Faulty model forces rapid retraction of paper on sea ice and climate change

Last month, researchers published a paper whose conclusions suggested that looking at Arctic sea ice in the autumn offers clues to winter temperatures in Europe.

The letter appeared — briefly, as this post will demonstrate — in Nature Geoscience. The letter, titled “High predictability of the winter Euro–Atlantic climate from cryospheric variability,” was written by Javier Garcia-Serrano and Claude Frankignoul, of the Université Pierre et Marie Curie. The journal published the letter on March 23 and retracted it on April 14.

Here’s the abstract, which can still be found online:

Seasonal prediction skill for surface winter climate in the Euro–Atlantic sector has been limited so far^{1, 2, 3}. In particular, the predictability of the winter North Atlantic Oscillation, the mode that largely dominates regional atmospheric and climate variability, remains a hurdle for present dynamical prediction systems^{4, 5}. Statistical forecasts have also been largely elusive^{6, 7, 8}, but October Eurasian snow cover has been shown to be a robust source of regional predictability^{9, 10}. Here we use maximum covariance analysis to show that Arctic sea-ice variability represents another good predictor of the winter Euro–Atlantic climate at lead times of as much as three months. Cross-validated hindcasts of the winter North Atlantic Oscillation index using September sea-ice anomalies yield a correlation skill of 0.59 for the period 1979/1980–2012/2013, suggesting that 35% of its variance could be predicted three months in advance. This skill can be further enhanced, at the expense of a shorter lead time, by using October Eurasian snow cover as an additional predictor. Skilful predictions of winter European surface air temperature and precipitation are also obtained with September sea ice as the only predictor. We conclude that it is important to incorporate Arctic sea-ice variability in seasonal prediction systems.

But that conclusion, it turns out, left other experts cold. According to the notice:

In our Letter, a conceptual error in the cross-validation approach led to an overestimation of the predictive skill of the winter (December–February) North Atlantic Oscillation (NAO) and Euro-Atlantic surface climate from Arctic sea-ice variability. The error does not affect the hindcasts based on the snow advance index.

Specifically, to produce the one-year-out cross-validated hindcasts based on Arctic sea-ice variability, we performed a maximum covariance analysis (MCA) between Arctic sea-ice concentration (SIC) anomalies and winter Euro-Atlantic sea-level pressure anomalies in the period 1979/80 to 2012/13. We then applied one-year-out cross-validation using subsets of years from the SIC time series derived from the whole period. Thereby, the regression coefficients (that is, slope and intercept) and predictor value of the statistical model were estimated assuming the knowledge of the MCA fields in the year out. This procedure overestimates the cross-validation skill.

One-year-out cross-validated hindcasts instead require cross-validation of the MCA pattern-generation in the year out, thus performing an MCA on the remaining years. Following this approach, the cross-validated skill in hindcasting the winter NAO index using September SIC over the whole Arctic is 0.08, indicating that there is no predictive skill from Arctic sea-ice variability. The cross-validated NAO skill using October or November SIC over the whole Arctic is 0.22 and 0.18, respectively, suggesting some skill.

Although our analysis reveals no skill in sea-ice-based NAO predictions with three months lead time, the limited skill from October–November sea-ice concentration supports the notion that sea-ice information should be incorporated in dynamical prediction systems to improve their skill at forecasting the surface winter climate in Europe.

Following the identification of the error in our cross-validation approach, this Letter has been retracted. We are grateful to Geert Jan van Oldenborgh (KNMI, De Bilt, The Netherlands) for identifying this error. We also thank Francisco J. Doblas-Reyes (IC3, Barcelona, Spain) for discussions.

Hat tip: Steve Forden