 Thanks, Jo Kannon. When I started putting together this presentation, I regretted giving this title to the organizers, because I realized that I wouldn't be able to talk about what I wanted to talk, which is showing our results. But I think it's likely to still be interesting, at least from my point of view, to try to make sense out of a very strange and difficult soup of acronyms that you might have seen this week. But if you have been attending these kind of meetings before, there are many more acronyms. And I'll try to make a bit of light on this, not easy and not obvious. Just to introduce myself, just briefly I come from the Barcelona Supercomputing Center in Barcelona. And I lead the departments with around 50 people who work on environmental forecasting. And these are the kind of things that we do atmospheric compositions, atmospheric chemistry, cloud prediction, computational earth sciences, and also services. And the two aspects, or the two groups that have been pulling some results for this presentation from are the one on services and the climate prediction one, of course. So this is one of the sentences that I took from the page that describes the workshop. And it mentions WCRP, which is the World Climate Research Program for those who don't know it. And its core projects and WCRP has four projects. Click here that works on the cryosphere, clivar, GWX that works on the hydrological cycle, and Spark that is interested in the stratosphere. And obviously, we are in clivar. This is a clivar workshop. The sentence also says that, well, WCRP has identified an interest in the K-dol climate for ability and predictability. And that brings along another acronym, DCVP. I warn you that the acronyms are going to crop up quite a lot in this presentation. So I'll try to spell them out as often as I can. So DCVP, which appears here as, well, an acronym for a concept, is also a group of people who are meeting this afternoon and tomorrow to discuss about, well, these kind of aspects. And it's one of the research foci of clivar. I'll show it a bit later. And also to continue this experimental climate prediction, which is something that is managed by another group that is called DCPP. So there is a P instead of the V. So also in WCRP, you have working groups. There are four. One of them is WGCM, the working group on coupled modeling. And this is the working group that is managing SIMIP. So SIMIP is part of WCRP. So SIMIP is not IPCC. SIMIP is WCRP. Mike will tell me off if I don't clarify this thing. So there is another working group that is relevant here, which is WGCIP. This is mainly the working group I'm involved with. And it's the working group on seasonal change to annual prediction. And for instance, WGCIP is hosting DCPP, the Decadal Climate Prediction Project, which is one of the modeling toward comparison projects of SIMIP 6. And again, it has a panel with lots of very interesting people. But WGCIP is also doing lots of other things. So for instance, one of the projects that is relevant to what has been discussed here is a project on drift and initial shock. And you've heard already a few presentations about drift and initial shock from Christoph just a few moments ago and also yesterday. So these are two of the four working groups of WCRP. WCRP also has grant challenges. And there are six here, one that I put here in white, because it's the most relevant to us. That is the grant challenge on near term climate prediction. If you want to know what the grant challenge is going to do, talk to Johanan. He has all the information. There is not much information yet, to be honest. So this gives you a fairly quick overview of what WCRP is doing. But there are other organizations inside the core projects, for instance, that are also relevant. Inside Pliver, for instance, we have a panel, which is the group on synthesis and observations, which is mainly looking into the different ocean reanalysis and reference products that we can use to validate and initialize our systems for forecasting. But there are also research full site. And one of the research focuses of a Cliver is DCVP, the Decadal Climate Variability and Predictability Group. So this is complex enough, but things are even more complex. And outside WCRP, we have WWRP, which is a World Weather Research Program. So the weather people, they get organized separately. And they are as numerous as we are. And there are some initiatives inside WRP. Some of them are coordinated with WCRP. One of them is the sub-seasonal to seasonal projects, again, concerned by some of the problems that we've been talking about here, initialization, for instance. PVP, which is a polar prediction project that has to do with, for instance, some of the things that Amy has described in her presentation, and the joint working group on focus verification research that people with this horrible acronym that you see there, they do really interesting things. I'll try to illustrate some of them. And then on this side, on the right, you have other initiatives and panels that are very, well, you might be familiar with, which is the IPCC, that takes a lot of information from SIMEP, but also from the rest of WCRP, at least for working groups one and two, and also GFCS, which is a global framework for climate services, which is, again, something, another initiative inside WMO. And WMO is outside all this and is much, much bigger. So there are many more panels and committees. And I'm sure that you will agree that it's quite difficult to make sense out of this. And I guess that people are already quite happy of being here. And, well, that's already organizing your work and your contacts enough. But let me be a bit controversial here. In my opinion, this is very compartmental. It's preventing progress. And I think we need to be a bit more promiscuous and try to be engaged as much as we can in activities that other communities are doing. Because somehow, this is a clever meeting where, for instance, the Spark people are not present or we're not people working on the grand challenge on cloud circulation and climate sensitivity are not here. But climate sensitivity is fundamental for the kind of thing that we are doing. Why this happens? Well, I guess that we all have ideas about why this is the case. So let me go a bit more into some of these initiatives and panels. Simip V was designed to feed the fifth assessment report of the IPCC. And at least for the decadal prediction parts that contributed to Chapter 11 of the Working Group 1 report, there were a few lessons. This is my selection. I'm sure that many of you will remove some of them and add many, many more that I'm sure that will be very important and very, very interesting. But just to start with, there is a climate prediction and a climate modeling community that got together to start the decadal prediction. And the first thing that we realized after two years of fights is that we were talking about different things with the same words. So there was a massive language problem. And I was talking about this with Susanna yesterday. It's a horrible, terrible situation. And something else that it was obvious is that the decadal prediction has skill. The skill, the skill over the continents, at least for temperature. For precipitation, it's not the case, except in certain regions. But there is skill. The decadal prediction has skill. Sorry if I repeat it, but many people say that the decadal prediction doesn't have skill. It's not true. There is skill. But most of the skill in temperature, at least in temperature, is forced by forcings, natural and external forcings. And we saw in Gavin's presentation, you talked on Monday, Gavin, I think it was, that if we had the right forcings, we could have predicted the, well, much better the slowdown of global mean temperature. Unfortunately, we don't have the forcings because we are doing prediction. So we have to live with that. We also need to have many more handcuffs than the ones that were done for CIMIP-5, starting every five years. Volvo yesterday showed how the skill scores go up and down. And we've seen other examples as well. And also, it's very important to use the same model configuration as the ones that are used for climate experiments. Christophe has been talking about the pacemaker and the fixed anomaly experiments. And he could use the same model that he was using for the historical experiments. So is it the same version that he used for the decadal predictions? And this is the way we can really learn about our decadal prediction systems. Because if we start changing the versions, it's going to be very difficult to make progress. So learn about your model and its drawbacks. And the other really tough thing was that drift, initial shock, and systematic error are hampering the progress of decadal prediction. And several of us have been talking here about why we didn't do more flux correction for decadal prediction. Well, I don't know. Maybe it's a time to do some work. Julia has been doing some. And E.C.M. Doliev as well. But it's not systematic. So this is one of the plots that was showed before. These are the decadal predictions in solid. Four from SIMIT 5, from systems that started every year. In gray, you have the observations in dashed. You have the historical runs for the left global mean temperature on the right AMV. It's a simple AMV index, which takes an aerial average of SSD over the North Atlantic minus the global mean temperature. And it's quite nice to see that the AMV is better predicted when you initialize than when you use the historical runs. And it's also striking the similarity between the global mean temperature predicted and also from the historical runs. But we have this. I think it was Susanna who mentioned that the historical runs were overdoing the global mean temperature in the last 10 years, while the initialized simulations didn't. And there might be two ways to explain what happened here. One is that the SIMIT 5 systems were very good at facing in the internal variability. The other one is that they are correcting the incorrect forced model response in our systems. And I would like to advocate for the second one. So it's true that SIMIT 5 didn't use the correct volcanic aerosol forcing, and the aerosols were not correct. But somehow, by initializing, we are correcting for these aeros in the forcing. So basically, there are two things that we should expect from the KDL prediction. One is the phase in of the internal variability, which is something that we've been talking about here, which is something really tough. But the other one is the correction of the forced model response up to a certain time into the future. And let me go a bit more in detail with this idea. But this is another sentence from the page describing the workshop that says that, well, the warming hiatus caught us by surprise. But there was a failure of the initialized coupled model predictions to detect it. In my opinion, this is not true. They detected it. They predicted it in hindsight. But was it for the right reasons or not? It depends on what the right reasons are for us. But for the users, the right reasons is whether they can really make better decisions based on the information that we give them. So more on the correction of the forced model response. These are results from a very nice experiment that was run years ago at the Met Office. A huge data set was constructed using the perturbed parameters, the perturbed parameter approach for the perturbation of the model, the several configuration of the model were run in the KDL focus mode with initialization and without initialization. And one of the interesting things about this system is that you have several configurations of the same model with different responses to the forcings. So these are, well, measures of scale in terms of correlation for temperature. And the vertical axis represents the aggregation of the full costs, especially while the horizontal axis represents the focus time. So we are aggregating in focus time or we are aggregating in spatially. So we are smoothing the maps, basically, as we go up in this axis here. So the first column corresponds to the uninitialized runs. The second one to the initialized runs. And this is the difference in correlation. And the difference between these three configurations, but in average, well, the configuration one, two, and three, is that this one, the one at the top, has the highest slope in global mean temperature in the historical run. The second and the third are the ones with the smallest slopes in the global mean temperature. So what we can see is that the impacts of the initialization is much bigger when we have shallower slopes in the global mean temperature than stronger slopes. The interpretation of this is that, again, the initialization is doing more than just phasing in the internal variability. And the reason why we are benefiting is because the models have the wrong force response to the 4.6. Starting every year, the hindgust, it makes the hindgusts much more expensive. It's not the same doing hindgusts every five years than running hindgusts every year. But it has an impact on what we can say about our systems. This is the skill for the AMV, it's a function of the focus time. And each color corresponds to a different time averaging of the forecasts. So in red is the root mean square error for the forecasts taken every year. And well, in green is the result for the forecasts when we have averaged them every three years into the forecast and so on. So we have more and more smooth version of the time series. So when we initialized every five years, which is what we see here, we see apps and downs. And we saw some of the results yesterday that, well, they seem to suggest that there was a recovery skill after a few years into the focus, which sometimes is a bit difficult to explain. However, if you take exactly the same period, so 1960 to 2005 for the verification, and look at what is the root mean square error when you initialize every year, basically we have exactly the same information here, we are not increasing the number of degrees of freedom of the skill. But we're having much smoother versions of the skill measures. And the reason is that we need to sample the internal variability during the period when we are making the forecasts. I'm not saying that it's absolutely necessary to do high gusts every year, but every five years is definitely too little. And, well, this is something similar when looking at individual models. So I'm talking all the time about prediction. And in prediction, we are initializing a climate model. And we do this, as I said before, to address internal variability, to take into account the incorrect force model response. And we initialize the model by using the best available observations. So we have to link to the observation lists, whoever they are. That's already a first challenge to identify them. To transfer that information into the couple model, avoiding imbalances, again, it's easier said than done, as you heard the discussion yesterday. And then we need to run with initial perturbations that, well, produce a spread that is representative of the uncertainties that we have, which are a lot. But unfortunately, the uncertainties are uncertain. So what we do is something like this. We run a set of high gusts with an ensemble, in this case, well, let's say five members for the second illustration. One year later, we do the same thing. And then five years later, we've done already six high gusts. And we carry on doing this thing every year. And we start from an observational data set, some measure of what the current state of the climate system is, not necessarily the best estimate of the state of the climate system for prediction. The reality is slightly different, because Susanna, I think, showed this, or Daniela showed this plot from the IPCC, where we can see that in the IPCC, we had models that used the anomalous initialization approach, the full field initialization approach. But even if they used the anomalous initialization approach, the systematic errors were massive. And those using the full field initialization approach, they're suffering from these drifts that you see here. You don't see the initial shocks, but they are astonishing. And it's really enlightening to have a look at the KDL forecast and see an animation of any variable using six-hourly or three-hourly values. With a month is enough. You don't need to go any farther. The model is all over the place. It's incredible. It destroys everything that you are putting in. It's so amazing. And it's very nice to see, because you see all the processes that we've been talking about here in action all at the same time, without interaction, with interaction. And the problem is that we are forecasting while the model is doing this. So we are trying to say something about the future while the model is in complete, well, going through a complete reorganization of the information in itself. So we need to do something. And some of the things that were done was a discussion about an evaluation of the relative merits of the anomaly initialization approach. And there are many approximations that have been implemented for the anomaly initialization approach. So several institutions carried out, again, as a result of SIMIT 5, but also as part of SIMIT 5. They carried out intercomparison experiments where they either initialized in the real world, which is here in green. And if you do this, then the model just goes to its own attractor, so it drifts. Or you initialize here, and you initialize in the model world, which is already biased. But at least you know that the initial shock is going to be smaller or inexistent. So there are many examples of this. I'm just giving one example here, but the Metaphys looked into it, and MPI and the University of Hamburg have looked into it. And many other people, for instance, Environment Canada has done the same thing. But unfortunately, until now, there are very few proofs or very few hints that anomaly initialization is giving us a bit of skill. In fact, it seems that full field initialization is giving better skill in spite of the initial shocks and the drift. And this is an example from Danila Volpi, who looked at different ways of implementing anomaly initialization in the ocean, taking into account just the correcting the anomalies in TNS or in T and density, or introducing some anomaly initialization as well in the CIs. And there are small differences and small improvements coming from the anomaly initialization. For instance, in the AMV or in the Arctic CIs volume, but after three years, no distinction, no difference whatsoever in terms of skill. And you have to really fight hard to get the anomaly initialization working, because you have all sorts of drifts and misplacements of the observed anomalies into the model. Basically, it's a data simulation problem in not in the observed attractor, but in the model attractor, far from Danila. So it's a really nice challenge, but we have to do something in the meantime to provide the users with some information. We also did some additional work. I won't go into the details. To try to understand a bit more why, in underserting conditions, we were having a better skill in the first one or two years with the full field initialization and why in other conditions we were having it with anomaly initialization. And we used a hierarchy of models, basically the Lorentz model, the Peñan-Kalne model, which is a version of the Lorentz model with nine variables, and also the Van Itzeman, the Cruz model, which is a simplified couple model with less than, well, it's 29 variables. And what we found out using a very simple mapping approach to implement the anomaly initialization is that when the attractors are incompatible, when the model attractor and the observed attractor are not compatible, then there is a massive initial shock. The initial information is lost after just a few time steps. And then the model is better in using an anomaly initialization approach. But most of the times, the attractors are not fully incompatible. They are compatible, but they have differences, for instance, in certain variables in the kurtosis or in the skewness. And that's making the linear approach for the implementation of the anomaly initialization, which is the one that has been implemented until now, really unfeasible in giving the, well, unsatisfactory results. These are two of the references of this work using simplified models. But it's an example of how we need a hierarchy of model that goes even beyond just looking at Amy runs or ocean only runs. We really need to put under test the approaches and the concepts that we are using. So there is going to be a bit more discussion about this next year. We're organizing a workshop in Barcelona. It's sponsored by the specs and the preface European projects. And it will also receive sponsoring from WGSEP, which is the working group on seasonal to inter-annual prediction. And basically the idea is to come up to an agreement in terms of the kind of experiments that we need to address this issue. The issue of the initial shock and the drift. But at the same time, discuss about the feasibility of coupled initialization. Note that I'm not talking here about coupled data simulation. I'm talking about coupled initialization, which is a different thing. To address the problem. And for this, we need representatives of the global, of the group on synthesis and observations panel from of Glybar. So again, trying to be a bit more promiscuous here. And the other objective of this workshop is to discuss about an eventual recommendation from WGSEP on the problem of bias adjustment. So in the end, the users, they need something that is not an anomaly. They need a field that they can use. And a field that resembles the reality. And this is something that still requires a lot of work. Symmep six, this is coming up. Christoph has already mentioned some of the activities. Of at least the decadal component of Symmep six, DCPP. In that context, decadal prediction can benefit from a better understanding, from the possibility of looking into the control runs and understanding the characteristics of the variability in our models. And also from the infrastructure. If we want to share data, it's all done. So it's already there. We'd rather be close to what WGCM is already doing. Other MIPs, other modeling into comparison projects can also benefit from the decadal prediction part of Symmep six, because there will be more people looking at the reduction of the systematic error. We have a massive interest in reducing this because it's destroying our initial condition information. Not all, but a lot of it. And also because we'll do a continuous verification of the models. We are continuously comparing with observations, with real observations. But it's very expensive. These are really huge experiments. At the same time, there will be a real-time decadal prediction component in Symmep six that is led by DCPPs, the one called component B. And there will be also hopefully a lot of work on looking at other processes that might bring additional skill. So before Symmep six, we have to work on issues like for instance, the impact of the volcanic aerosol and how we project the volcanic aerosol into the future. It's one of the tasks of component C. Because in the end, we have to recognize that we can't predict volcanoes. And even if a volcano went off, we won't have the real estimates of the four things or the aerosol loads before a few years after the volcano has gone off. We also need to set benchmarks. It's very good to keep working with the couple models. They are very complex. We've seen lots of their limitations. But at the same time, there are people working on very interesting alternatives that can really help the development of the model, but at the same time, be used as a benchmark. And a good example is you could see it in an image-suckling poster this week. This is an example of the skill score in terms of correlation for the critical decadal prediction system that they've built. This system has the additional interest that it's a system that is calibrated. So it's probabilistic. It's not just giving you a value for the next five or 10 years. It's giving you a range of values that are calibrated with the observations. And then we have observational uncertainty. The observational uncertainty is hurting us in creating the initial conditions. And it was very nice to hear the argument yesterday between Dirk and Alicia. Not that I like to hear you arguing, but I like what you say when you argue. But they also limit our ability to know what the skill of our system is. And this is an illustration from seasonal forecasting. It's from a set of seasonal forecasts. We run with East Earth, the new version of East Earth. And this is the skill for Neo 3.4. For Hankus, started on the 1st of May and running for four months. And it's a skill for each month for the same system but evaluated against four different observational data sets. The one from ESA, it's a 25 kilometer data set that has been recently released. All the others are low resolution data sets. And we found out that when validating forecasts from high resolution systems, well, you have to take into account also the uncertainty that is associated to the resolution of the observations, which is again something quite fascinating to look at. Sorry, GFCS, again, another component. This one is outside WCRP. So you don't need to feel concerned about it, but it's a very nice thing. It stands for Global Framework for Climate Services. And basically we are here in the research, modeling and prediction part. But just to give you an idea of what climate services is about, we are just in a corner. And the observationalists are a bit here, next to us, but the bulk of the climate services is a completely different kind of problem, where we have something to say and where some communities are waiting for us to interact with them. These communities, they get together and one of them is getting together in the Euphoria project, which is another European project. This one led by the MetaFest. And they came up with these principles for a successful climate service. So it's actually quite nice. It looks funny, but it has a lot of depth into it. So it's a matter of defining your problem, having a roadmap, being flexible, verify and evaluate and monitor, being transparent to avoid tensions, being able to listen and use the principles of research and development. These principles that you might think that they are very obvious and that they've been put there for the users, actually they are not for the users. They put them together for us. So this is what they expect from us to have a successful climate service. And on top of this, it's actually interesting to see what they are coming up with. They're the right, really interesting reports. There is a very nice one that was recently released. It's the ethical framework for climate services. I know it sounds a bit pompous, but it's really interesting reading, very short. And it's based on four core elements that they want to propose that everyone interested in climate services abide, well, follows. Integrity, transparency, humility and collaboration. And as obvious as they are and as much as we think that we follow them, I think I did my own exercise of, well, critical exercise to try to say if I was following them and I failed in two of them. So I'd suggest or I would like to encourage you to try something similar. In any case, something, a mistake that we are doing, coming from the WCRP side, is to think that climate data is climate information. And it's not true. Climate data is just a piece of the climate information puzzle. So something very nice about climate services is one initiative that is taking place within the working group on seasonal to intranet prediction and in particular, as part of the Decadal Climate Prediction Panel, which is the multimodal Decadal focus exchange. If you go to this website that the Met Office, you'll find plenty of information and it will be kept under development by DERC, DERC's group in the next few years. We also have, especially from our community, downstream services. So portals that distribute data and make very nice plots and they provide some information. Well, they have a link, for instance, to the ESGF data. For instance, now we have Decadal data on Decadal predictions on ESGF. And it's a key aspect for success. It's good to have this thing to engage with climate services, with the users. But there is a bot. And the problem can be illustrated with this plot from Andrea Taylor, who is at the University of Leeds. So this is a very interesting plot. Andrea is a psychologist. So, and she writes very nice papers that I really recommend you to have a look at. She works on seasonal prediction. And she, as part of Euphoria's conducted a set of interviews, a few hundred. And among those interviews, she analyzed the users of seasonal forecasts that responded to a set of questions. And that these are the forecasts that indicated in which way they received information about uncertainty. So here we have uncertainty in the form of range of values. So an interval, which is something that we do very often in certain, sorry, range of values or confidence intervals or variable descriptions of likelihood. Very tough to communicate. Role data, well, that's something that they already say that they received. But what they didn't receive is information about how good those forecasts are or indicators of the signal strength and what it means or information about possible sources of error. And in portals like this, we can't do this because that goes down to human interaction. And unfortunately, human interaction takes time and engagement from people. On top of this, we are providing information that this information needs to have a quality measure. And one of the groups inside WMode that is really doing a lot of work in this direction is the joint working group on focus verification research. And the address questions like how do we interpret something like this contingency table that we have here. So this is a table where we put the, well, let's say we have a number of events that we've been looking into and we have also forecasts for those events. And usually we look at this column here. We look at the events that were observed, I don't know, the hiatus, for instance, or the shift in the IPO. And we look at whether the model was successful or missed it in the focus. And we also look a lot at this one. So there are plenty of presentations at the EGU or the AGU saying, well, I focus this thing. But it was so good that I got the shift at the right time. It's really good. But what we never see is that box. That's a box in which you have all those cases that the model reproduced or predicted an event that actually didn't happen. And this is the one that is hurting us. What is hurting us is not this one, but this one because it's the one that is affecting the users. The users are affected by cases in which we cry wolf and the wolf doesn't come. This is the, that's a case. And let me just wrap up and conclude with an example for instance, a user of decadal predictions. And this is the case that we built in collaboration with a reinsurance company. And there are other examples along the same line in which we demonstrated to the company that the advantage of the decadal predictions with respect to the historical simulations for the next 10 years, for their trade, for in particular for the prediction of the impact of tropical cyclones in the North Atlantic, is that the decadal predictions are not just more skillful because we can predict the AMV, but they are also more credible because we have a better estimate of the actual uncertainty of these forecasts. Because we have reduced in a certain way, in a meaningful way, the uncertainty associated to the initial conditions. And I'll jump that one and go to just last overhead in which I would like to make a plea for all of you to get engaged into activities like this one, this is a fact sheet series that we started as part of this specs project in which we try to explain with very simple words and using a vocabulary that has been said previously what we are doing, the kind of problems that we are dealing with. And I know that Ed is doing plenty of things along these lines, both inside the community and outside the community as part of his gliver activities. There is a pressing need for things like this because otherwise we might be out of business sooner than you expect, unfortunately. And this is something that, well, we might see before we retire. Just a summary, there is a complex ecosystem of international activities and many of them are relevant to decadal prediction and predictability, but we need to know who they are to take advantage of what they are doing. And at this stage, there is a lot of information missing about this. There is a broadening range of users who is asking for information and giving them information about the decadal predictions even if the decadal predictions don't have skill is a fair thing to do because it will be a matter of providing them with a good climatological focus which is, again, not as trivial as we might think. Decadal prediction is showing signs of providing useful information. We've seen examples here. Emerging all this information into a reliable source of information is not a trivial task either. The users don't want to have five different focus systems and then find a way of what to do with it. It doesn't work this way, unfortunately. The models have drift and there is a massive initial shock in a full field initialization. The initial shock doesn't go away even when you go into anomaly initialization and it's something that we have to understand a bit better and basically we need to keep asking for investment in observational networks. We've seen already how important the lack of observations is. Increased collaboration and a reduction of all aspects that are at the origin of model error. And I think I'll leave it here. Thank you.