Sampling biases in CMIP5 decadal forecasts

Abstract

Recent studies examining the fidelity of decadal hindcast experiments from phase 5 of the Coupled Model Intercomparison Project have highlighted the need for larger ensembles of forecasts, compared to the initial five yearly spaced initializations, to help correct for model biases (drift). This study quantifies differences in the two drift estimates in sea surface temperature (SST) and SST anomaly (SSTA) predictions, between experiments initialized every 5 years and those initialized every year. The effect of the recommended mean drift correction, on the two sets of predictions, is also analyzed. Our results indicate that differences between the SST drift estimates are largest over the tropical Pacific. Moreover, this difference is large for Niño 3.4 and almost negligible for the global average SSTA. Drift correction as per the mean drift from the 5 year case leads to spurious peaks in the drift‐corrected Niño 3.4 (and the tropical Pacific) and sporadic improvements in skill. This problem with Niño 3.4 stems from an aliasing that occurs during the drift calculation that results from a combination of the timing of major El Niño events in relation to the initialization dates. The study recommends accounting for such sampling effects while considering any subset of the full data set.

Assessing the Decadal Predictability of Land and Ocean Carbon Uptake

“…Other sources of uncertainties might arise from the relatively small sample of both starting dates and ensemble size. For example, Choudhury et al () show that sampling biases might impact predictive skill assessment, especially at a regional scale. Similarly, a larger ensemble size would be expected to refine our estimate of model predictability (Deser et al, ).…”

An Assessment of Drift Correction Alternatives for CMIP5 Decadal Predictions

“…Although there are other models available from the CMIP5 archive, only these models provide annual initializations over this period (other models only have initializations every 5 years). The issues arising from using an undersampled data set have already been well documented in the literature (e.g., Choudhury, Sharma, et al, ; García‐Serrano & Doblas‐Reyes, ; Smith et al, ). Choudhury, Sharma, et al () showed that using model data sets with 5 year spaced initializations leads to very large biases in the estimated drift over the tropical Pacific because of a spurious imprinting of the observed ENSO signal and, thus, leads to biased results even after drift correction.…”

“…For anomaly initialization, models are initialized by adding the observed anomalies (relative to the observed mean state) to the model climatology (Smith et al, ). Theoretically, anomaly‐initialized models should not drift (Smith et al, ), although some drift generally still occurs (Choudhury, Sharma, et al, ; Kim et al, ), albeit substantially reduced. This is because the model may not simulate the long‐term trend correctly (Kharin et al, ) and the model's erroneous mean state can lead to dynamical imbalances that lead to biases and poor prediction skill (Boer et al, ).…”

“…Inset shows the ensemble mean biases (gray) for the four initializations and the mean model drift (black; i.e., the average of all initialization ensemble mean biases). The ensemble mean bias from the decade starting in 1960 is shown in brown and that in 1990 is shown in green (modified version from Choudhury, Sharma, et al, ).…”

Predictability of Decadal Atlantic Meridional Overturning Circulation Variations

“…In the general case, the model attractor will be related to the real one by some combination of these three transformations. Figure 6 illustrates that the model bias may vary, not only with lead time, but with model state (e.g., whether or not the model is in an El Niño state; Choudhury et al, 2016) and whether the local model trend is different from the true trend (see Meehl, Teng, & Arblaster, 2014). For decadal forecasting in particular,…”