A series of 46-day ensemble integrations starting on the 15th of each month from 1989 to 2008 has been completed with the European Centre for Medium-Range Weather Forecasts (ECMWF) forecast system. The Madden– Julian Oscillation (MJO) simulated by the hindcasts is diagnosed using an index based on combined empirical orthogonal functions (EOFs) of zonal winds at 200 and 850 hPa and outgoing long-wave radiation. Results indicate that the dynamical model is able to maintain the amplitude of the MJO during the 46 days of integrations and the model displays skill for up to about 20 days to predict the evolution of the MJO. However, the MJO simulated by the model has a too slow eastward propagation and has difficulties crossing the Maritime Continent.
The Madden– Julian Oscillation (MJO) is the main source of potential predictability in the Tropics on time-scales exceeding one week but less than a season (Madden and Julian, 1971). It has a large impact on the Indian (Murakami, 1976; Yasunari, 1979) and Australian (Hendon and Liebmann, 1990) monsoons. It plays an active role in the onset and development of an El-Niño event (e.g. Kessler and McPhaden, 1995) and has an impact on tropical cyclogenesis (Maloney and Hartmann, 2000a,b; Mo, 2000). It also impacts on the extratropical weather (Ferranti et al., 1990; Cassou, 2008). Therefore it is important for a monthly forecasting system to have skill not only in predicting the evolution of the MJO, but also in simulating the MJO teleconnections. The goal of the present paper is to evaluate the skill of a large set of 46 d hindcasts using a version of the ECMWF Integrated Forecast System (IFS) known as Cycle 32r3 (Cy32r3) (Bechtold et al., 2008) to simulate MJO events and its teleconnections in the Tropics and in the northern Extratropics.
A first goal of this paper is to assess the skill of the numerical model to predict MJO events. The skill of the ECMWF model has so far been assessed routinely for each new model cycle using a series of 5-member ensemble integrations of 32 d forecasts for each day from 15 December 1992 to 31 January 1993 (46 cases), during the Intensive Observing Period of the Tropical Ocean– Global Atmosphere Coupled Ocean– Atmosphere Response Experiment (TOGA COARE). This experimental set-up has the advantage of being cheap enough to be performed for each new version of the model, and to give a good indication of the MJO representation in each new version of the ECMWF model (example in Bechtold et al., 2008). However the serial experiments cover only one single MJO event. The skill of the model to predict this specific MJO event may not be representative of the general skill of the model. In the present study, a large number of model hindcasts covering a 20-year period (1989– 2008) is used to assess the skill of the ECMWF monthly forecasts. This should give a more reliable evaluation of the skill of the model than the serial experiments.
A second goal of this paper is to evaluate the impact of the MJO on the atmospheric processes in the Tropics (precipitation) and Extratropics (North Atlantic weather regimes and probabilistic skill scores). Such study would have been difficult before Cy32r3, since the model could not maintain the amplitude of an MJO event for more than a few days. Over the recent years, the representation of the MJO has improved dramatically, thanks mostly to changes in the model's physics introduced in Cy32r3 (Bechtold et al., 2008). Now IFS is able to maintain the amplitude of the MJO for more than 30 d, which makes it possible to evaluate the teleconnections associated with the MJO in the model integrations.
After this introduction, section 2 will describe the experimental set-up. The skill of the model to predict MJO events will be evaluated in section 3. The impact of the MJO on precipitation will be discussed in section 4. Section 5 will show the impact of the MJO on the North Atlantic weather regimes in the model simulations. The impact of the MJO on monthly forecast probabilistic skill scores will be evaluated in section 6. Finally, section 7 will summarise the main results of this paper.
2. Experimental set-up
A series of hindcasts has been performed for the 20-year period 1989 to 2008. The hindcasts start on the 15th of each month and are 46 d long in order to cover the full next calendar month. For each starting date, the hindcasts consist of an ensemble of 15 members: a control and 14 perturbed forecasts. The version of IFS used in this experiment is Cy32r3, which was operational from November 2007 until June 2008. As mentioned in the introduction, this version of IFS showed a clear improvement in the representation of the MJO over the previous versions. The configuration of the hindcast is the same as the one used in the operational monthly forecast at ECMWF (Vitart et al., 2008) except for the length of the forecasts (46 d instead of 32 d for operational monthly forecasts). In this configuration, IFS is first integrated for 10 d with a resolution of T399 (about 50 km resolution) and 62 vertical levels. At day 10 the horizontal resolution is lowered to T255 (about 80 km resolution) until the end of the forecast. During the first 10 d, IFS is forced by persisted SST anomalies. After day 10, IFS is fully coupled to the HOPE (Hamburg Ocean Primitive-Equation model) ocean GCM (Wolff et al., 1997). The frequency of coupling is 3 h.
The initial conditions are taken from the ECMWF re-analyses ERA-40 (Uppala et al., 2005) until 2001 and from ECMWF operational analysis after 2001. The 14 perturbed integrations use slightly different atmospheric and oceanic initial conditions which are designed to represent the uncertainties inherent in the operational analyses. The 14 atmospheric perturbations are produced using the singular vector method (Buizza and Palmer, 1995). These include perturbations in the Extratropics and perturbations in some tropical areas by targeting tropical cyclones (Puri et al., 2001). In addition, in order to take account of the effects of uncertainties in the model subgrid-scale parametrizations, the tendencies in the atmospheric physics are randomly perturbed during the model integrations (Buizza et al., 1999; Palmer, 2001). Different ocean initial conditions are produced by applying a set of wind stress perturbations during the ocean data assimilation (Vialard et al., 2005). More details about the monthly forecast configuration at ECMWF can be found in Vitart et al. (2008).
3. Main characteristics of simulated MJO events
The methodology for assessing the skill in predicting the MJO follows Gottschalk et al. (2009). The Wheeler and Hendon index (Wheeler and Hendon, 2004) has been applied to all the model hindcasts and to the 1989– 2008 reanalyses, ERA-Interim (Simmons et al., 2007), to evaluate the skill of the monthly forecasting system to predict MJO events and to produce composites for different phases of the MJO. The Wheeler and Hendon index is calculated by projecting the forecasts or analysis on the two dominant combined EOFs of outgoing long-wave radiation (OLR), and zonal wind at 200 and 850 hPa averaged between 15°N and 15°S. The index has been applied to daily anomalies relative to the 1989– 2008 climate instead of the absolute value of the field, in order to remove the impact of seasonal cycle. In addition, a 120 d running mean has been subtracted to remove the variability associated with El Niño Southern Oscillation (ENSO). The positive (negative) phase of EOF2 describes suppressed (enhanced) convection over the Indian Ocean and enhanced (suppressed) convection over the West Pacific. The positive (negative) phase of EOF1 describes enhanced (suppressed) convection over the Maritime Continent region. Analysis and forecasts can be projected onto those two EOFs to describe the phase of the MJO in terms of two time series, PC1 and PC2. The two time series can be plotted as a succession of points in the PC1– PC2 phase space, in such a way that the MJO is described by a clockwise propagation in the phase space. The PC1– PC2 phase space can be divided into eight sections representing a specific phase of the MJO (e.g. Figure 2 in Gottschalk et al., 2009). Phases 2 and 3 (negative EOF2) correspond to enhanced convection over the Indian Ocean, phases 4 and 5 (positive EOF1) correspond to the MJO over the Maritime Continent, phases 6 and 7 (positive EOF2) correspond to the MJO over the Western Pacific and phases 8 and 1 (negative EOF1) correspond to the active phase of the MJO in the Western Hemisphere.
To evaluate the skill of the monthly forecasting system to predict the MJO, a linear correlation is performed between the observed time series of PC1 and PC2 with the forecast time series at different lead times. We consider that the forecast is skilful when the anomaly correlation is higher than 0.6.
According to Figure 1, the model has skill to predict the evolution of the MJO up to about day 20. The anomaly correlation falls to 0.5 around day 23. PC1 and PC2 display similar scores. The RMS error of the ensemble mean reaches the RMS error obtained with climatology after day 20 for both PC1 and PC2 (Figure 1), indicating also that the model has useful skill up to day 20. The amplitude of the MJO simulated by the model starts to increase after day 5 and becomes too strong by about 20% after day 10 (Figure 2).
The monitoring of real-time forecasts from Cy32r3 indicates that the model often fails to propagate the MJO across the Maritime Continent, and that it is often too slow over the Indian Ocean. To check if this is the case in the set of 46 d ensemble hindcasts, a composite of all the MJO events in phase 2 (convection over the western Indian Ocean) has been produced and its daily evolution calculated for the analysis and the ensemble hindcasts. A Hovmöller diagram of the percentage of cases in a given phase of the MJO shows the eastward propagation of the MJO in both model and reanalysis, but the propagation is slower in the model than in the reanalysis (Figure 3). For instance, the majority of MJO events in the reanalysis reach phase 4 within 10 d after phase 2, instead of 14 d in the model simulations. All the 15 ensemble members display a too slow MJO propagation, which suggests that the difference in MJO speed between the model and reanalysis is statistically significant. This slow propagation of the MJO in the model is confirmed by the fact that the number of days spent on average in each individual phase of the MJO is higher in the model than in the reanalysis. It is not limited to the Indian Ocean, but also takes place over the Pacific and the Western Hemisphere.
The percentage of events propagating from one phase of the MJO to the next phase is displayed in Table I The table shows that the percentage of events propagating over the Indian Ocean to the west of the Maritime Continent (from phase 1 to 4) in the model hindcasts is close to the reanalysis. However, the percentage of MJO events crossing the Maritime Continent (phase 4 to 5) and propagating into the West Pacific (phase 5 to 6) is lower in the model than in the reanalysis. In the reanalysis, 30% of MJO events do not propagate from the west part of the Maritime Continent into the Western Pacific (from phase 4 to phase 6). This percentage climbs to 50% in the model. This confirms that the model has difficulties in propagating the MJO across the Maritime Continent. In addition, the model tends to regenerate too many MJOs from a previous MJO event: in the reanalysis, only 25% of MJOs in phase 7 start a new MJO event over the western Indian Ocean (phase 2). In the model, this percentage climbs to 40%.
Table I. Percentage of MJO events moving from one MJO phase to another for reanalysis and model hindcasts.
1 ⇒ 2
2 ⇒ 3
3 ⇒ 4
4 ⇒ 5
5 ⇒ 6
6 ⇒ 7
7 ⇒ 8
8 ⇒ 1
4. Impact of the MJO on precipitation in the Tropics
This section evaluates the impact of the MJO on precipitation in the Tropics in the model and in ERA-Interim. The first 15 d of the forecasts have been excluded of the calculation of the composites in order to minimise the impact of initial conditions. In order to reduce the number of panels in the figures, phases 2 and 3, 4 and 5, 6 and 7, and 8 and 1 have been merged. Phase 2 + 3 corresponds to the active phase of the MJO over the Indian Ocean, phase 4 + 5 corresponds to the MJO over the Maritime Continent, phase 6 + 7 corresponds to the convection over the West Pacific, and phase 8 + 1 corresponds to the MJO over the Western Hemisphere and suppressed convection over the Maritime Continent. Phases 2 + 3 and 6 + 7 correspond to opposite phases of the MJO, as do phases 4 + 5 and 8 + 1.
Figures 4 and 5 show the composites of total precipitation anomalies for phases 2 + 3, 4 + 5, 6 + 7 and 8 + 1 for the period November to April (Figure 4) and June to August (Figure 5) 1989 to 2008. The eastward propagation of the MJO from one phase to another is visible over the Indian Ocean and West Pacific in both seasons. The eastward propagation of precipitation in the model (panels (e) to (h)) is expected since the propagation of OLR anomalies is contained in the definition of the MJO indices used to create the composites. The model represents well the differences between summer and winter, with precipitation anomalies propagating northward in summer and southward in winter. The northward and southward propagations of precipitation are not present in the MJO index, which is based on fields averaged between 15°N and 15°S.
For the period November to April (Figure 4), the MJO composites from the reanalysis show an impact of the MJO on precipitation over Equatorial Africa and Brazil. Phase 2 + 3 is associated with drier conditions over Brazil and Equatorial Africa, whereas phase 8 + 1 is conducive to wetter conditions over Brazil and Equatorial Africa. Those MJO teleconnections are generally well captured by the model hindcasts, although the anomalies over Africa tend to extend rather more to the east than in ERA-Interim. The model simulates wet conditions over the southeast of the US during phase 8 + 1. ERA-Interim displays a similar pattern but more to the northeast. Those patterns are not a trivial consequence of an MJO event because they are located east of the dateline, where the eastward propagation of convection associated with the MJO stops. Therefore the model seems to capture reasonably well the general impact of the MJO on the atmosphere in the Tropics in the period November to April.
In summer (Figure 5), the MJO has an impact on the precipitation anomalies over Central America, with wet anomalies during phases 2 + 3 and 8 + 1 and dry anomalies during phases 4 + 5 and 6 + 7 in ERA-Interim. The model reproduces well those teleconnections (Figure 5(e) to (h)). This suggests that the model should have enhanced skill to predict precipitation over Central America a few weeks in advance. Interestingly the precipitation anomalies over Central America are consistent with the impact of the MJO on tropical cyclone activity over the Atlantic and the eastern North Pacific. During MJO phases 2 + 3 and 8 + 1 (phases 4 + 5 and 6 + 7), the tropical cyclone activity is enhanced (reduced) over the Gulf of Mexico and eastern Pacific in observations and in the model simulations (Figure 3 in Vitart, 2009).
Over Equatorial Africa, the impact of the MJO during the summer season is mostly visible in phases 2 and 6. During phase 2, ERA-Interim displays an increase of precipitation over the equatorial band of Africa, although the signal is noisy (Figure 6(a)). During phase 6, the impact of the MJO is the opposite to its impact in phase 2, with drier conditions over most of Equatorial Africa except for East Africa (Figure 6(c)). In the model hindcasts, the impact of the MJO on African rainfall is generally consistent with ERA-Interim in phases 2 and 6 (Figures 6(b) and (d)), although the dry anomalies in phase 6 extend too far to the east in the model. The amplitude of the precipitation anomalies is also weaker in the model than in observations during phase 6. The patterns in the model are smoother than in ERA-Interim, but this can be due to the fact that the model integrations represent the equivalent of 300 years, instead of just 20 years for ERA-Interim.
In the model, the impact of the MJO on precipitation anomalies in the Extratropics is generally much weaker than in ERA-Interim. For instance, the positive anomaly over China in phase 8 + 1 in ERA-Interim during summer is not reproduced by the model. This is also the case for the precipitation anomalies over the North Atlantic in winter.
5. Impact of the MJO on the Northern Hemisphere weather
5.1. Impact on 500 hPa geopotential height anomalies
Figure 7 shows the composites of 500 hPa geopotential height anomalies for phases 4 + 5 and 8 + 1 of the MJO in the model hindcasts at the time range 16– 45 d and in ERA-Interim for the period November to April 1989– 2008. The composites are performed only on cases where the amplitude of the MJO index exceeds one standard deviation. The anomaly patterns are generally consistent between model and reanalysis and are generally of opposite sign between phase 4 + 5 and phase 8 + 1. However, the amplitude of the anomalies can be different. For instance, the impact of the MJO on the North Pacific sector is stronger in the model than in the reanalysis during phases 4 + 5 and 8 + 1. Over the Euro-Atlantic sector, the impact of the MJO on 500 hPa geopotential height anomalies is generally weaker in the model, particularly over Southern Europe and North Africa. For phases 2 + 3 and 6 + 7 (not shown), the patterns of the MJO teleconnections are also generally consistent with observations, but the amplitude of the anomalies is much weaker in the model, including over the North Pacific.
Using reanalysis data covering the period 1974– 2007, Cassou (2008) showed that the impact of the MJO on European weather is the strongest about 10 d after the MJO is in phase 3 or phase 6 (Figure 3 of Cassou, 2008). The probability of a positive phase of the North Atlantic Oscillation (NAO) is significantly increased about 10 d after the MJO is in phase 3 (phase 3 + 10 d), and significantly decreased about 10 d after the MJO is in phase 6 (phase 6 + 10 d). The probability of a negative phase of the NAO is decreased (increased) about 10 d after the MJO is in phase 3 (phase 6). The impact of the MJO on two other Euro-Atlantic weather regimes, the Atlantic ridge and Scandinavian blocking, is much weaker. This section will focus on phase 3 + 10 d and phase 6 + 10 d to evaluate if the model can reproduce this MJO impact on the NAO.
Figure 8 shows a 10 d lagged composite of 500 hPa geopotential height anomalies in the hindcasts and in ERA-Interim for phases 3 and 6. The impact over the Euro-Atlantic sector is much weaker in the ensemble forecasts than in the 20 years of ERA-Interim reanalysis. For the 10 d lagged composite of phase 3, the amplitude of the 500 hPa geopotential height anomalies is stronger in the model than in the reanalysis over the Pacific sector, but weaker over Northeast Canada and Europe.
The fact that the MJO propagation in the model is too slow may have an impact on the MJO teleconnections. To test this hypothesis, the phase 3 + 10 d composites have been produced, but only for the MJO events simulated by the model which have a similar propagation speed to that in the reanalysis (Figure 9(b): the criterion is that the MJO starting in phase 3 reaches phase 5 by day 8). Those composites are then compared to the MJO composites obtained from MJO events in the model which propagate slowly (Figure 9(a): the criterion is that the MJO is still in phase 3 by day 8). The Hovmöller diagram of 500 hPa geopotential height averaged between 40 and 60°N indicates that the fast-propagating model MJOs display a faster and stronger wave propagation of 500 hPa geopotential height anomalies than the slow-propagating MJOs (Figures 9(c) and (d)). The lagged composites at day 10 are shown in Figures 9(e) and (f) for slow- and fast-propagating MJOs respectively. The 500 hPa geopotential height anomaly composite obtained with the fast-propagating MJOs (Figure 9(f)) looks more consistent with reanalysis (Figure 8) than the composite obtained with the slow-propagating MJOs (Figure 9(e)); over the North Pacific sector, the positive anomalies are located more to the west in the fast-propagating MJO cases than in the slow-propagating MJO cases as in the reanalysis. Over Northeast Canada and Europe, the anomalies are stronger with fast-moving MJOs than with slow-moving MJOs and are therefore more consistent with the reanalysis. This result suggests that some of the discrepancies between the model simulations and ERA-Interim in Figure 8 are due to the too slow MJO propagation in the model. However even the fast-propagating MJO events simulated by the model fail to produce 500 hPa height anomalies over Europe as strong as in the reanalysis and the maximum of the positive anomaly is located too far south, over North Africa.
The model composites are produced over the equivalent of 300 years (15-member ensemble × 20 years), whereas the reanalysis dataset used in this study includes only 20 years. Therefore, it is possible that the strong signal over Europe in the reanalysis (Figure 8) is due to sampling errors. To test if this is the case, the model composites have been computed for the combination of two individual ensemble members, instead of the combination of the 15 ensemble members. The combination of two ensemble members represents 40 years of forecast, which is twice the number of years in the reanalysis dataset used to produce the composites in Figure 8 and about the same number of years as the analysis dataset used by Cassou (2008). Figure 10 shows the phase 3 teleconnections after 10 d for some combinations of two ensemble members. Some of the realisations (perturbations 0 + 1 for instance) show a maximum positive anomaly over North Africa, when another one (perturbations 2 + 3) shows a maximum over northeast Europe. Another realisation (perturbations 13 + 14) has a maximum anomaly over the Atlantic, and another one (perturbations 7 + 8) has a maximum anomaly over south-central Europe as in the 20-year ERA-Interim reanalysis (Figure 8). This suggests that there is considerable uncertainty in the impact of the MJO over the European sector, even in a 40 year simulation. The Pacific sector and western Canada display considerably less variability from one combination to another.
5.2. Impact on weather regimes
Another method to investigate the impact of the MJO on the Northern Extratropics weather is to project all the forecasts into pre-defined weather regime patterns. This is the method Cassou (2008) applied to an analysis covering the period 1974– 2007. He found that the frequency of positive NAO regimes (NAO+) increases 10 d after an MJO event in phase 3 and decreases 10 d after an MJO event in phase 6. In the present paper, the model hindcasts and the ERA-Interim reanalyses are projected onto four pre-defined weather regimes (NAO+, NAO-, Blocking, Atlantic ridge) for the period December-January-February 1989– 2008. The pre-defined weather regimes have been computed by Corti and Ferranti (private communication) from ECMWF reanalysis data using the algorithm developed by Straus et al. (2007).
Figure 11 indicates that the probability of a NAO+ event increases with time during the 15 d following an MJO event in phase 3 in the majority of ensemble members. Only three ensemble members (in 15) show a decrease of probability of NAO+ by day 15. This result suggests that the MJO simulated by the numerical model has an impact on the weather over Europe. The increased probability of a NAO+ event in the ensemble mean (about 20% at day 10) is only about half the increase obtained with ERA-Interim (dashed line in Figure 11). However the increase obtained with ERA-Interim is within the spread of the 15-member ensemble, and there are a couple of ensemble members which show an increase of the probability of a positive NAO similar to the one obtained with ERA-Interim. Therefore it is not clear if the discrepancy between the ensemble mean and ERA-Interim is due to an inadequacy of the model to represent correctly the MJO teleconnections or if this is due to the fact that 20 years of reanalysis is too short to evaluate the impact of the MJO on weather regimes. The categorical definition of weather regimes may also contribute to the surprisingly large spread in Figure 11, since two relatively close weather patterns can be identified as different weather regimes. If we consider only the fast-propagating MJO events (dotted curve in Figure 11), the probability of NAO+ events increases slightly more than when considering all the MJO events, but the increase still remains lower than in ERA-Interim.
The probability of a positive NAO diminishes during the days following an MJO event in phase 6 in both ensemble mean and ERA-Interim (Figure 11(b)). The amplitude of the decrease of the probability of a positive NAO is of the same order of magnitude in the ensemble mean and in ERA-Interim after day 10. The decrease obtained with ERA-Interim is about half the amplitude of the increase obtained after an MJO in phase 3 and it is weaker than the decrease displayed in Figure 3 of Cassou (2008) where it reaches about 40%
Overall, the model displays a 10% decrease in the probability of a negative NAO (NAO– ) in the 15 d period following an MJO event in phase 3 and a 12% increase in the 15 d period following an MJO event in phase 6 (not shown). The sign of this variation of NAO− probability is consistent with ERA-Interim and Cassou (2008), but the spread in the model ensemble distribution is particularly large. Only half of the ensemble members display a decrease of NAO− probability following an MJO in phase 3 or an increase following an MJO in phase 6. The amplitude of the changes in the probability of a negative NAO displayed by ERA-Interim lies within the model ensemble distribution. As for the positive phase of the NAO, a few ensemble members show an evolution of the probability of a negative NAO similar to ERA-Interim.
Nine ensemble members in 15 (60%) simulate a steady decrease of the probability of a Scandinavian blocking following an MJO in phase 3 and an increase after phase 6. On average, the frequency of blockings is reduced 15 d after phase 3 by about 6% and increases 10 d after phase 6 by about 10%.
The model also simulates an impact of the MJO on the probability of an Atlantic ridge, with an overall decrease after an MJO in phase 3 (66% of ensemble members) and an increase following an MJO in phase 6 (75% of ensemble members). Overall this represents a decrease or an increase of about 10% in the probability of an Atlantic ridge by day 15 following an MJO respectively in phase 3 or 6.
As for NAO+ and NAO– , the impact of the MJO on the frequency of blockings and Atlantic ridge is consistent with ERA-Interim, although the amplitude is on average lower in the model simulations than in ERA-Interim.
6. Impact of the MJO on the monthly forecast probabilistic skill scores
Previous studies (Ferranti et al., 1990; Jones et al., 2004; Jung et al., 2010) have suggested an improvement of forecast skill during MJO events. Ferranti et al. (1990) showed that a 20 d forecast was improved when the Tropics were relaxed towards observations during an MJO event. Jones et al. (2004) used a 10-year run of the NASA GCM to show that the potential predictability is increased by 2– 3 d in that model. More recently, Jung et al. (2010) showed a reduction of extratropical forecast errors for periods with active MJO events in an experiment where the Tropics are relaxed towards observations as in Ferranti et al. (1990).
In the present paper, a different approach has been used to provide a more quantitative assessment of the impact of the MJO on the ECMWF monthly forecast probabilistic skill scores. The 120 15-member ensemble forecasts (all the forecasts starting on 15 October, November, December, January, February and March 1989– 2008) have been classified as a function of the presence or not of an MJO event in the initial conditions. About 55% of the 120 cases have an MJO in the initial conditions (this MJO event can be in any phase). Probabilistic skill scores computed for all the cases with an MJO event in the initial conditions are then compared to the probabilistic skill scores computed for all the cases with no MJO event in the initial conditions. The probabilistic skill scores applied include the Relative Operating Characteristic (ROC; Stanski et al., 1989; Mason and Graham, 1999) and Brier Skill Scores (Wilks, 2005) of the probability that 500 hPa geopotential height, 850 hPa temperature or total precipitation over the Northern Extratropics (north of 30°N) are in the upper or lower tercile, for the weekly periods (days 5– 11, 12– 18, 19– 25 and 26– 32). For precipitation and temperature, only land points have been considered. The definition of the weekly periods corresponds to the one used in the operational ECMWF monthly forecast products (Vitart, 2004).
The ROC areas and Brier Skill Scores for the probabilities to be in the upper tercile are displayed in Tables II to IV. The results for the low tercile probabilities (not shown) are similar. According to Tables II to IV, the ROC areas and Brier Skill Scores are higher when there is an MJO in the initial conditions than when there is no MJO in the initial conditions (numbers in parentheses), except for 500 hPa geopotential height and 850 hPa temperature at days 5– 11 where the scores are very close. The difference is statistically significant within the 10% level of confidence using a 10 000 bootstrap re-sampling procedure for the 500 hPa geopotential height skill scores for days 12– 18 and 19– 25 (Table II). For temperature at 850 hPa (Table III), the difference is statistically significant for days 19– 25. For precipitation (Table IV), the difference is statistically significant within the 10% level of confidence for the Brier Skill Score for days 5– 11. The difference of probabilistic skill scores between the cases with and without an MJO in the initial conditions is not statistically significant for days 26– 32.
Table II. ROC and Brier Skill Scores for the days 5– 11, 12– 18, 19– 25 and 26– 32 for the probability that 500 hPa geopotential height is in the upper tercile when there is an MJO in the initial conditions. The number in parentheses indicates the score when there is no MJO in the initial conditions.
Days 5– 11
Days 12– 18
Days 19– 25
Days 26– 32
Table III. As Table II, but for temperature at 850 hPa.
Days 5– 11
Days 12– 18
Days 19– 25
Days 26– 32
Table IV. As Table II, but for total precipitation.
Days 5– 11
Days 12– 18
Days 19– 25
Days 26– 32
The impact of the presence of an MJO in the initial conditions is particularly important for the period 19– 25, which is a time range often considered as having very low predictability and reliability in the Extratropics (e.g. Vitart, 2004; Weigel et al., 2008). Therefore it is interesting to notice that, when there is an MJO event in the initial conditions, the forecasts over the Northern Extratropics have a positive Brier Skill Score for 500 hPa geopotential height and temperature at 850 hPa for days 19– 25, suggesting that those probabilistic forecasts are likely to be useful at this time range. When there is no MJO in the initial conditions, the 19– 25 forecasts have very low ROC area (close to 0.5) and negative Brier Skill Score, indicating that those forecasts have low skill and are not reliable. This result is confirmed by the ROC and reliability diagrams (Wilks, 2005) (Figure 12) of the probability that 850 hPa temperature is in the upper tercile for various regions, including Europe. Over Europe, the day 19– 25 probabilistic forecasts display some reliability, with a reliability curve close to the diagonal, when there is an MJO in the initial conditions. However, the probabilistic forecasts are unreliable (almost flat curve) when there is no MJO in the initial conditions (dotted line in Figure 12(c)). This result suggests that the MJO represents a major, if not the main, source of predictability in the Northern Extratropics at this time range. This also demonstrates that the skill at this time range is not always as low as previous studies have suggested, and forecasts at this time range can be potentially useful over the northern Extratropics. From a practical point of view, this result also suggests that the users of the ECMWF monthly forecasting system could use the presence of an MJO in the initial conditions to decide if the monthly forecasts at days 19– 25 should be trusted or not.
For days 12– 18, the presence of an MJO in the initial conditions enhances the skill of the model, but the model already displays significant skill at this time range even when there is no MJO in the initial conditions. For days 26– 32, the presence of an MJO in the initial conditions also improves the probabilistic skill scores, but the probabilistic scores are very low, even with an MJO in the initial conditions.
This paper has documented the main characteristics and impacts of the MJO in a set of 15-member ensemble hindcasts. The model displays some notable skill to predict the evolution of the MJO (about 20 d of predictability). However, the MJO simulated in this set of hindcasts suffers from the following problems:
The MJO simulated by the model tends to be too strong, but this has been partially solved in the subsequent version of IFS.
The simulated MJO tends to be too slow. This is not a systematic problem, since some MJOs in the model propagate at the same speed as observed MJOs, but on average the simulated MJO events tend to stay longer in each phase of the MJO.
The simulated MJOs often have difficulties crossing the Maritime Continent. Statistically the percentage of MJO events which do not cross the Maritime Continent is higher in the model than in observations. In these cases, the convection can be locked over the Maritime Continent until the end of the 46 d forecast.
The simulated MJOs tend to regenerate a new MJO too often.
Of all those problems, the too slow propagation of the MJO is probably the most serious issue for the current ECMWF monthly forecasting system, particularly for the longer time range (days 19– 25 and 26– 32). The too slow propagation of the MJO and its difficulty to cross the Maritime Continent may cause the forecast to be out of phase with observations after 20 d on some occasions.
The MJO teleconnections are well reproduced by the model in the Tropics. For instance, the impact of the MJO on the precipitation over South America and Equatorial Africa are generally consistent with reanalysis. Vitart (2009) showed that the impact of the MJO on tropical cyclone activity and risk of landfall was also realistically simulated by the ECMWF model. In the Extratropics, the model simulates an increase in the probability of a positive NAO following an MJO in phase 3 (enhanced convection over the eastern Indian Ocean) and a decrease following an MJO in phase 6 (suppressed convection over the eastern Indian Ocean). Overall, the model teleconnections in the Extratropics are generally consistent with ERA-Interim, except over the Euro-Atlantic sector where they are weaker than in ERA-Interim. Fast-propagating MJOs in the model display stronger teleconnections over Europe than slow-propagating MJOs, but the teleconnections are still weaker than in ERA-Interim. However the observed impact of the MJO on the Extratropics remains uncertain since 20 years of reanalysis may be too short to assess the MJO teleconnections in midlatitudes.
Finally, the impact of the MJO on the extratropical forecast skill was investigated. Results show that the MJO has no significant impact for the period 5– 11 (except for precipitation), but has a a positive impact for days 12– 18, 19– 25 and 26– 32. This impact is statistically significant for days 12– 19 and 19– 25. The impact of the MJO is particularly important for days 19– 25 with the model showing almost no skill at all when there is no MJO in the initial conditions, but the day 19– 25 probabilistic forecasts become reliable and skilful when there is an MJO in the initial conditions. This suggests that it is possible to know apriori if a monthly forecast will be reliable or not. Those results also suggest that improvements in the representation of the MJO in the ECMWF model are likely to lead to improved monthly forecast skill. Woolnough et al. (2007) have shown that coupling the atmospheric model to a high-vertical-resolution ocean mixed-layer model can impact on the speed of the simulated MJO events through its impact on the SST diurnal cycle and intraseasonal variability. Therefore coupling the version of IFS used in this study to an ocean mixed-layer model, as in Woolnough et al. (2007), may help the atmospheric model to produce faster MJO events, which could lead to more realistic MJO teleconnections and enhanced skill in the Extratropics. This will be explored in future studies.
We are grateful to Rob Hine, who has helped to improve the quality of the figures, and to the two anonymous reviewers whose comments proved invaluable in improving the presentation of the material.