Seasonal climate prediction for North Eurasia

An overview of the current status of the operational seasonal climate prediction for North Eurasia is presented. It is shown that the performance of existing climate models is rather poor in seasonal prediction for North Eurasia. Multi-model ensemble forecasts are more reliable than single-model ones; however, for North Eurasia they tend to be close to climatological ones. Application of downscaling methods may improve predictions for some locations (or regions). However, general improvement of the reliability of seasonal forecasts for North Eurasia requires improvement of the climate prediction models.


Introduction
The socio-economic development of the countries located in North Eurasia is strongly climate and weather dependent. The importance of seasonal climate prediction has been realized by the world meteorological community. The World Meteorological Organization (WMO) initiated the establishing of the Lead Center for Long Range Forecast Multi-Model Ensembles (WMO LC LRF-MME) in Seoul, Korea, and the Lead Center for the Long Range Forecast Verification System (WMO LC SVS-LRF) in Melbourne, Australia. Twelve operational centers issuing global seasonal model predictions have achieved the status of WMO Global Producing Centers (GPC). In March 2012, WMO plans to start issuing a global seasonal climate update: a climate monitoring and seasonal climate forecasting bulletin with global coverage. The forecasting section of the bulletin will be based on the results produced by the Global Producing Centers and both WMO Lead Centers. This activity essentially improves the provision of the world's societies and economies with seasonal forecasts.
A brief overview of the current status of seasonal forecasting for North Eurasia and possible ways of improving it is given in this work.

Single-model forecasts
In order to improve the regional interpretation of the global long range forecasts produced by the Global Producing Centers, WMO established a set of WMO Regional Climate Centers, with the North Eurasia Climate Center (NEACC) being in charge of the operational long range forecasts for North Eurasia. NEACC issues a set of long range forecasts based on the ensemble predictions from two global circulation models, the model of the Main Geophysical Observatory (Mirvis et al 2006a(Mirvis et al , 2006b and the model of the Hydrometeorological Research Center of Russia (Tolstykh 2010). The one-month-lead seasonal climate prediction is based on the outputs from the single model of the Hydrometeorological Research Center of Russia (Tolstykh 2010). It should be noted that according to the WMO regulations (WMO 2002; see their figure 1), a one-month-lead seasonal forecast is a forecast of the seasonal mean values issued one month in advance of the beginning of the forecast period. For example, the seasonal forecast for wintertime (December through February, DJF) temperature implies the issuing of the forecast of the seasonal (DJF) mean temperature values by the beginning of November, with model integrations being initialized using October observations. In Although being successful in the prediction of the extremely low Arctic Oscillation index and the extremely cold North Eurasia winter of 2009/10, in general, NEACC's seasonal climate predictions, similarly to those from all the seasonal climate prediction models, are quite poor in the middle and high latitudes. Figure 1 shows the rank probability skill score (RPSS; Wilks 1995) of the seasonal forecasts of tercile probabilities interpolated from the GPC-Moscow model outputs as regards climatological forecasts for 70 weather stations more or less uniformly distributed throughout North Eurasia. These 70 stations comprise a standard set of the stations used at the Hydrometeorological Research Center of Russia and NEACC for verification of the long range forecasts. The postprocessing of the model forecasts includes correction of the model bias in accordance with the WMO requirements (WMO 2010), that is, the systematic errors in model climatology are corrected on the basis of the climatology of the historical forecasts and that of the observations.
The verification assessments shown in figure 1 have been performed on the cross-validated forecast series with five years withheld. WMO recommends (WMO 2002) that any influence of the data from the years which follow the year of forecast should be prevented in a cross-validation procedure. An analysis of the autocorrelation structure of the station temperature series has revealed that for almost all of them an autocorrelation function becomes negative at the time shift of one-two years. So, the verification assessments performed on the cross-validated series, with the data for five years (the year of forecast and the following four years) being withheld, match the WMO recommendations.
The RPSS values exceeding approximately 0.08 (orange and red points) are statistically significant at the 2.5% significance level in the one-tailed test as revealed from 1000 Monte Carlo runs. The threshold value has been exceeded for several stations. Since the total number of the stations is 70, there is a probability that these significant RPSS values have been obtained occasionally (Wilks 1995, chapter 5.4) and field significance tests accounting for spatial dependences between the stations have been conducted. Indeed, the probability of obtaining of such numbers (11 for winter and 8 for summer) of 2.5% significant RPSS values occasionally exceeds 5% and 8% for winter and summer, respectively, as revealed from 1000 Monte Carlo tests, with spatial dependence between the stations being accounted for by scrambling in the time domain of the whole yearly sets of station forecasts. Unfortunately, as was mentioned above, a poor performance of seasonal climate predictions for the middle and high latitudes is a common shortcoming of all the seasonal climate prediction models, even the state-of-the-art ones, both two-tier and one-tier (Kharin and Zwiers 2001, Yun et al 2003, Kryjov et al 2006, Bundel et al 2011. The conclusion of the poor performance, particularly for North Eurasia, is supported by the verification assessments of the seasonal forecasts from the twelve WMO Global Producing Centers which are posted on the Web-site of the WMO LC SVS-LRF (www.bom.gov. au/wmo/lrfvs/index.html).

Multi-model forecasts
The tendency in the past decade has been to use multi-model ensembles in seasonal climate prediction. Nowadays, a number of (mainly international) climate centers issue multi-model seasonal climate predictions. For North Eurasia, multi-model ensemble predictions issued by Asia Pacific Economic Cooperation Climate Center (APCC, Busan, Korea), the International Research Institute for Climate and Society (IRI, Palisades, New York) and WMO LC LRF-MME are available. As has been shown by numerous studies (Kharin and Zwiers 2002, Peng et  In this formula X is a single-model ensemble member forecast; y is a signal-a fraction of the forecast positively correlated with observations Y; err cl is a (constant) bias of the model climatology with respect to observed climatology caused by the model formulation; err m is a varying model systematic error caused by the model formulation; err n is a 'weather noise'-normally distributed random error. err cl is considered constant and it is easily eliminated by adjustment of the model climatology to the observed one on the basis of historical forecasts. The last term, the random error, is reduced by the use of an ensemble of model integrations started from different initial conditions, with this random error converging to zero with increase of the ensemble size. However, as is obvious from the equation, no single-model ensemble size increase leads to a reduction of err m . The purpose of the use of a multi-model ensemble is just the reduction of this error. This is based on the hypothesis that these errors from individual models are independent of each other. Under these conditions, since signal y is common to all the models, the averaging of the model forecasts leads to an increase of the signal to error ratio because the errors from individual models tend to mutually offset each other.
Indeed, in the regions for which predictability is high, e.g., the equatorial Pacific, such offset of the model errors leads to extraction of a useful signal common to all the participating models and improvement of the seasonal forecast. Meanwhile, for the regions for which atmospheric predictability is low and the errors are larger than the useful signal, mutual offset of the errors results, on the one hand, in increasing of the reliability of forecasts, but, on the other hand, in the tendency to lead to forecasts close to climatology ones (Min et al 2009). In particular, on the basis of seven-model ensemble historical forecasts for 21 yr, Min et al (2009) have shown that about 70% of the multi-model forecasts of the tercile probabilities of temperature for the northern extratropics are within the range of 0.20-0.40, which is close to the climatological tercile probability value of 0.33. In comparison, the fraction of such forecasts for the Tropics, for which predictability of temperature is high, does not exceed 40%. North Eurasia is an area of low predictability (see, e.g., Robertson et al 2004, Min et al 2009, Bundel et al 2011, with the fraction of the forecasts close to climatological ones being about 80%. In the global multi-model ensemble seasonal forecasts issued by the climate centers which operate large multi-model ensembles, APCC (www.apcc21. org), IRI (portal.iri.columbia.edu) and WMO LC LRF-MME (wmolc.org), forecasts for North Eurasia tend, as a rule, to be close to climatology ones, that is, uncertain. The multi-model ensemble approach in seasonal forecasting is a great step forward as compared with a single-model one. For the regions with appropriate predictability, the use of multi-model ensemble forecasts provides a more reliable basis for decision-making than single-model forecasts and the advantage of the multi-model approach is entirely obvious for these regions. Meanwhile, for the regions with low predictability, particularly for North Eurasia, the multi-model ensemble forecasts prevent us from making wrong decisions, which is a definite advantage; however, support for correct decisions based on them is quite limited.

Future development
A reasonable way to improve seasonal climate prediction for North Eurasia is to apply a downscaling. Developed for quantification of the local effects of global change processes (von Storch 1995, Zorita andvon Storch 1999), nowadays statistical downscaling is becoming a standard and routine technique in long range forecasting for regional/local scales (see, e.g., Feddersen and Andersen 2005, Kang et al 2009, Juneng et al 2010. The downscaling approach is based on the view that the regional climate is conditioned by two factors: the large scale climate state and regional (local) physiographic features. However, the use of the downscaling technique in seasonal climate prediction differs from that in climate projection for which it was initially developed. Whereas in climate projections the relationships are established between the observed local variables and large scale fields derived from observations (e.g., by reanalysis), in seasonal predictions the relationships are established between the observed local variables and model simulated forecast fields in which the large scale atmospheric patterns are usually spatially shifted as compared with observations. Therefore, downscaling in seasonal prediction yields model forecast correction rather than localization.
Correcting the raw model forecasts with the downscaling technique, it is possible to improve to some degree the forecasts and forecast performance. In particular, this is shown by the results from the tests with the downscaling method suggested by Min et al (2011) modified for a single-model version. This method is based on regression with a probabilistic interpretation of the forecasts based on the estimation of forecast uncertainty associated with both regression random errors and the ensemble spread. Another peculiarity of the method is the selection of the model predictor (model grid-point series), the predictand being the station temperature, with the use of the cross-validation cycle, which allows one to improve the stability of the selection of predictors. This 'inner' cross-validation cycle, which is used for the selection of the model predictors, should not be confused with an ordinary verification cross-validation cycle, which in this case is an 'outer' one. Forecasts are formulated in terms of tercile probabilities based on the Gaussian probability distribution function (PDF). The parameters of the forecast PDF are estimated as an arithmetic mean of the mean values and standard deviations of the Gaussian PDFs predicted from the simple linear regressions (with estimated uncertainty) of the station temperature on the selected model grid-point series which feature a congruence coefficient significant at the 5% level (depending on the effective number of degrees of freedom, the grid-point threshold coefficient value slightly varies around 0.45). It should be noted that the use of the arithmetic mean of the standard deviations implies that the grid-point series are not considered independent. The details of the downscaling method applied are described by Min et al (2011).
The verification assessments of the downscaled forecasts have been performed on the cross-validated forecast series (the 'outer' cross-validation) with five years withheld, that is, similarly to the verification assessments of the interpolated forecasts shown in figure 1.
Results from the tests are shown in figure 2. The basic model forecast is the same as for figure 1-it is the single-model GPC-Moscow global forecast-and the assessment procedures are absolutely the same so this figure can be compared with figure 1. The number of stations for which RPSS exceeds 0.08 is 29 for winter and 21 for summer. The probability of obtaining such figures occasionally is less than 0.1%.
In general, the number of stations (out of 70) for which downscaled forecasts outperform climatological ones is 44 for winter and 43 for summer forecasts. For interpolated forecasts those figures are 28 and 27, respectively. So, downscaling improves to some degree the performance of the seasonal predictions. However, it cannot be assessed as reliable yet. Even improvement by downscaling based on the multi-model ensemble  does not achieve very much.

Conclusion
Reliable seasonal climate prediction for middle and high latitudes, particularly for North Eurasia, is a great challenge for modern climate science. The work goes in two directions: improvement of the climate models and improvement of the methods of postprocessing the model output, with the former being the basic one. Both simple and sophisticated postprocessing methods are aimed at the correction of the varying systematic model errors. It is required that the model outputs at least contain the useful signal and that this signal is sufficiently large to be extracted from the raw model forecast. Nowadays, the seasonal climate prediction models are governed by the boundary conditions, mainly confined to sea surface temperature in the eastern equatorial Pacific (Shukla et al 2000). The area of reliable seasonal prediction is also mainly confined to the eastern equatorial Pacific. However, many empirical studies (see, e.g., Mosedale et al 2006, Sinha and Topliss 2006, Pan 2007 reveal that some seasonal predictability may arise from the heat and mass exchange between the ocean and atmosphere in other regions, including the extratropics. In addition, inclusion of the highly resolved stratosphere may also improve the long range simulation of the atmospheric processes in the extratropics (see, e.g., Baldwin and Dunkerton 2001, Perlwitz and Harnik 2004, Kryjov and Park 2007. General improvement of the reliability of seasonal forecasts for the Earth regions with low atmospheric predictability requires improvement of the climate prediction models. Meanwhile, reliable seasonal climate prediction for North Eurasia remains a great challenge.