Seasonal prediction is based on changes in the probability of weather statistics due to changes in slowly varying forcings such as seasurface temperature anomalies, most notably those associated with El Niňo-Southern Oscillation (ENSO). However, seasonal weather can be perturbed by many factors, and is very much influenced by internal variability of the atmosphere, so comprehensive models are needed to identify what can be predicted. The predictability and probabilistic nature of seasonal

more »

... sts is explained with suitable examples. Current capabilities for seasonal prediction that have grown out of work done in the research community at both national and international levels are described. Dynamical seasonal prediction systems are operational or quasi-operational at a number of forecasting centres around the world. Requirements for seasonal prediction include initial conditions, particularly for the upper ocean but also other parts of the climate system; high quality models of the ocean-atmosphere-land system; and data for verification and calibration. The wider context of seasonal prediction and seamless forecasting is explained. Recommendations for the future of seasonal prediction and climate services are given. 56 T. N. Stockdale et al. / Procedia Environmental Sciences 1 (2010) 55-80 1.1.1. Forcings external to the climate system Volcanoes and solar variations have been recurring themes historically in discussions of seasonal prediction. Variations in solar forcing are, however, generally comparatively small and tend to operate on long timescales with the most notable being the 11-year solar cycle. Van Loon et al. [1] review some aspects of solar forcing but the deduced effects are not strong on seasonal timescales. Volcanoes, on the other hand, clearly affect climate (for example, Robock,[2]; Stenchikov et al. [3] ) and have the potential to add skill to certain seasonal forecasts made after large eruptions. However, Collins [4], in a study of the effects of the El Chichón and Pinatubo eruptions, notes that "the volcanic signal is only robust on the spatial scale of continents and, moreover, the signal can easily be contaminated or completely obscured by climate variability". Anthropogenic forcing effects from greenhouse gases and aerosols are often neglected in seasonal prediction with the presumption that the effects are small compared to natural variability and that the global warming signal is, in any case, largely incorporated into the forecast in the initial and boundary conditions. However, Liniger et al. [5] and Boer [6] show that specification of anthropogenic forcing does influence seasonal forecasts, and the time-variation of greenhouse gas and aerosol forcing are now increasingly being introduced into seasonal prediction schemes. Predictability from internal variability of the climate system The natural workings of the climate system give rise to internally generated variability on all timescales. Some of the processes, such as the development of synoptic systems in the atmosphere, happen on short timescales, and are simply a source of unpredictable noise for seasonal prediction. However, slow variations of the climate system are often a source of predictability on seasonal timescales. Examples include the longer timescales of the ocean and the coupled atmosphere-ocean system, and also other components such as sea ice, soil conditions, snow cover and perhaps the Quasi-Biennial Oscillation (QBO) and the state of the stratosphere. The most important source of seasonal predictability is El Niňo and the Southern Oscillation. Walker [7] [8] detected a see-saw between the sea level pressure at Tahiti and Darwin, known as the Southern Oscillation (SO) that was later linked to variations in equatorial sea-surface temperature in the eastern Pacific by Bjerknes [9]. This combined ocean-atmosphere ENSO phenomenon is the dominant mode of Earth's interannual climate variability, and its associated global teleconnections produce important temperature and precipitation anomalies across the globe (for example, Ropelewski and Halpert [10]; Rasmusson and Carpenter [11] ). The prediction of ENSO variability with a simple coupled ocean-atmosphere dynamical model was first demonstrated by Zebiak and Cane [12] . Subsequent work has led to the development of today's sophisticated operational ENSO forecast systems using coupled ocean-atmosphere General Circulation Models (GCMs). Oceanic anomalies other than ENSO can also drive temperature and precipitation anomalies on seasonal timescales (for example, Goddard et al. [13]; Cassou et al.[14]), including the connection of the tropical Atlantic with North-east Brazil rainfall.[15][16] the tropical Indian Ocean.[17][18] and the extra-tropical Atlantic (for example, Rodwell and Folland,[19]). Land surface processes have the potential to provide a signal on seasonal timescales, at least in some areas and under some circumstances (for example, Koster et al, [20]; Senevirante et al, [21]) and anomalous snow cover/amount may also have an effect (for example, Fletcher et al..[22] and references therein). Other phenomena that vary on seasonal timescales may be identified in the system and may enhance (or detract from) predictive skill depending on how they are invoked and treated in the forecast procedure. They include the Northern Annular and Southern Annular modes (NAM and SAM), the Pacific North American (PNA) pattern and the North Atlantic Oscillation or NAO, (Hurrell et al.[23] Bojariu and Gimeno [24], among others). These latter are generally thought to be internal dynamical modes of the atmosphere, but with the potential sometimes to act as mediators of seasonal predictability. In some cases the links back to oceanic forcing are clear, in others less so. Much of the variability of the modes may in any case be unpredictable. Recent investigations also suggest that the stratospheric may prefigure and affect tropospheric anomalies (for example, Baldwin and Dunkerton [25]; Ineson and Scaife [26]). The long timescales of the stratospheric QBO could also have an effect under some circumstances (for example, Boer and Hamilton [27]; Marshall and Scaife [28]). Predictability and the probabilistic nature of seasonal forecasts The mechanisms mentioned above, as well as others implicit in complex coupled atmosphere-land-ocean models, offer the possibility of a predictable signal on seasonal timescales. However, a predictable signal always co-exists with unpredictable noise in the seasonal forecasts, due to the effects of synoptic activity and/or other sources of unpredictability. The fraction of the total variance accounted for by the signal component is generally modest, and thus so is the predictability, particularly away from the tropics. Given the limited predictability, it is generally most appropriate to characterize a forecast in terms of a probability distribution. While a deterministic forecast essentially assigns a probability of 1 to a particular forecast result, a probabilistic forecast also indicates the uncertainty of the result. Probabilistic forecasts can be produced in various ways. For dynamical seasonal forecasting systems, the starting point is an ensemble of forecasts Yi, i = 1 ... n, produced using a set of initial conditions that are intended to reflect the uncertainty in these conditions. The forecasts follow different evolutions because of their differing initial conditions, resulting in a spread of trajectories at forecast time. If the trajectories spread apart widely, the inferred probability distribution is also wide and the forecast is uncertain, while a bundle of close trajectories might suggest less uncertainty. However, this neglects error in the forward calculation of the model. Experience has shown that dynamical seasonal forecast models are overconfidenttheir spread is too narrow to match the range of observed outcomesand there is often little relationship between ensemble spread and the error in the forecast. The reason for this is believed to be large model error. Multi-model approaches, where ensembles from different state-of-the-art models are