Time-space distribution of long-range atmospheric predictability

Thomas Reichler and John O. Roads

Scripps Institution of Oceanography

University of California San Diego, 9500 Gilman Dr., La Jolla, CA 92093-0224,

Correspondence to: John O. Roads (jroads@ucsd.edu)

Draft from Tuesday, January 28, 2003

The global 3-dimensional structure of long-range (2 weeks to 1 season) atmospheric predictability was investigated with a general circulation model. Main focus was to find out the role of initial conditions for such predictability as a function of lead time and space. Four types of predictability experiments with different types of initial and boundary conditions were conducted to this end. The experiments were verified against model data and reanalysis to determine perfect as well as real world forecast skill. Spatial maps and vertical cross sections of predictability at different lead times and for the two contrasting seasons were analyzed to document the varying influence of initial and boundary conditions on predictability. It was found that the atmosphere was remarkably sensitive to initial conditions on the week 3-6 forecast range. Particularly, the troposphere over Antarctica, the region over the tropical Indian Ocean, and the lower stratosphere were affected. It was shown that most of the initial condition memory was related to the persistent nature of the atmosphere in these regions, which in turn was linked to the major modes of atmospheric variability. Possible influences from stratosphere-troposphere coupling were also discussed.

Table of Contents

Abstract 2

Table of Contents. 3

1.???? Introduction. 4

2.???? Methodology. 7

a.??? Model 7

b.??? Experimental setup. 7

c.??? Data. 10

d.??? Analysis procedure. 11

3.???? Horizontal structure of predictability. 13

a.??? Perfect world. 13

b.??? Zonal averages. 15

c.??? Seasonality. 16

d.??? Real world. 17

4.???? Vertical structure of predictability. 18

a.??? Perfect world winter 18

b.??? Perfect world summer 21

c.??? Seasonality. 22

d.??? Real world. 23

5.???? Persistence and initial condition memory. 23

a.??? Persistence of daily heights. 24

b.??? Relationship to major modes. 27

c.??? Antarctic Oscillation. 28

6.???? Summary and Discussion. 30

References. 36

Table and Figure Captions. 39

Table and Figures. 42

1. Introduction

Three fundamental sources affect atmospheric predictability: the model, which contains the physical principles that govern atmospheric flow, the initial atmospheric state, and the atmospheric boundary conditions during the forecast. Predictability is limited because none of these are perfect. Small unavoidable errors grow rapidly through non-linear interactions and lead to useless forecasts after some time. Classical studies of predictability were concerned with the loss of predictability due to the growth of initial errors. Divergence of identical twin experiments, which started from slightly different initial conditions, was used to estimate predictability. The associated limit in forecasting the instantaneous atmospheric state (or weather) was established with simplified models to be about 2-3 weeks (e.g., Lorenz, 1969; Lorenz, 1982).

Although chaotic, the atmosphere can be forced externally by its boundaries towards a particular state. This effect offers the potential for predictability on longer time scales, which was the focus of more recent studies of predictability. Palmer (1993), for example, used the Lorenz (1963) model to show that the probability of being in one regime of phase space or another was no longer equally probable in the presence of external forcing. In particular, sea surface temperature (SST) variations in the equatorial Pacific associated with the El Ni?o Southern Oscillation (ENSO) phenomenon have significant impacts on tropical (e.g., Webster and Chang, 1988) and extratropical planetary circulations (e.g., Horel and Wallace, 1981; and many others).

There is reason to believe that initial conditions can also affect long-range predictability (2 weeks to 1 season), since this time scale represents a mixture between the weather and climate prediction problem. Recently, Reichler and Roads (2003a) (hereafter RR) showed that initial conditions dominated a forecast during the first four weeks, and that even thereafter initial conditions contributed to improved predictability. There are certainly several physical mechanisms that might contribute to long-term predictability from initial conditions, associated with long-lived persistent and periodic phenomena. One example is the intraseasonal oscillation (Madden and Julian, 1972), which is established through a certain combination of atmospheric initial and ocean boundary conditions. There are other examples of how initial conditions might affect long-term predictability, like blocking events, or slow shifts of major modes. Recently, observational evidence suggested that the downward propagation of long-lived stratospheric anomalies into the troposphere might have implications for long-term predictability (Thompson et al., 2002).

The main goal of this paper was to determine how important initial conditions are beyond the classical predictability limit of 2-3 weeks. Predictability is in general a function of space, season, and lead time. Therefore, we investigated globally the full 3-dimensional structure of predictability, and examined how it changes with lead time, season, and strength of the boundary forcing. As will be shown, there exists considerable initial condition produced predictability at the week 3-6 time range. This prompted further analysis, aimed at finding possible physical mechanisms behind the large sensitivity to initial conditions. In particular, it is investigated in how far this effect is related to low-frequency variability associated with major atmospheric modes.

Understanding the role of initial conditions is important for progress in the long-term predictability effort. It is, for example, still an open question as to just how sensitive the atmosphere is to initial conditions, and when and where initial conditions can be important for long-range predictability. For operational seasonal forecasts, for example, the various centers use different initialization strategies. At the International Research Institute for Climate Predictions (IRI) a previous forecast rather than analysis are used to initialize the atmospheric model (see Mason et al., 1999). However, at the National Centers for Environmental Prediction (NCEP) (Kanamitsu et al., 2002b), and at the Scripps Experimental Climate Prediction Center (ECPC) (Roads et al., 2001) the initial conditions used for weather predictions (analysis) are also used for seasonal forecast models.

We used a model-based approach and developed four basic types of predictability experiments. Each experiment was forced with different types of initial and boundary conditions, so that we were able to determine their individual contribution to predictability. In most cases, the experiments were verified under the so called perfect model assumption against simulations with the same model. This way we avoided complications with model related errors and, as will be discussed later, could reduce substantially the sampling problem. Since we were also curious as to whether our idealized findings were still applicable in the real world, we also compared the experiments against real world observations.

The paper is structured as follows: In Section 2, model, data and analysis method is described. Section 3 presents a 3-dimensional analysis of monthly mean predictability for the 500 hPa height using model as well as observational data for verification. In section 4, the relationship between initial condition memory, atmospheric persistence, and major modes of variability is established. A summary and discussion is provided in section 5.

2. Methodology

a. Model

The AGCM of this study is the NCEP seasonal forecasting model (SFM) described by Kanamitsu et al. (2002b) and by RR. The model uses the spectral transform method to solve the primitive equation system for vorticity, divergence, virtual temperature, specific humidity, and logarithm of surface pressure. A triangular truncation of 42 spherical harmonics (T42), equivalent to a horizontal resolution of about 280 km, and a vertical sigma coordinate system, which contains 28 layers and 29 levels from the surface to zero hPa, was used. 18 layers are below 200 hPa (i.e., the extratropical troposphere), and 10 layers are above. The SFM was mainly based upon the 1995 version of the Medium-Range Forecast (MRF) model used at NCEP for making analysis and medium-range predictions. Earlier versions of the SFM were used for the NCEP/NCAR reanalysis (Kalnay et al., 1996; Kistler et al., 2001), and the NCEP/DOE reanalysis-2 (Kanamitsu et al., 2002a). The model version for this study was released in October 2000, and differs from the reanalysis-2 model mainly by its parallelized code structure and improved physical parameterizations.

b. Experimental setup

Detailed specifications of the simulations for this study are provided in Table 1. Experiments are designated specific acronyms, which indicate the type of initial (IC) and boundary conditions (BC). Experiment ICBC (good ICs and BCs) used perfect initial and boundary conditions, which is the upper predictability limit for a specific model. Experiment BC (good BCs) used perfect boundary but neutral initial conditions, experiment IC (good ICs) used perfect initial but climatological boundary condition, and experiment iBC (adjusted ICs, good BCs) used AMIP-type initial and perfect boundary conditions.

Each experiment was carried out for a winter and summer season for 22 years from 1979 to 2000. The simulations were started at December (June) 15^th and run continuously through the 14 week long period until the end of the following March (September). To avoid any ambiguity, from now on winter and summer refer to the northern hemisphere (or boreal) seasons. For each experiment, year, and season, an ensemble of 10 simulations were produced. For the reference experiment ICBC, the ensemble size was increased to 20 members. This experimental setup was identical to that of the ?dynamical seasonal prediction? (DSP) project (Shukla et al., 2000), which focused on the winter season.

Individual forecasts of one ensemble experiment were forced with identically evolving boundary conditions, but were started from slightly different initial conditions. These initial conditions were derived from one of the two base runs or NCEP/NCAR reanalysis using the breeding technique (Toth and Kalnay, 1993; Toth and Kalnay, 1997). Breeding calculates the fastest growing modes for a particular initial state using the atmospheric model itself. The resulting perturbations correspond to the observational uncertainty in atmospheric analysis. These initial errors amplify and lead to the subsequent divergence of different ensemble members. Breeding cycles were started five days ahead of the actual initialization date (i.e., Dec. or Jun. 10^th). The initial perturbations came from those used operationally at NCEP, but truncated to T42 resolution (the breeding algorithm and the initial perturbations were made available to us by Z. Toth from NCEP).

We first carried out two continuous AMIP-type base runs to derive initial conditions for the subsequent experiments. Run BASE-O (base run with observed SSTs) was forced with observed global SSTs and sea ice over the 1948-2000 period, and with model generated land boundary conditions. Run BASE-C (base run with climatological SSTs) was forced every year with the same seasonal cycle of climatological SSTs and sea ice, and land boundary conditions were again generated internally by the model. According to the kind of ocean boundary conditions, we designated the initial conditions that were derived from BASE-C as climatological; those derived from BASE-O were designated anomalous.

The perfect experiment ICBC was forced with observed global SSTs and sea ice, but again the model generated the land boundary conditions. ICBC was started from anomalous initial conditions of BASE-O. This was equivalent to using ?observed? initial conditions, but under a perfect model assumption. Experiment IC was designed to measure the effect of initial conditions on predictability. It used the same initial conditions as ICBC, but was forced with climatological ocean and land boundary conditions derived from reanalysis-2. We repeated ICBC and IC for the winter season with initial conditions from reanalysis. We named these experiments IC-r and ICBC-r. To study the effects of boundary forcing alone without the possible influences from initial conditions, experiment BC was started from randomly chosen ?climatological? initial conditions from BASE-C, but forced with the same perfect boundary conditions as ICBC.

Experiment iBC was very similar to ICBC. iBC was forced with observed ocean boundary conditions. The initial conditions were produced by integrating ICBC for one whole year, i.e., from Dec. (June) 15^th of the current year to Dec. (June) 15^th of the next year. The final state was then used as initial conditions for iBC. The influence of the original initial conditions was therefore of little importance, but the atmosphere has adjusted to the boundary forcing at the new initialization time. Therefore it should be expected that the amplitude and phase of the large scale waves were similar to ICBC, at least to the extent that they were controlled by boundary conditions. The synoptic and smaller scales, however, should be completely different. Experiment iBC is comparable to an ensemble of continuous AMIP-type integrations, and to the current operational seasonal forecasting methodology at the IRI. The motivation for this experiment was to find out how much predictability might be lost by excluding the beneficial effects of synoptic scales in the initial conditions on dynamical predictability.

c. Data

The experiments of this study were forced with prescribed ocean boundary conditions, which were derived from observations. SSTs were taken from the UKMO Global Ice and Sea Surface Temperature (GISST) (e.g., Folland and Parker, 1995) data set for the 1948-1981 period, and from satellite-in situ blended SST analysis based on the method of optimum interpolation (Reynolds and Smith, 1994) after 1982. The sea ice distribution was taken from the NCEP/NCAR reanalysis. Some experiments were forced with a climatological seasonal cycle of SST and sea ice. Climatologies were calculated by averaging the observed monthly mean fields over the 50-year period 1950-1999. The monthly mean data were then linearly interpolated to daily values. For most experiments, land boundary conditions were determined internally by the land surface scheme of the model. In cases where observed land boundary conditions were needed, they were replaced by daily fields of soil moisture and snow cover taken from NCEP/DOE reanalysis-2. They are close to observations since the reanalysis-2 land surface model was driven by observed precipitation (Kanamitsu et al., 2002a). At the time of this study, these fields were the best available global land surface analysis. The climatological mean seasonal cycles of soil moisture and snow cover were derived from NCEP/DOE reanalysis-2 by averaging over the period 1979-1998.

d. Analysis procedure

Spatial maps of predictability were constructed by calculating at each grid point the temporal anomaly correlation (AC) between the time series of a verification and a prediction data set. The temporal AC was used since it represents a common measure of atmospheric predictability (e.g., Kumar et al., 2002; Peng et al., 2000). Anomalies for each experiment were taken with respect to the individual 1979-2000 climatology. Each anomaly time series consisted of seasonal or monthly mean fields over 22 years (1979-2000) of the same lag. 10-member ensemble means of the experiment under consideration were used as prediction time series, i.e.,

,???????????????????????????????????????????????????????????????????? (1)

where ?is the monthly mean anomaly of the prediction of member r during month m and year y. Under the perfect model assumption, we selected experiment ICBC as verification since it had perfect initial and boundary conditions. Since verification and prediction time series came from the same model, model dependant systematic errors were eliminated. If ?is the ensemble mean anomaly of the prediction, and ?the anomaly of a single member s of the corresponding verification, then the temporal anomaly correlation for month m and verification member s is given by

,????????????????????????????????????????????????????? (2)

where ?means averaging over Y years. Each of the S=20 members of experiment ICBC could be selected as verification, so that the calculation was repeated 20 times. The average of the resulting 20 AC maps were used as the final estimate of predictability, i.e.,

.??????????????????????????????????????????????????????????????????? (3)

The verification of ICBC against itself gave the upper bound of predictability with this model, since the model as well as initial and boundary conditions were perfect. To avoid any bias towards higher predictability, again 10-member (instead of the possible 19-member) ensemble means were taken from ICBC as prediction, and another arbitrarily chosen simulation was taken as verification time series. Some experiments were also compared against real world observations. In this case, observational estimates in the form of NCEP/NCAR reanalysis were used as verification time series. Since only one such time series was available, the resulting predictability maps were considerably noisier.

Since AC values are bounded by ?1, they can be non-normally distributed. To make differences between AC values more meaningful, we applied a simple Fisher Z transformation (e.g., Roads, 1988) to the correlations, that is,

?????????????????????????????????????????????????????????????????????? (4)

before taking the differences, and then again transformed back the differences. One should also keep in mind that correlation coefficients are statistically significant only when they exceed certain values. At the 95% error level, the critical AC values were about 0.35 for all years (n=22). In some of our analysis we excluded certain years, so that the critical values were correspondingly larger (e.g., 0.5 for n=14, 0.6 for n=8).

3. Horizontal structure of predictability

In this section we present maps of monthly mean predictability at the 500 hPa level for different experiments, seasons and verification methods. We focused mainly on monthly mean predictability during forecast week 3-6, which corresponds to January and July. At this time range, the high deterministic predictability period at the beginning of a forecast is mostly excluded, although the effects of initial conditions are likely to be still important, and the averaging period is long enough for boundary effects to be detectable. Furthermore, we focused on predictability of the 500 hPa height, since this variable is widely used in this kind of studies. The predictability of other levels will be described in the following section.

a. Perfect world

The geographical distribution of temporal anomaly correlations of January mean 500 hPa heights is presented in Fig. 1 under perfect model assumption, where ensemble means of each experiment were verified against individual simulations of ICBC. Panels on the left are absolute AC scores, and panels on the right show differences to experiment ICBC. Simulation ICBC (top) shows the upper predictability limit with this model, since model, initial conditions and boundary conditions were all perfect. Main regions of high predictability were the tropics, the Pacific-North-American (PNA) region (e.g., Wallace and Gutzler, 1981), and the Pacific-South-American (PSA) region. The PSA region is the southern hemispheric counterpart to the PNA region (e.g., Karoly, 1989). Interestingly, predictability was also enhanced over the Antarctic continent. Independent experiments by Kumar et al. (2002) with the NSIPP model (Bacmeister et al., 2000) confirm the existence of a high predictable region over the Antarctic continent, indicating that this feature was not just a product of our particular model. Experiment iBC (2^nd row) had in general lower skill scores than ICBC, and predictability for BC (3^rd row) was even lower, showing the beneficial effect of initial conditions on long-term predictability. This was particularly evident over the Southern Hemisphere, but also to some extent over all other areas.

Experiment IC, on the other hand, had over some regions relatively high predictability just from using initial conditions. These were in particular the Southern Hemisphere, but there was also some skill over the PNA region, the PSA region, and the tropics. The latter was surprising since the tropics are usually regarded as being dominated by influences from boundary forcing. When the AC values were calculated only from years where ENSO was strong (not shown), then the skill of experiment IC even increased over the PNA and PSA region, but it decreased over the tropics and the Antarctica. On the other hand, when considering only neutral to weak ENSO years (not shown), then the AC scores of experiment IC over the PNA region decreased, and those over the tropics became more evenly distributed. During February and March (not shown), the simulation skill of IC dropped to insignificant values for most areas, whereas the patterns of predictability for iBC and BC became increasingly similar to that of ICBC. The only exemption was the region over the Antarctic continent, where the initial condition influence was still detectable during March.

During July (Fig. 2), predictability was generally lower than during January. This was probably related to the relative weakness of the ENSO signal during this time of the year. The only exception was the PSA region over the south Pacific with higher skill during July than during January. This might be due to seasonal changes in the atmospheric background state, which led to a stronger ENSO teleconnection response over the PSA region during southern hemisphere winter. The reduction in skill for iBC and BC, and some areas of skill for IC, showed that initial conditions were also important during July. An important difference from January was that the Antarctic region was much less predictable during July.

b. Zonal averages

Zonally averaged ACs were calculated to examine better the latitudinal differences in predictability. Since strong boundary effects during ENSO years are likely to override the more subtle initial condition effects, we analyzed the data separately for strong and neutral-to-weak ENSO years. The classification of ENSO years was taken from the Climate Prediction Center at NCEP (published on the internet), which provides a season-by-season breakdown of the SST conditions of the tropical Pacific. The chosen ENSO years, which include both warm and cold events, are presented in Table 2; neutral-to-weak ENSO years are all the other years.

Fig. 3 shows zonally averaged ACs for January (top) and July (bottom). During January, most notably, experiment ICBC showed at most latitudes the highest forecast skill. During neutral-to-weak ENSO years (left panel), there existed large differences in skill over the tropics, with simulation ICBC clearly the best, iBC better than BC, and IC having surprisingly good skill. Over the northern extratropics, overall skill values were low, but again ICBC had the best skill. South of 30?S, AC scores of IC were very close to that of ICBC, whereas iBC and BC had almost no forecast skill. During strong ENSO years (right panel), initial conditions were, as expected, less important. They were most relevant south of 60?S. Over the tropics and the northern extratropics, the skill scores of ICBC, iBC and BC were very close, indicating that boundary forcing dominated predictability during those years. The bottom of Fig. 3 shows the zonally averaged ACs for July. As for January, experiment ICBC had the highest skill values at most latitudes. The differences between ICBC and iBC/BC indicated that initial conditions had long term effects also in summer. These differences were largest during weak ENSO years (left panel) over the tropics. Over the extratropics, overall skill values were too small to be significant.

To determine whether initial condition were still important on a seasonal time scale, we repeated the above analysis using January, February and March (JFM) seasonal mean 500 hPa heights (not shown). As one might expect from the decaying influence of initial conditions with lead time, the effect of initial conditions was much weaker for seasonal means. Only over the Antarctic continent were initial condition effects important, and over the tropics during weak ENSO years, initial conditions played some role.

c. Seasonality

We also examined experimental differences between zonally averaged ACs for the two contrasting seasons (Fig. 4). The difference BC-ICBC (left panel) is a measure of how much skill was lost by having imperfect initial conditions. At all latitudes and during both seasons, the differences were negative, indicating that initial conditions were important. The reduction in skill was for most latitudes around ?0.1, with some regional and seasonal differences. One important exception was the Southern Hemisphere during January with initial conditions clearly dominating. There was also a tendency of initial conditions over the Northern Hemisphere to be less important, and over the tropics initial conditions were somewhat more important during January than during July.

The right panel of Fig. 4 shows the difference in AC between experiment BC and IC. Both curves are mostly positive, showing that the overall contribution of boundary forcing to predictability was larger than that of initial conditions. Again, the curves for winter and summer are quite similar. Note that the small latitudinal variations of the curves for the two seasons are mostly out of phase. This is true for BC-ICBC as well as for BC-IC. Arrows in both diagrams indicate two relative maxima at 30-40?N (January) and 30-40?S (July), which are accompanied with relative minima during the opposite month. Comparing with Figs. 1 and 2 shows that these latitudes were related to the PNA and PSA teleconnection regions. This indicates that the predictability of these patterns is dominated by boundary forcing during winter of the respective hemisphere. This was the time when the jets were well developed, and the resulting atmospheric mean state favored a strong teleconnection response from the tropical ENSO region (e.g., Trenberth et al., 1998). During summer, just the opposite was true, and initial conditions were quite important for the predictability of these patterns. This may be related to the long time scale of these modes at times when eddy kinetic energy was small, and the ENSO teleconnection response was weak.

d. Real world

To determine whether the long-term effects of initial conditions on monthly predictability were real or artifact, we replaced ICBC and IC by ICBC-r and IC-r and verified against reanalysis instead of model data. In this case, model errors and the high sampling uncertainty complicated the analysis. Fig. 5 shows the zonally averaged results for January. During neutral-to-weak ENSO conditions (left panel), initial condition effects were clearly important over the tropics. Over the extratropics, however, skill values are too small. During strong ENSO years (right panel), boundary forcing tended to completely dominate forecast skill over the tropics and northern extratropics. Remarkably, over the southern hemisphere, initial conditions were similarly important as in the perfect model world, as was shown by the strong increase in temporal ACs of ICBC and IC over the Antarctic continent. Overall, the main findings from the perfect model still hold when using observational data, although there was the difficulty of having only one verifying realization.

4. Vertical structure of predictability

The previous section was focused on differences in predictability at one level (500 hPa) and one lead time interval (week 3-6) or seasonal means. Since other levels may have important differences in their sensitivity to initial conditions and boundary forcing, we investigated the predictability of zonal mean heights for the whole atmospheric column from the surface to the lower stratosphere and for three different lead time intervals. Instead of zonal mean height predictability, we could have analyzed zonal averages of total predictability. We chose to use zonal mean heights, since they were simpler to analyze, and since they had a somewhat better signal to noise ratio than zonal averages of total height predictability. As in the previous section, predictability was measured by the temporal anomaly correlation of monthly mean heights.

a. Perfect world winter

Fig. 6 presents latitude-height cross sections of perfect model predictability during January (week 3-6), February (week 7-10) and March (week 11-14) for the four experiments. The predictability fields for experiment ICBC and IC are displayed in terms of their absolute correlation values. The fields for iBC and BC, however, show differences to ICBC. We chose this particular format to make the month to month changes in predictability clearer. The black curve shows the approximate location of the tropopause, which was calculated from the thermal tropopause criterion (e.g., Reichler et al., 1996; Hoinka, 1998) using monthly mean temperature fields from experiment ICBC.

During January, experiment ICBC (Fig. 6, top left panel) had positive AC scores everywhere, except over the lower Arctic. Predictability was generally higher over the southern than over the northern hemisphere, and it increased from lower to higher levels. Maxima (AC>0.9) were found in the tropical troposphere, where the atmosphere was most sensitive to boundary forcing. Further, predictability was large (AC>0.7) in the lower stratosphere, especially over the tropics and the southern hemisphere. The region over Antarctica had an equivalent barotropic structure with increasing predictability upwards into the lower stratosphere. The troposphere at 30-40?N was also well predictable. This represented the vertical signature of the PNA pattern, which can be seen by comparing with Fig. 1. The signature of the less well predictable PSA pattern, the southern hemispheric counterpart to the PNA, can be seen at around 40-50?S. During February and March, AC scores for ICBC generally decreased. The decrease affected most of the atmosphere, but it was most notable in the lower stratosphere. Some of these monthly changes in predictability might be explained by seasonalities in the strength of boundary forcing and in the atmospheric mean state, but most of them were related to the loss of initial condition memory. This became clear from the full predictability fields of experiment iBC and BC (not shown), which were quite constant during different months.

The results for experiment iBC (2^nd row) and BC (3^rd row) are shown as differences to ICBC. They exhibited at all lead times lower predictability than ICBC. The differences were strongest during January, and weakened during the subsequent months. The loss in skill from not having perfect initial conditions was most noticeable in the lower stratosphere and over Antarctica, but during January it also affected the tropics and the northern hemisphere. During January, experiment BC had at all levels higher deficits in skill than iBC, owing to the adjustment of BC to observed boundary conditions. During the other two months, the differences in skill became smaller (-0.1 - -0.3 correlations), but the structures were spatially very coherent, indicating that this was more than noise. It was particularly surprising that these differences were even detectable during March. Remember, iBC started from AMIP-type initial conditions that were fully adjusted to the boundary forcing, indicating that even minor differences in initial conditions can impact predictability out to many weeks.

During January, simulation IC (bottom panels) exhibited significant predictability over the southern hemisphere and in the lower stratosphere. The skill of IC over Antarctica was almost identical to that of ICBC. Further areas of significant skill from initial conditions were the PSA region, the lower troposphere over the tropics, and the PNA region. During February, IC still had some skill over Antarctica, but during March predictability from initial conditions alone was practically zero. It is interesting that the negative predictability patterns displayed by iBC and BC were strikingly similar to that of IC at all three lead times. This means that to first-order the loss in skill from not having perfect initial conditions was similar in strength and location to the skill of just having perfect initial conditions. In other words, long-term effects of initial and boundary conditions on atmospheric predictability are to a large degree linear. This is plausible, since non-linear effects grow usually fast and their time scale may be too short to be relevant for sub-seasonal forecasts.

Both experiments ICBC and IC, which started from perfect initial conditions, had in the lower stratosphere and over Antarctica relatively high predictability. The opposite was true for the iBC and BC which came from non-perfect initial conditions. This means that the stratosphere and the Antarctic regions respond very slowly to external forcing from the boundaries, and that the initial condition memory is very long. The predictability patterns even suggest that the Antarctic region may be influenced from the stratosphere above. This picture is consistent with recent observational findings (e.g., Robock, 2001; Thompson et al., 2002; Baldwin and Dunkerton, 2001; Baldwin and Dunkerton, 1999) that stratospheric anomalies affect tropospheric circulation, and that these anomalies can be used as predictors for tropospheric weather regimes with lead times of several weeks.

b. Perfect world summer

Fig. 7 presents the zonal mean predictability during boreal summer. In this case, the reference experiment ICBC contained only 10 members, so that the results were somewhat nosier than for winter. In general, the patterns of predictability were more symmetric to the equator than during winter. During July, predictability for ICBC was highest over the tropics and in the stratosphere and was lowest over Arctic and Antarctic regions. The separation into different well predictable areas as during winter was less pronounced. The vertical signatures of the PNA and PSA patterns can still be seen at around 40?N and 30?S, respectively. The decrease in skill for ICBC during August and September was again related to the loss of initial condition memory, as the comparison with experiment iBC shows.? Experiments iBC (2^nd row) and BC (3^rd row) exhibited during summer a stronger loss in predictability than during winter. During July, most of the atmosphere showed sensitivity to initial conditions, but during the later months, only the stratosphere and the region over the Antarctica were affected. Again, the loss for BC was larger than that for iBC. Experiment IC (bottom row) shows that initial condition effects during summer were most important in the stratosphere. The most important difference to winter was the high predictable area for ICBC and IC over Antarctica during winter, which had no counterpart during summer over the Arctic. During July, other areas with significant skill were the PNA region, the lower tropics, and the PSA region. The initial condition effect on the PNA region was larger than that on the PSA region, which was opposite to the winter season (see Fig. 6). This finding was consistent with that shown in Fig. 4. It confirms that initial conditions were more important for the ENSO teleconnection regions during summer than during winter of each hemisphere. During August, the initial condition effect was confined to small regions of the lower stratosphere, and during September it almost completely disappeared. Note that the patterns of IC were again quite similar to that of iBC.

c. Seasonality

Next, we explore seasonal differences in the magnitude and spatial characteristics of predictability. It is well known that the strength of the tropical ENSO signal peaks during late winter/early spring, but from this alone it is not obvious how the magnitude of predictability between winter and summer compares. Predictability depends not only on the seasonal cycle of ENSO, but also on the atmospheric mean state and the magnitude of atmospheric internal variability. Fig. 8 shows the seasonal differences in AC scores for experiment ICBC between winter and summer. Taking the atmosphere as whole, predictability during winter was clearly higher than during summer. This result contradicts recent findings from Kumar et al. (Kumar et al., 2002). This was true in particular for the tropics and the Antarctic regions, owing to the stronger ENSO signal and the strong initial condition effect during winter. For the midlatitudes of both hemispheres, the situation was more complicated, and a strong function of latitude and height.

d. Real world

Fig. 9 shows vertical cross sections of winter zonal mean predictability for experiment ICBC-r and IC-r verified against reanalysi s. The patterns were more noisy since only one member was used for verification. Nevertheless, similar conclusions found previously for the perfect model world also hold under the real world verification. During January, experiment ICBC-r showed high predictability in the tropics, in the stratosphere, and over the Antarctic continent. During February and March, the predictability for areas other than the tropics gradually reduced. Predictability patterns for IC-r were similar to ICBC-r except for the tropics. The loss in skill for experiments iBC and BC was similar to that in the perfect model world. Differences in skill to ICBC were particularly large during January, in the stratosphere, and over Antarctica.

5. Persistence and initial condition memory

In this section we explore further the long memory of initial conditions seen in the previous sections. We were particularly interested in determining whether this effect was due to dynamical predictability in the sense that daily variations in height were forecast well, or whether this was rather due to the strong atmospheric persistence over particular regions.

a. Persistence of daily heights

Atmospheric persistence can arise from persistent forcing by anomalous boundary conditions or from low-frequency variability within the atmosphere itself. Some of the many factors which may be important for persistence are trends, persistent boundary forcing, baroclinicity and advection (Trenberth, 1985b). Baroclinicity causes development of new disturbances, and advection moves disturbances over a given location. Both effects lead to a faster decay of persistence. Consequently, regions where baroclinicity and wind speeds are small are likely to be more persistent.

There are many ways to measure persistence. We use as measure the ratio between the low- and high-frequency variability of daily 500 hPa heights. This method is similar to the concept of signal to noise. First, daily climatological means were calculated at each grid point by averaging over all R ensemble members and all Y years of one experiment, i.e.,

,?????????????????????????????????????????????????????????????????? (5)

where represents the height of the 500 hPa surface at one grid point for the ith day and yth year of ensemble member r. Anomalies z? were calculated by removing the climatological mean from the raw data, i.e.,

.??????????????????????????????????????????????????????????????????????? (6)

Note that this method does not remove the interannual variability related to persistent boundary forcing (e.g. ENSO) and therefore increases persistence. The low-frequency component (or signal) during a particular year y and for a particular member r was given by the mean anomaly over the T=48 day long period from Dec. 15^th to Jan. 31^st, i.e.,

.?????????????????????????????????????????????????????????? (7)

The high-frequency part (or noise) was derived from the daily standard deviation around this mean over the same period by taking

.?????????????????????????????????????????????????? (8)

The ratio

?????????????????????????????????????????????????????????????????????????? (9)

gave the measure of persistence for one member and year. The average over all Y and R was taken as the final results, i.e.,

?????????????????????????????????????????????????????????????????? (10)

This quantity is shown in Fig. 10 (a) for reanalysis, and for experiments ICBC and IC. In the reanalysis, persistence was high over the tropics, in particular over the Indian Ocean. This reflected mostly the effects of persistent tropical forcing on the atmosphere. Another area of strong persistence was found over the Antarctica. There, the high-frequency variability was particularly small (not shown), which is probably linked to the low baroclinicity and the low wind speeds typical for Antarctica during January. This result was also consistent with Trenberth (1985b). Further regions of enhanced persistence can be found over the North Atlantic, Greenland, the North Pacific off the coast of the U.S., and the Arctic. When persistence was calculated from simulation ICBC, very similar patterns emerged (middle panel). Simulation IC, on the other hand, had important differences with the reanalysis. This allowed to determine regions where atmospheric persistence was caused by persistent boundary forcing, and where it was related to the slow time scale of the atmosphere itself. The main differences were over the tropics and the two ENSO teleconnection regions PNA and PSA, where persistence is related to persistent boundary forcing. IC had persistence over the Indian Ocean and over Antarctica, indicating that the long initial condition memory over these regions was related to enhanced atmospheric persistence.

To verify our results, a second form of persistence measure was derived from the autocorrelation between daily 500 hPa heights. For each ensemble member r and year y, the autocovariance at lag L was computed over the I=107 days of the winter season, i.e.,

.????????????????????????????????????????????? (11)

The mean autocovariance over all years and members was calculated by

.????????????????????????????????????????????????????? (12)

Then, the autocorrelation at lag L was given by

.????????????????????????????????????????????????????????????????????????????? (13)

In Fig. 10 (b), the lag L for the autocorrelation to reach 1/e is shown. This measure of persistence provided similar patterns as before. For reanalysis and ICBC, largest values of L (>20 days) can be seen over the tropics. For simulation IC, L was relatively short over the tropics and the North Pacific, and it was long over Antarctica.

b. Relationship to major modes

Atmospheric persistence in the extratropics was also closely related to the major modes of variability. This is demonstrated by Fig. 11, which shows a composite of the leading modes of variability calculated from individual members of experiment ICBC. The modes were derived from monthly anomalies using Empirical Orthogonal Function (EOF) analysis. The analysis was carried out separately for the tropics (30N-30S), and the northern (90N-20N) and southern (20S-90S) extratropics. The first three major modes over the north were derived by rotated EOF analysis technique (e.g., von Storch and Zwiers, 1999). Rotating the EOFs led to physically more meaningful patterns. These well known modes were the North Atlantic Oscillation (NAO) (Walker and Bliss, 1932), the PNA pattern, and a pattern similar to the Northern Asian pattern (NA) (Barnston and Livezey, 1987). They explained 12.5%, 11.6%, and 11.0% of the total variance, respectively. For the southern hemisphere, unrotated EOF analysis was used, and only the first mode is shown. This is the Antarctic Oscillation (AAO) (Gong and Wang, 1999), which explains 32% of the total variance. Over the tropics, unrotated EOFs were also used. The tropical mode represents the Southern Oscillation (SO) (Walker, 1924), which explains 32% of the total variance.

All of the above modes were characterized by enhanced persistence, as can be seen by comparing them with the persistence maps for ICBC shown before (e.g., Fig. 10). This relationship was, of course, not a coincidence, since these modes represent large-scale and low-frequency patterns of variability. EOF analysis from simulation IC (not shown) did not reveal the PNA mode. This means that most of the persistence or low-frequency variability associated with the PNA mode was caused by remote SST forcing, which is consistent with IC not exhibiting much persistence over the North Pacific. Simulation IC exhibited a weak SO mode (not shown) as the third tropical EOF mode. It explained only about 10% of the total variance. This mode was again very similar to the patterns of persistence and predictability exhibited by simulation IC over the tropics. The AAO, on the other hand, was a very robust feature. It was found in all simulations and existed independent of SST forcing. This might explain why large persistence and high initial condition predictability were found over Antarctica.

c. Antarctic Oscillation

The relationship between initial condition memory, persistence and the AAO mode of low-frequency variability is now further described. Fig. 12 shows the pattern of the AAO, derived from 500 hPa height fields from January monthly and ensemble means of simulation ICBC. Only the first 10 members of ICBC (=ICBC-A) were used to construct the EOF. Since ensemble means were used, the EOF explained almost 60% of the total variance. As pointed out by Thompson and Wallace (2000), the AAO is characterized by an equivalent barotropic, zonally symmetric structure. A very similar pattern in terms of shape and strength emerged if reanalysis or any other experiments were used (not shown). The zonally symmetric structure, with one sign over the pole and the opposite sign at mid-latitudes, indicates that this mode involves exchange of mass between the high and mid-latitudes.

To determine how important initial conditions were for the simulation of the AAO, we examined how well the EOF was reproduced by the other experiments. To this end, the reference EOF from ICBC-A was projected on the data from the other experiments. The resulting time series were a measure of how well the AAO was? simulated by the other simulations (left panels of Fig. 13). The year-to-year fluctuations of the original AAO are shown in the top panel. This time series was well reproduced by ICBC-B, which were the other 10 members of simulation ICBC. The good relationship is indicated by the rather large correlation shown on the right. Remarkably, the time series from simulation IC correlated also quite well with ICBC-A. On the other hand, the time series for iBC and BC correlate less well with ICBC-A. This confirms our previous finding that the predictability of the AAO during January was dominated by initial conditions. The relationship between AAO and initial conditions was also determined from reanalysis and simulations ICBC-r, IC-r, iBC and BC (Fig. 13, right panels). The correlations between the time series were now smaller, which may not be surprising since model errors affect the forecast. However, a similar relationship as before held in terms of the different sizes of the correlations. Again, IC-r was as good as ICBC-r, and iBC and BC showed much lower correlations. It should be noted, however, that the relatively good correlations between reanalysis, ICBC-r and IC-r were to some extent due to a common long-term trend exhibited by all three time series.

As seen before, the AAO was very persistent during January. This raised the question as to whether there existed a relationship between strength and persistence and hence predictability of the AAO. Such a relationship would make sense, since a strong mode involves a large exchange of mass between polar and mid-latitudes, which may take long to equilibrate. That this was indeed the case is demonstrated by Fig. 14, which shows for particular years of the 1979-2000 period the relationship between January mean skill over the southern extratropics (20?S-90?S) and the strength of the AAO index. The strength of the AAO was measured by the value of the index time series for experiment ICBC at January 1^st. Skill for particular years was determined from the spatial anomaly correlation over the southern extratropics between experiment and reference simulation ICBC. As with the calculation of the temporal ACs, the ensemble mean of the experiment was verified against individual members of ICBC, and the mean over all members was taken as final result.

Experiments ICBC and IC showed a good relationship between magnitude of the index and the corresponding forecast skill. The correlation coefficients were 0.53 and 0.75, respectively. Thus, simply from knowing the strength of the AAO at the beginning of a forecast one can predict how skillful a forecast will be. For experiments iBC and BC, this relationship did not hold, since no good initial condition information were provided. Interestingly, ENSO also had some influence on the polarity of the AAO, presumably through the teleconnection into the PSA region. From the limited number of samples it seems that a positive (negative) AAO phase was more likely during warm (cold) years.

6. Summary and Discussion

This study investigated the spatial and temporal structure of predictability with a complex numerical model. The main goal was to find out how important initial and boundary conditions were for long-range predictability. To this end, four predictability experiments with different types of initial and boundary conditions were designed. Predictability was measured from the temporal anomaly correlation, and by verifying either against model data or against reanalysis.

Boundary forced predictability was high over the tropics. There, it affected the whole atmospheric column from the surface to the lower stratosphere. Boundary forcing was also essential for predictability of the PNA region during northern winter, and the PSA region during northern summer. At the week 3-6 forecast range, initial conditions were remarkably important for predictability. This could be detected almost everywhere and in both seasons, but it was strongest during January, during weak ENSO years, and over the Indian Ocean, the PSA region, Antarctica, and in the lower stratosphere. Both perfect and real world verification showed that predictability over Antarctica during southern hemispheric summer was elevated, and other modeling (e.g., Kumar et al., 2002) and observational (e.g., Trenberth, 1985a) studies showed similar results. Therefore, we are convinced that this effect is real. During July, the effects of initial conditions were somewhat weaker. For the PNA and PSA region, initial conditions were more important during summer of the respective hemisphere than during winter.

The initial condition effect on predictability was intimately related to atmospheric persistence. Persistence was large over the Indian Ocean and over Antarctica, exactly where the impact of initial conditions on predictability was strong. Over Antarctica, persistence was in turn connected to the AAO, the leading mode of low-frequency variability over the southern hemisphere. Only with good initial conditions was the time series of the AAO index well predicted, whereas boundary forcing alone was unable to produce such results. It was also found that forecast skill over the southern extratropics was elevated when the AAO was either in its strong negative or strong positive phase, which could be useful for predicting the expected forecast skill. Most of these results were also found when reanalysis was used to verify the experiments. At lead times of more than 6 weeks, the initial condition effect was small, but nevertheless still detectable.

The importance of initial conditions for tropical forecasts was somewhat surprising, since it is generally believed that the tropics are dominated by boundary forcing. Reichler and Roads (2003b) showed that initial conditions are important for the MJO, although in this study most of the MJO related variability was removed by taking monthly averages. There were other processes that might be responsible for the long initial condition memory. For example, atmospheric persistence, which was found to be quite high over the Indian Ocean region, is a likely possibility.

There are two caveats to this study. First, the adequacy of reanalysis over Antarctica and for the lower stratosphere is doubtful. There, observations are sparse, and artificialities of the assimilation model may influence the reanalysis. However, geopotential heights are smoothly varying fields, so that good results may be achieved with relatively few observations. This assumption was confirmed by comparing reanalysis with radiosonde data, which showed that reanalysis from 1979 and on are an excellent proxy for real heights over the southern hemisphere (pers. comm. D. Thompson, 2002). Second, it is unclear how well the model performs in the stratosphere. The model was not specifically designed for this domain, but it has an adequate vertical resolution in the stratosphere (11 layers above 200 hPa), and modern radiation schemes capable of simulating the effects of radiatively active gases on the temperature structure of the stratosphere. Preliminary analysis showed that basic stratospheric features were simulated well by the model.

We also found that initial conditions strongly impacted the lower stratosphere. Presumably, this was related to the small spatial complexity and to the large persistence of the stratosphere (Perlwitz and Graf, 2001). The stratosphere of the model was very slow in responding to boundary forcing from the tropics. This was demonstrated by the month-to-month evolution of predictability from boundary forcing alone. During January, the signal rapidly filled the entire equatorial troposphere, but the lower stratosphere and the southern hemispheric extratropics showed little response. Only during the following months was the boundary forced signal able to propagate up into the lower stratosphere. This delayed response was likely related to the stable stratification of the stratosphere, which causes vertical transports through the tropopause and into stratosphere to be extremely slow (e.g., Holton et al., 1995). To the extent that the model dynamics are realistic, the signal may have reached the stratosphere by means of vertically propagating equatorial wave modes. It has been observed that these modes are excited by large scale convective heating anomalies in the equatorial troposphere (e.g., Andrews et al., 1987).

The region of high predictability over Antarctica showed an equivalent barotropic structure with increasing skill upward and maxima in the lower stratosphere. This raised the question whether tropospheric and stratospheric circulations are connected. Indeed, Thompson and Wallace (2000) found from observations that during certain ?active seasons?, the tropospheric circulation over the southern hemisphere is coupled to the stratosphere, and that anomalies amplify with height upward into the stratosphere. The ?active season? is a 6-8 week interval centered in November, which may still include the month of January. During this time, the polar vortex is in the process of breaking down, thus allowing strong interaction of vertically propagating planetary waves from the troposphere with the flow of the lower stratosphere (e.g., Charney and Drazin, 1961). This mechanism may help to explain why troposphere and stratosphere over Antarctica had a similar predictability structure.

During July, we could not find an equivalent effect over the Arctic. Probably, this was related to the sharply contrasting land-sea distribution and stationary wave climatologies of the two hemispheres. The circulation over the southern hemisphere is more zonally symmetric, with fewer baroclinic disturbances and less meridional exchange than over the northern hemisphere. Another important difference is that the Arctic is not snow covered during summer, so that strong convective disturbances can develop. The permanent snow and ice cover over Antarctica, on the other hand, might provide very constant boundary conditions with little or no anomalies. And lastly, July is not an active season for strong troposphere-stratosphere coupling over the Arctic. All these factors together may help to understand why high initial condition predictability could only be found over Antarctica.

Another interesting phenomenon is the long-term trend of the AAO index time series, which was found in reanalysis, and which was well simulated by experiments with good initial conditions. It would be important to find out why this trend could be simulated successfully by the model. Thompson and Solomon (2002) argued that in the real atmosphere this trend was related to photochemical ozone loss in the stratosphere. This would cause circulation changes in the lower stratosphere, which in turn propagate downward into the troposphere. The simulations of this study were forced with climatological ozone conditions, so that the observed trend must have entered the simulations through their initial conditions. Several recent observational and modeling studies have shown evidence for a downward propagation of stratospheric anomalies into the troposphere (e.g., Shindell et al., 1999; Baldwin and Dunkerton, 2001; Polvani and Kushner, 2002; Black, 2002), and it has even been suggested that such a mechanism could improve predictability on longer time scales (e.g., Thompson et al., 2002; Robock, 2001). It is unclear at this point whether the model of this study was able to simulate this mechanism, and whether this contributed to the good predictability over Antarctica. Clearly, more work needs to be done to achieve more clarification.

Acknowledgements: We thank Masao Kanamitsu for many useful comments and for his help with the model. We would also like to thank Dietmar Dommenget for his careful review of the original manuscript and for insightful discussions. Funding for this research was provided by a cooperative agreement with NOAA (NA77RJ0435 and NA17R1231). The views expressed herein are those of the authors and do not reflect the views of NOAA. We thank the Maui High Performance Computing Center (MHPCC) and the San Diego Supercomputing Center (SDSC) for providing computing time for some of the experiments. This work is part of the PhD thesis of T. Reichler.

References

Andrews, D. G., J. R. Holton, and C. B. Leovy, 1987: Middle Atmosphere Dynamics. Academic Press, Orlando, Florida.

Bacmeister, J. T., P. J. Pegion, S. D. Schubert, and M. J. Suarez, 2000: An atlas of seasonal means simulated by the NSIPP 1 atmospheric GCM. NASA Tech Memo No 104606, 17, Goddard Space Flight Center, Greenbelt, MD 20771.

Baldwin, M. P. and T. J. Dunkerton, 1999: Propagation of the Arctic Oscillation from the stratosphere to the troposphere. J Geophys Res-Atmos, 104, 30937-30946.

Baldwin, M. P. and T. J. Dunkerton, 2001: Stratospheric harbingers of anomalous weather regimes. Science, 294, 581-584.

Barnston, A. G. and R. E. Livezey, 1987: Classification, Seasonality and Persistence of Low-Frequency Atmospheric Circulation Patterns. Mon Weather Rev, 115, 1083-1126.

Black, R. X., 2002: Stratospheric forcing of surface climate in the Arctic oscillation. J Climate, 15, 268-277.

Charney, J. G. and P. G. Drazin, 1961: Propagation of planetary-scale disturbances from the lower into the upper atmosphere. J Geophys Res, 66, 83-109.

Folland, C. K. and D. E. Parker, 1995: Correction of Instrumental Biases in Historical Sea Surface Temperature Data. Quart J Roy Meteorol Soc, 121, 319-367.

Gong, D. Y. and S. W. Wang, 1999: Definition of Antarctic Oscillation Index. Geophys Res Lett, 26, 459-462.

Hoinka, K. P., 1998: Statistics of the global tropopause pressure. Mon Weather Rev, 126, 3303-3325.

Holton, J. R., P. H. Haynes, M. E. McIntyre, A. R. Douglass, R. B. Rood, and L. Pfister, 1995: Stratosphere-Troposphere Exchange. Rev Geophys, 33, 403-439.

Horel, J. D. and J. M. Wallace, 1981: Planetary-Scale Atmospheric Phenomena Associated with the Southern Oscillation. Mon Weather Rev, 109, 813-829.

Kalnay, E., et al., 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull Amer Meteorol Soc, 77, 437-471.

Kanamitsu, M., W. Ebisuzaki, J. Woolen, S.-K. Yang, J. Hnilo, M. Fiorino, and G. L. Potter, 2002a: NCEP/DOE AMIP-II reanalysis (R-2). Bull Amer Meteorol Soc, 83, 1631-1643.

Kanamitsu, M., et al., 2002b: NCEP dynamical seasonal forecast system 2000. Bull Amer Meteorol Soc, 83, 1019-1037.

Karoly, D. J., 1989: Southern Hemisphere Circulation Features Associated With El Nino - Southern Oscillation Events. J Climate, 2, 1239-1252.

Kistler, R., et al., 2001: The NCEP-NCAR 50-year reanalysis: Monthly means CD-ROM and documentation. Bull Amer Meteorol Soc, 82, 247-267.

Kumar, A., S. Schubert, and M. Suarez, 2002: Variability and predictability of 200-mb seasonal mean heights during summer and winter. J Geophys Res, (in press).

Lorenz, E., 1963: Deterministic non-periodic flow. J Atmos Sci, 20, 130-141.

Lorenz, E., 1969: The predictability of a flow which possesses many scales of motion. Tellus, 21, 289-307.

Lorenz, E., 1982: Atmospheric predictability experiments with a large numerical model. Tellus, 34, 505-513.

Mason, S. J., L. Goddard, N. E. Graham, E. Yulaeva, L. Q. Sun, and P. A. Arkin, 1999: The IRI seasonal climate prediction system and the 1997/98 El Nino event. Bull Amer Meteorol Soc, 80, 1853-1873.

Palmer, T. N., 1993: Extended-Range Atmospheric Prediction and the Lorenz Model. Bull Amer Meteorol Soc, 74, 49-65.

Peng, P. T., A. Kumar, A. G. Barnston, and L. Goddard, 2000: Simulation skills of the SST-forced global climate variability of the NCEP-MRF9 and the Scripps-MPI ECHAM3 models. J Climate, 13, 3657-3679.

Perlwitz, J. and H. F. Graf, 2001: The variability of the horizontal circulation in the troposphere and stratosphere - a comparison. Theor Appl Climatol, 69, 149-161.

Polvani, L. M. and P. J. Kushner, 2002: Tropospheric response to stratospheric perturbations in a relatively simple general circulation model. Geophys Res Lett, 29.

Reichler, T. and J. O. Roads, 2003a: The role of boundary and initial conditions for dynamical seasonal predictability. Nonlinear Proc Geophys, (in press).

Reichler, T. and J. O. Roads, 2003b: Long range predictability in the tropics. Part II: 30-60 day variability. J Climate, (submitted).

Reichler, T., M. Dameris, R. Sausen, and D. Nodorp, 1996: A global climatology of the tropopause height based on ECMWF-analyses. Report 57, DLR, Institute of Atmospheric Physics, 23 pp.

Reynolds, R. W. and T. M. Smith, 1994: Improved Global Sea Surface Temperature Analyses Using Optimum Interpolation. J Climate, 7, 929-948.

Roads, J. O., 1988: Lagged average predictions in a predictability experiment. J Atmos Sci, 45, 147-162.

Roads, J. O., S. C. Chen, and F. Fujioka, 2001: ECPC's weekly to seasonal global forecasts. Bull Amer Meteorol Soc, 82, 639-658.

Robock, A., 2001: Stratospheric forcing needed for dynamical seasonal prediction. Bull Amer Meteorol Soc, 82, 2189-2192.

Shindell, D. T., R. L. Miller, G. A. Schmidt, and L. Pandolfo, 1999: Simulation of recent northern winter climate trends by greenhouse-gas forcing. Nature, 399, 452-455.

Shukla, J., et al., 2000: Dynamical seasonal prediction. Bull Amer Meteorol Soc, 81, 2593-2606.

Thompson, D. W. J. and J. M. Wallace, 2000: Annular modes in the extratropical circulation. Part I: Month-to-month variability. J Climate, 13, 1000-1016.

Thompson, D. W. J. and S. Solomon, 2002: Interpretation of recent Southern Hemisphere climate change. Science, 296, 895-899.

Thompson, D. W. J., M. P. Baldwin, and J. M. Wallace, 2002: Stratospheric connection to Northern Hemisphere wintertime weather: Implications for prediction. J Climate, 15, 1421-1428.

Toth, Z. and E. Kalnay, 1993: Ensemble Forecasting at NMC - the Generation of Perturbations. Bull Amer Meteorol Soc, 74, 2317-2330.

Toth, Z. and E. Kalnay, 1997: Ensemble forecasting at NCEP and the breeding method. Mon Weather Rev, 125, 3297-3319.

Trenberth, K. E., 1985a: Potential predictability of daily geopotential heights over the Southern Hemisphere. Mon Weather Rev, 113, 54-64.

Trenberth, K. E., 1985b: Persistence of daily geopotential heights over the Southern Hemisphere. Mon Weather Rev, 113, 38-53.

Trenberth, K. E., G. W. Branstator, D. Karoly, A. Kumar, N. C. Lau, and C. Ropelewski, 1998: Progress during TOGA in understanding and modeling global teleconnections associated with tropical sea surface temperatures. J Geophys Res-Oceans, 103, 14291-14324.

von Storch, H. and F. W. Zwiers, 1999: Statistical Analysis in Climate Research. Vol. 12, Cambridge University Press, Cambridge.

Walker, G. T., 1924: World Weather II. Mem India Meteor Dept, 24, 275-332.

Walker, G. T. and E. W. Bliss, 1932: World Weather V. Mem Roy Meteor Soc, 4, 53-84.

Wallace, J. M. and D. S. Gutzler, 1981: Teleconnections in the Geopotential Height Field During the Northern Hemisphere Winter. Mon Weather Rev, 109, 784-812.

Webster, P. J. and H. R. Chang, 1988: Equatorial Energy Accumulation and Emanation Regions - Impacts of a Zonally Varying Basic State. J Atmos Sci, 45, 803-829.

Table and Figure Captions

Table 1: Boundary and initial conditions, ensemble size and simulation period for each experiment. ?r-2? means NCEP/DOE reanalysis-2. Winter refers to Dec. 15^th ? Mar. 31^st of the following year, and summer refers to Jun. 15^th ? Sep 30^th. ?rndm.? indicates randomly chosen initial conditions, ?obs.? means observed (i.e. reanalysis-1), ?cont.? indicates continuous base run over all years, and ?clim.? indicates climatological boundary conditions.

Table 2: Classification of strong ENSO years during January and during July.

Fig. 1. Temporal anomaly correlations of January mean 500 hPa heights assuming perfect model. Left panels are absolute AC scores, and right panels are Fisher Z transformed differences to ICBC.

Fig. 2. As Fig. 1 but for July.

Fig. 3. Zonally averaged ACs from January (top) and July (bottom) mean 500 hPa heights for neutral-to-weak ENSO years (left) and strong ENSO years (right).

Fig. 4. Difference (Fisher Z transformed) of zonally averaged ACs from monthly mean 500 hPa heights from all years under perfect model assumption.

Fig. 5. As Fig. 3 but for ICBC-r and IC-r and verification against reanalysis.

Fig. 6. Vertical cross sections of temporal ACs of monthly mean zonal mean heights under perfect model assumption. For experiment ICBC (top row) and IC (bottom row), absolute AC values are shown. For iBC (2^nd row) and BC (3^rd row), differences (Fisher Z transformed) to ICBC are shown. The black line indicates the location of the thermally defined tropopause. Vertical axis is pressure in hPa, and horizontal axis is latitude in degrees.

Fig. 7. As Fig. 6 but for boreal summer.

Fig. 8. Differences in seasonal predictability for ICBC between JFM and JAS.

Fig. 9. As Fig. 6 but for ICBC-r and IC-r and verified against reanalysis.

Fig. 10: (a) Ratio between mean and spread of daily 500 hPa anomaly fields of individual members from Dec. 15^th ? Jan. 31^st. Shown are averages over the 1979-2000 period and over all ensemble members. (b) Lag L in days for autocorrelation of daily wintertime 500 hPa height anomalies to reach e^-1. A 3x3 moving box average was applied.

Fig. 11. Composite view of the leading modes of extratropical variability of monthly mean 500 hPa height anomalies during winter (JFM), derived from experiment ICBC. The first 3 major modes (NAO, PNA, NA) are shown for the Northern Hemisphere, and the first leading mode (AAO and SO) are shown for the Southern Hemisphere and the tropics. Contour levels are ?40 and ?80 meters (?6 and ?10 in the tropics). Negative contour lines are dashed. The corresponding PC time series were normalized.

Fig. 12. Leading EOF (AAO) over the southern extratropics (20?S-90?S) derived from 500 hPa January monthly mean fields of the ensemble mean of the first 10 members of simulation ICBC. The corresponding time series was standardized, so that the pattern represents average anomalies. Units are geopotential meters.

Fig. 13: (left) AAO index time series derived from projecting reference EOF (shown in Fig. 12) from experiment ICBC-A on monthly mean data from the corresponding simulations. Numbers on the right denote correlation between time series of reference and experiment. (right) As (left), but for reference EOF from reanalysis, and for ICBC-r and IC-r instead of ICBC-B and IC.

Fig. 14. Relationship between January mean skill over the southern extratropics (20?S-90?S) and AAO index at January 1^st. The corresponding EOF was normalized, so that units of index (abscissa) are in meters. Ordinate denotes mean pattern correlation between ensemble mean of experiment and individual members of ICBC. Numbers in lower left corner denote correlation coefficient between skill and magnitude of index. Squares in red (blue) denote warm (cold) ENSO years.

Table and Figures

	boundary conditions		initial conditions
name	ocean	land	atmosphere	land	size	period	years
BASE-O	observed	model	obs. 1/1/48	obs. 1/1/48	1	cont.	?48-2000
BASE-C	clim.	model	obs. 1/1/48	obs. 1/1/48	1	cont.	?48-2024
ICBC	observed	model	BASE-O	BASE-O	20 10	winter summer	?79-2000
ICBC-r	observed	r-2	r-2	-	10 ?	winter	?79-2000
IC	clim.	r-2 clim.	BASE-O	-	10 10	winter summer	?79-2000
IC-r	clim.	r-2 clim.	r-2	-	10	winter	?79-2000
BC	observed	model	BASE-C rndm.	BASE-C rndm.	10 10	winter summer	?79-2000
iBC	observed	model	ICBC, 1 yr lag	ICBC, 1 yr lag	10 10	winter summer	?80-2001