An assessment of seasonal predictability using atmospheric general circulation models

Abstract

AbstractSeasonal predictability is investigated using a 15‐year set of 4‐month range, 9‐member ensemble integrations from atmospheric general circulation models (AGCMs) involved in the European PROVOST project (PRediction Of climate Variations On Seasonal to interannual Time‐scales). The integrations were performed using prescribed ideal (observed) sea surface temperatures (SSTs), therefore skill attained (referred to as ‘potential’ skill) represents an estimated upper bound on skill achievable with current models using predicted SSTs. Most analysis is presented for The Met. Office Unified Model (UM), the European Centre for Medium‐Range Weather Forecasts (ECMWF) T63 model (T63) and an 18‐member multiple‐model ensemble (JT2) constructed from these individual models. The benefits of higher‐order multiple models (employing all four participating PROVOST AGCMs) are also investigated. Evaluation is focused on four assessment regions: the tropics; the northern extratropics; Europe; and North America. Probabilistic skill is assessed for the basic events: 3‐month mean 850 hPa temperature above/below normal; 3‐month mean precipitation accumulation above/below normal. Deterministic (ensemble mean) skill is also assessed. A summary of the main results is provided below.• Potential skill: Skill scores for months 1–3 forecast 850 hPa temperature and precipitation calculated for the entire tropical and northern extratropical regions indicate that, while skill is highest in the tropics, it is also available over the northern extratropics in all seasons. Scores for the northern extratropics are highest in spring (March‐April‐May; MAM). Scores for precipitation are generally lower than for 850 hPa temperature, however, there is evidence of substantial potential for rainy season predictions in some tropical regions. Over Europe and North America skill scores for 850 hPa temperature are (for at least one of the UM, T63 and JT2 models) comparable to those of the northern extratropics in all seasons. Peak skill occurs over Europe in MAM (as found for the northern extratropics). In contrast, peak skill over North America occurs in December‐January‐February (DJF), apparently as a result of enhanced predictability during El Niño Southern Oscillation (ENSO) events. In non‐ENSO years skill over Europe and North America is similar, suggesting that the greater predictability often attributed to the North American region relative to Europe may apply only during ENSO events. Skill for months 2–4 is generally lower than for months 1–3, though there is evidence that during ENSO events levels of skill in the first three months are maintained into the second three months. For precipitation, best skill over Europe and North America is found in MAM and DJF, with little evidence of any skill over Europe in summer and autumn.• Skill prediction: Largest ENSO‐related skill enhancements over North America are found in DJF and over Europe in the following (post‐ENSO peak) MAM. Ensemble spread appears a useful indicator of ensemble‐mean skill in some seasons over Europe and North America. Thus prospects for skill prediction appear promising, perhaps using strategies which combine information on both the state of ENSO and ensemble spread.• Benefits of multiple‐model ensembles: Multiple‐model ensembles enhance prediction capabilities, allowing the strengths of the individual AGCMs to be exploited without extensive a priori calibration of each model. The multiple‐model ensembles frequently provide a filter for the more skilful individual model (the identity of which varies with season and region). The key factor determining the skill of the multiple model appears to be the skill of the most skilful component ensemble, and does not appear to be strongly connected with the increased ensemble size.• Use of persisted SST anomalies: Tests indicate that a substantial proportion of the skill achieved using observed SSTs is retained using persisted SST anomalies (SSTA) from the month preceding the initial date of the integration, indicating that use of persisted SSTA is a viable method for real‐time seasonal prediction, at least for up to one season ahead.• User value: A methodology for linking technical forecast quality with financial value for users has been outlined using the relative operating characteristic and the user cost/loss matrix. Results indicate promising potential for user value of probabilistic seasonal predictions not only over tropical areas but also in some extratropical areas, including Europe.