether the complexity of climate models justifies their use over simpler statistical approaches for hurricane activity forecasting on the 0–5 year timescale.
• To investigate the role of internal climate variability and external forcings (e.g., volcanic activity) in determining model performance over decadal prediction horizons.

Proposed method

• Uses historical SST data from the HADISST dataset (1860–2005), focusing on the Main Development Region (10°–20°N, 15°–70°W) for July–September.
• Applies a statistical model—the '8-year flat-line'—which predicts future SSTs as the simple average of the previous eight years’ SSTs.
• Employs an ensemble mean of 22 coupled ocean-atmosphere climate models from the IPCC 20th-century simulations, split into those with and without volcanic forcing.
• Applies a bias correction to model predictions using the last eight years of observed SSTs to align model means with current conditions, enabling fair comparison with the statistical model.
• Calibrates both model types using the same 8-year window to ensure that identical predictions would result if models predicted constant SSTs.
• Evaluates prediction skill over a 5-year lead time using root-mean-square error (RMSE) comparisons between predicted and observed SSTs.

Experimental results

Research questions

• RQ1 Can IPCC climate models with perfect knowledge of future forcings outperform a simple statistical model in predicting five-year-ahead tropical Atlantic SSTs?
• RQ2 To what extent does the inclusion of realistic volcanic forcing improve the predictive skill of climate models compared to statistical methods?
• RQ3 Does the internal variability simulated by climate models provide any predictive advantage over statistical models that rely on historical averages?
• RQ4 How does the performance of climate models degrade over time relative to statistical models, and what does this imply about their long-term reliability?
• RQ5 Under what conditions might climate models eventually surpass statistical models in predictive skill for decadal SST forecasts?

Key findings

• The 8-year flat-line statistical model consistently outperforms climate models that do not include volcanic forcing, even when those models have perfect knowledge of all other climate forcings.
• Climate models that include perfect volcanic forcing predictions show a slight improvement over the statistical model at lead times up to five years, but this advantage diminishes rapidly beyond that.
• Climate models without volcanic forcing perform only slightly better than the statistical model over the first five years, but their error increases more quickly with lead time, indicating a tendency to drift from observed conditions.
• The primary source of skill in both model types is the bias correction, which aligns predictions with the current level of decadal variability—suggesting that the real predictive power comes from capturing the current phase of internal variability.
• The climate models’ poor performance is attributed to random initial conditions and an inability to sustain realistic internal variability patterns, leading to systematic drift that degrades forecast accuracy over time.
• Given the massive computational cost and complexity of climate models versus the simplicity and comparable performance of statistical models, the authors conclude that statistical methods are preferable for 0–5 year hurricane activity predictions.