so today's paper is titled conditional portfolio optimization I felt like it was a very interesting paper because it reveals a lot but at the same time doesn't reveal much at the same time um much more to that will come I'll explain further so a little bit about me my name is I am a junior qualitative researcher at arts and teams and I'm a proud VTC meaning that I graduated at the University of Pittsburgh's right here in Johannesburg I graduated with a master's mathematics and decided to then switch to quantitative Finance simply because I wanted to try and explore how I can add more value in the space given my strong background mathematics but I slowly realized that that's quantity Finance is just one third mathematics two-thirds finance and actually take the skills but over the past few months and years or so I've gained enough experience enough knowledge to actually have a say or actually have an opportunity like this one to be working for us in the teams and actually conducting a reading like I have today so like you have on the screen yeah I derived a interesting quote from the paper itself whereas where they say that fast performance is not indicative of future results and I feel like quantitative analysts or researchers should live by then but anyways let's go on so I partitioned the slides or the presentation into eight main sections the first section is going to be the introduction we're just gonna here to do a bit of repetition maybe things that you already know but we're just going to talk about traditional optimization method and then we're then going to compare those to what the authors the label this machine learning happens and see how the two actually flow against each other and then we're going to talk about the core concept of what exactly is CPO and how for instance perhaps this paper adds to the community then when you look at the methodology and looking at how CPU actually works it's a two-step approach and then after they want to look at the process exactly how it works the bare bone try to add a bit of Flesh to it and try to understand the actual thought process behind the construction of the model processing and so forth and then we're going to look at the results um some may be controversial we'll look at those and finally I have my own personal Computing remarks and my own personal thoughts through that in the paper so I think without further Ado let's start so like I said perhaps this is a bit of repetition in itself but nevertheless I think it's worth mentioning that um traditional approaches to portfolio optimization typically involve using mathematical models using to construct uh portfolio to maximize um your given risk also to maximize your returns given a level of risk and I think that's textbook that's something if you're probably learning quantitative Finance you will learn that's essentially what um sorry that's essentially what um traditional method is important in terms of trying to optimize your portfolio now in order of maybe more naive to more advanced some of these methods involved equal waiting risk parity and mean variance AKA markovics but now all of these approaches do not typically account for changing the market or consider the impact of Market regimes when you say Market regimes we're just looking at ships in the market and how for instance investor contribution changes and how the markets just generally change and I think this has become even more relevant in our times where we see sudden shifts in markets that happens so happens so rapidly for instance I think in the space of five years we've had the covid-19 pandemic and we also have the war on Ukraine just two significant Market shifts that have happened and the authors are suggesting that CPO will be able to pick up those regimes and those chains and be able to adapt and there's an interesting things up or they will go through that actually shows results in which this actually occurred and the managed the model managed to actually adapt and still maintain a good return okay so just the comparison now uh we've already talked about traditional methods as in the sense that the focus on construction postponing that maximize returns or minimize risk under a single assumption or constraints AKA this could be a sharp Ratio or if you think about the sharp ratio trying to maximize the denominator of minimize the numerator now these methods assume that acid returns are normally distributed or usually assume that assets returns they normally distributed I think there are methods to combat this but it's not always foolproof so that poses as an issue at times especially when you're doing Financial modeling using traditional methods and not just in portfolio uh optimization I think this applies a lot in a lot of quantitative finance and another big assumption which is why then the authors wrote the paper is that traditional methods don't sorry traditional methods assume the market conditions are stable over the time um perhaps maybe this is more naive because that has never been the case history has shown there's been big Market changes or Market regimes I think a very simple Market regime or context could be just a bull market versus bear market we've seen shifts of up and down bull markets Bay markets and a lot of traditional methods always assume that the markets are stable so this almost trans Factor traditional methods perhaps should be scraped away and we should throw focus on machine learning methods so in contracts regime change methods AKA machine energy methods take into account change in the market conditions but actually identifying different regimes or states that characterize Market at different times and this is huge this is why for instance the authors release the paper and this is why I feel like it actually adds value to the enrollment because of that fact that by taking some sort of inputs we can essentially bring out a way we could optimize the portfolio by accounting for the current market changes or Market uh shifts that are happening now another point to consider is that reading changes methods recognize the asset returns and correlations can vary significant across different market regimes and this is also a big one too because I think with traditional methods especially when you look at the naive one which is um a kind of fun correlations I think I remember when I was starting one quantity Finance just reading from the textbook a simple way to diversify portfolio or two essentially optimize the portfolio is just to account for different correlations so having assets that are uniquely different from each other so as to diversify your risk now machine learning takes that a step further by actually aligning those with the market shifts that are happening as they are happening or at least try to do that and then another point is that regime changes can provide more robust and adaptive portfolio allocation that can perform better in Dynamic monthly conditions where traditional approaches may actually not be able to do so this is also quite significant and robust can also be taken on the pinch assault because we know that whenever a model in machine learning is a bit complex or complicated it tends to exhibit signs of eye variance or it tends to overfit but we'll see examples whereby when they tested their model with in-sample databases or sample data they may not be actually a big change or shift but however there's more to come and I think we just talk about that when we get to that station so now the question is what is CPO and what is the concept behind the paper now I decided I decided to ask the question of how to actually answer the whole concept of CPU which is given a hypothetical portfolio now this portfolio could be anything assuming could be Forex it could have um bonds you could have whatever you want there's no there shouldn't be any limits to what you can include in your portfolio and so given a hypothetical portfolio and information about the market environment key point can I predict an objective function that I care about now the objective function can be anything you want given your investment preference as your as a manager or as a researcher whatever the case is cannot manage to predict or optimize we'll see that an objective function given Market environment and this has a difficult portfolio and CPU turned out to be the answer to that question so the methodology now now at the core concept CPO is a two two-step approach and firstly involves regime identification so in the first stage the machine learning model is trained to identify different regimes based on historical Market data which is the first part now the authors mentioned that they used a neural network but they did use the fact that they did mention sorry the fact that they use that lightly or to just maybe per se um to make it easier to maybe to explain but in actual fact they use some sort of gradient boost integration tree I.E potentially in XG boosting model and this model will be used to be trained on historical data to try and actually pick up the regime changes in the market and and for my own research I feel like this could be also substituted with potentially any uh model that you prefer I personally prefer the hidden markovian model because of its ability to identify latent States and time series and the fact that it actually tries to put more weights or emphasis on recent events rather than the past events but that's just the personal preference and then the second part of the two-factor approach is actually portfolio optimization this is where now after the model has been trained on the historical data historical data with your two inputs being your like you see on the diagram here your control features and Market features I will explain those it then is now a trained model and then after it's been trained you can then now try to use that instead of constraints to then actually optimize the portfolio and see which allocations are optimal for let's say trying to get the best objective function in the case of the paper they used a sharp Ratio or rather the one-day sharp ratio in most of the examples but like I said this objective function can be anything you want it's just based on your preference as an investor and so on which is why I feel like the actual methodology and method is actually more of a skeleton and can be filled in anything you want in terms of your the flesh you'd want so and the core concept is using machine learning to identify and adapt to different market regimes and improve portfolio performance over time and we know that with machine learning models they actually get better with actually with more time they learn to learn especially when you use XG boosting models or even hidden microbial models or neural networks or deep learning of some sort anything that has some sort of reinforcement learning in it we know that machine learning models get better with time and with more data of course so that is at the core concept of what CPO is in context of actual paper um yeah so um I I just want to interject so um I I think maybe just said slightly differently a very this is probably the most important thing to understand about the concept is um so you have two steps what essentially you you say is given um my control features and the market features I want to I want my model to learn to predict some kind of objective function and the example they used here was the the one month um one month into the Future Shop ratio and so the first step isn't actually about doing any kind of optimization in a in a later slide you'll see there's actually multiple steps so the first step is can I just create some kind of model that given a hypothetical portfolio the control features and um the current state of the market can I just take those two things and predict what my Sharp ratio will be in the future one month from now um so that's the first step and so the idea is just can I map a portfolio Plus Market features to a sharp ratio and and you figure out how to do that across all different portfolios given the current given the current market features um and then once you have that model that's trained that now understands how you know Market features will um affect the portfolio then you start saying okay how can I tweak my control features essentially because the control features are these levers you can pull the weights that you can shift around how do I tweak those so that I end up with a maximum sharp ratio at the end so it's it's slightly different to how a lot of models would typically um or to a lot of approaches that you typically take it's not going you know um find me the maximum return it's going find me any return given a control features and any set of Market features and so that's why you can see in this diagram you know the control features are all different it's different weightings of the portfolio but the market features are all the same and um that's the that's the kind of key idea here and so the the aim and the hope is that internally the the um model will develop some kind of understanding of market regimes and how it will affect the portfolio um so that's the kind of key idea behind um CPO cool thank you thank you Michael Yes um yeah I love I love that the reiterating that there is no optimization in the first step it's just training and learning and just teaching the model given these market regimes so for instance if you see the diagram here on our right um we have two inputs we have control features control features will be essentially what will make up your actual portfolio so in this case we have three assets you have your Google stock your Microsoft stock and your Apple stock and then you have your Market feature and these will be directly linked to the market um perception or essentially how the the market State exactly and in each row is essentially an allocation or a potential allocation that the model will be something through so you look at here in the first row we could either assign 20 of our money to Loosely say to Google 60 to Microsoft Apple 20 and so on and so on and so on so from the paper you can see that for you potentially will see that there's infinite possibilities to allocating for instance let's say for instance you have 500 stocks like you in the managed to partition this S P 500 into is separate 500 stocks and you assign let's say you have two positions you can either go short or you either go long so essentially you have two to the power 500 possible allocations and the model has go through all of them so it can then try to predict or try to understand which one would actually be the best allocation at the other steps now of course this poses as a big problem because I don't think we have the computing power to go through two support 500 possible allocation I think when I saw the number myself I was shocked how big the number is and is this not possible so the authors said that to overcome this they came up with something they called intelligent sampling they didn't actually specify what exactly it is and anything we can come up with this is only speculation I potentially think it's some sort of random walk sampling of some sort whereby it tries to live put this focus on the ones that seem to be using less um potential allocation and just try to focus on the ones that are maybe yielding a higher allocation sharp ratio of some sort something like that but I think there's a big mystery there in the fact that we don't know what exactly intelligent sampling is but they said that this is for instance I quoted here overcoming this curse of dimensionality by intelligence is something the grid is one of the major breakthroughs of our team as our country but which if they put a black box they don't intentionally um because we don't know what intensity intelligence offering is um but it's brilliant that it passed down potentially something through to the power 500 if you have 500 stocks in touchdown I think we don't even know how much it costs down the time by we just know that according to their work it works so I think we can go to the next slide now so then now we can start to look at the bare ball process of how CPO this is essentially CPO and its skeleton you start off with the training stage which is step one you take your historical Market features the ones I label here as or sorry the label gears Market features you take your control features the ones that are in your actual portfolio which is your in that case the Google and the Microsoft and so on you put that into some sort of pre-processing stage and you then you get your historical features right and then you add your historical portfolio in the ShopRite because remember you want to train your model as much as possible and then with that then it does what in the models then that's what they said intelligent software which is to start through all possible roles to see which is the best or potentially the best one-day shop ratio for example there's no optimization here it's just giving results of which ones are potentially the best one-day shop ratio and then you train the model and then you have a trained machine learning model then you move to stage two which is the inference stage you take the current market features so here you're looking at the Historical looking at maybe you can look at your I don't know three years or two whatever period I think maybe it's just your preference and your computing power and then you take your current features and you combine those as the control features remember your model is trained now you apply that to the train model and then you click get your predicted end date sharp ratio so the objective function in this case or in this example is trying to predict the end date sharp ratio if anything they made it more specific that they're trying to predict a one-day shop ratio and then you get to the optimization stage which so essentially the stage is one and three are what I'll characterize as what CPO is really all about picking up the regime changes and now optimizing optimizing the actual portfolio so now you you have your predicted ending sharp ratios you have your different allocation and then you add your constraints constraints are essentially your way so for example you don't want all yours you don't want your model to end up putting a certain amount of stock certain amount of money let's say interested in stock and try and actually potentially put you at risk so you put like a constraint to say that on stock or any stock I need to make sure that the model does not go above 40 and with cash I need to make sure that the model doesn't hold more than 50 cash for example so this will be the constraint and then they do another intelligent sampling essentially what it's doing is given those constraints which is actually used in this sharp ratio in this case sharp ratio but it can be any objective function you want this is the optimization part and then it base gives you essentially what is a base sharp ratio and the actual base allocation in this case would have been the first one which is this one 20 Google Microsoft 60 and Apple 20. uh Michael do you want to add anything on the slide uh yes so so the only thing um I I kind of wanted to point out here is uh those labels that they use to actually train the model uh it's important to understand that those were probably back tested so so um what they would do is they would take their their Market State and they would well they'll take their hypothetical portfolio and they would then back test that portfolio and the sharp ratio one month into the future that would become the label for that particular row and that will become a training sample right so you have your your um your portfolio you have your historical Market features and then you have the predicted shop rate all your back tested sharp ratio and you do that for all possible combinations hypothetically and then that's your training set so I think at the end we're going to have a little bit of a discussion on how that intelligent sampling works I I think Miss Simba had a good idea for how you'd maybe do that um there's a bunch of ideas when we were talking offline uh of of how you could possibly do sampling over a very large kind of sample space with lots of Dimensions how do you reduce that down um have a little bit of experience solving problems like this so we were kind of brain brainstorming how you do that and then once you've kind of got that then you can actually train your model um and then like you see then there's an inference stage where you make your predictions and then based on your constraints you can actually optimize um so yeah nothing more to add the only thing um symbols I would say I think the example they used is a one month Future Shop ratio that they look at not one day but once again you can choose uh any optimization objective function that you'd want so for example if you wanted to minimize variance you know across your portfolio or drawdown or something you can make that your objective function um and then that's that's kind of how this would work so it's a very general approach and um yeah uh we'll we'll get into how well it works in a second but that's basically all I all I have to add in the process at this point so yeah thanks Mr but you can carry on thank you Michael um so I think the big takeaway for me just on this because I feel like it offers a nice skeleton to actually add on however you please so for instance unfortunately we couldn't get access to the source code or any coding work regarding the paper because the authors actually are implementing these um the strategy with clients and potentially I'm sure there's some MBA agreement or something along those lines they're just not comfortable so it's uh giving us access to the code work but it's nice to get a diagram like this because it gives you a nice blueprint of how you can actually approach machine learning methodologies when it comes to portfolio optimization and you can then fill in the blanks however you like of course that would depend on your research team how much time you have all of these factors come into play but I think this is good enough as some sort of blueprint to say I could have a three-stage process in trying to optimize my portfolio and trying to actually start off with the training stage then go to inference stage then give the optimization stage of course there's a lot of um uncertainty when it comes to intelligent sampling and so on but like Marco said we will talk about that at the end on potentially what ideas could have come across in terms of what intelligence cycling is but I just like the fact that this is a nice blueprint to how you can tackle portfolio optimization in general anyway so let's go to the results now um so we have exhibited one we have an ETF I I didn't know what this ETA was I had to Google especially Canadian exchange traded fund but you know it's good to learn and the constraints they put was maximum weight on any stock was 10 and maximum any cash is 10 notes cash does not necessarily need actual cash it could mean that but also cash could mean an investment in a risk-free asset for example your um treasury bond for example so in this case out of sample between 2021 August to about 2022 um beginning of August um you see that the ones that are the the naive sorry the naive traditional method could potentially performed amongst the worst and funny thing I was surprised to see them appropriate as she performed the worst I thought maybe being fed is that more advanced technique would perform maybe the second best but markovitz actually was a worth the naive one was amongst the worst but CPO actually performed slightly better than the other one the interesting part now is that by slightly adjusting the same period same out of sample by slightly adjusting the weights here we now have 25 on any stock 50 on actually on cash CPO performs significantly better than the other ones and that's that's big because I think it's natural to know that during bad times you'd rather hold on to the most amount of cash we'd rather than have your money or your Investments dwindling to inflation to recession and so on so the model is able to pick that up by just simply adjusting cash we see a significant changing in our results sorry to interject I think maybe just an interesting comment about this um observation in some ways that's kind of not related to the paper but interesting nonetheless is um that's your constraints which typically are not data driven right you will typically have either a law or a mandate that will tell you how much of what you're allowed to actually hold um it's actually quite surprising how big of an outsized effect that can have on returns across all kinds of strategies um so I just think that's quite an interesting thing to note in general um and and I've not actually seen a lot of talk about maybe setting constraints or optimizing constraints in a certain way right so that could actually be something I mean the paper doesn't talk about this but um the thinking in the future maybe it starts making sense to have constraints be data driven maybe a certain set of constraints actually leads to really excellent returns and so that's yet another facet you can maybe optimize on in the future um but yeah I just think it's important to kind of notice this uh in this particular case thank you thank you so moving on further now we get a bigger picture of how CPU is actually performing versus the more traditional art approaches to portfolio optimization so we here we have a private Tech portfolio saving new stocks two Canadian and on the y-axis is a portfolio value um I don't know what exactly that means but portfolio value assume the higher the value the more sharp Ratio or whatever the case is and then on the XX is your time so you can see that for instance during bull markets perhaps between January 2020 and maybe up until January 2022 CPO actually capitalizes well in these bull regimes so if in this maybe 2017 to about 2020 there's not much of a difference interchanging you can't really tell which one is doing well but the moment there's a bull market or a bull run rather the the model is able to capitalize more on these markets right and then as you can see potentially around 2022 January or maybe February which is in line with the actual uh Ukraine Russia wall breakup now we have some sort of change in Market regime and we have a dip and CPO seemingly is aging well against this regime change yeah compared to looking at your equal weights your max shop ratio your minimum volatility and so on and this is quite significant because we'll see in the next slide here that at any point with those some sort of Max drawdown CPO allocated the most to cash as you can see here with these I mean on this diagram here versus the max rule down here on an S P 500 and this is essentially what maybe Michael was also iterating that the constraints are significant in trying to optimize or trying to protect your investment in a portfolio because the more months you Society catch like Naturally Speaking you don't want your money or your Investments during me when it's times of that so you want to hold as much cash as possible and CPO is managing to do it almost hedging on things bad times and I think that's quite significant in how the authors brought this to light and actually perhaps introduced a new way of portfolio optimization and here um we had two two diagrams so the top diagram they try to essentially maybe test CPO against these new age assets like your cryptocurrency so the top diagram has eight cryptocurrency portfolio it's a portfolio with eight cryptocurrency sorry and it's alive for short positioning aiming to maximize it's seven day four sharp ratio and I mean there's a significant change here uh sorry a significant difference here sharp ratio being significantly higher the CPO and sorry the out of sample period here was between during 2020 and June 2021 ideally I would wish there was a longer period but we take what we have and this is the period they use for this experiment and below is a bit of one of their clients which they said that they had to ask for permission from wsg investment so this results from a portfolio of seven trading strategies and I think wsg is a Forex uh trading uh company I'm not quite sure but they essentially do a Forex so even though they're trying to show how CPO deals with different types of assets and seems to actually do well with those different types of assets again the auto sample period for the Ws SG investment result is between during 2020 and July 2022 not the best in terms of the length I did it would have been better if it was longer but they also tried they included June July 2022 to try and show how it flares even when there's some sort of um dip in the market which is different excuse me which is demonstrated by the Ukraine war breakdown in 2010 in February 2022 now then they were not shy to show some interesting results per se whereby CPO may not actually being the best in terms of um the optimum um strategy in terms of portfolio optimization so they created what they labeled tactical asset allocation TAA this is inspired by The Golden Butterfly portfolio if you're aware of them and essentially the portfolio comprises of old small cap stocks large cap stocks which is your spy you want one to three year to rebounds shy or shy and your 20 plus here every bonds your TLT and in thoughtful here's a surprising fact here in Sample equal weight and a sharp ratio of 0.51 CPO had a sharp rate of 0.63 between the period of 2015 January and 2018 December 2019 uh January right you wouldn't think C you wouldn't think equal weight would end up performing better out of some but when they then took it out of sample incorporating that regime change that I keep mentioning the Ukraine Russia war that broke out in February 2022 equal weights actually outperformed CPO CPU I keep performing second base and they did mention that this was a bit of a surprise because they wouldn't have thought that in Sample equal waste didn't show any signs of potentially being the best beta option in terms of Ops portfolio optimization but out of sample in that same in that significant out of sample period equal weights was actually the best and CPO was second best but also trying to reiterate that CPO still will be able to perform well in this customized uh portfolio that it was created so I found that very interesting as if they're sharing the same sentiment with the authors that I wouldn't have thought equal weights would have been the one that would perform the best out of sample in that period given that we know equal weight is slightly seen as a naive approach or some sort of naive strategy because it's um it's entry level but maybe this thing gives us an idea amongst many others that we need to take the results of the pinch of salt but anyway we'll talk about that in the conclusion and so on and this is just to reiterate the out of sample allocations that CPU would have made of course maybe it would have been better to also see equal weightings and how that was happening but we're just gonna have to look at um our CPO allocation so this is also trying to give an idea of how it rebalances or readjust itself over time and they did mention that the model rebalances I think every seven days or something like that uh correct me from Rock Michael but I think it's something like that so between 2021 January we start off with a large chunk of our portfolio in cash which is shy we said cash does not necessarily mean actual cash it actually means a treasury bond or something that's risk-free it then readjusts that towards the end of 2021 or Midway rather to have more in actually in your S P 500 or in some sort of optimistic stock to see maybe potentially arrive or potentially invest the confidence rising or something like that and then looking at the beginning of 2022 a lot of our acid was in gold and this dramatically changes towards mid-2022 remember after the significant event now we have a significant of a significant amount of money in in cash and I feel like this is just to show how it keeps rebalancing and re-changing itself to always have the optimal um allocation towards which assets should in terms of which with the constraints you'd have said how much of the assets should we hold and so on and so on and this is I think I feel like this is also proven why CPO is potentially a game changer in this optimization industry uh or in the uh in the capture optimization industry and trying to actually see the optimum number or sorry the optimal portfolio that can be held because when let's start it off as a point in trying to understand Quant all I knew or what we're told about is equal waiting which is just trying to use correlation Matrix make sure that you hold assets that are opposite correlated to each other or inverse equal related to each other to try and diversify your your your risk so now we talk about conclusion and we'll come to the end so I think the main takeaway that I would want you to go with is that CPO is a two zip approach right first approach is the actual regime identification second um the second step is actually the optimization process there's an intermediate step but just at the core those are the two um findings of what the papers bring to the table interesting results now we have a mixed bag we have a situation whereby uh we have a situation whereby CPO has shown great Potential from the results which I do believe they're not enough but anyway from the results CPO is actually performing well against what's most fund managers will be using to optimize their portfolios however which is a bit fair that they did include a scenario whereby CPO wasn't actually the best although it did perform well it wasn't the best and this also the important question of timing is everything when it comes to uh portfolio investing in portfolio management the dock number two machine learning or cooperation equals significant Improvement I think one thing we share here it has to maintains is we're trying to be the Forefront in understanding machine learning algorithms and how they actually can influence our world today and the Beautiful part about today is that most of our computers can handle so much workload in terms of computing power and so on so machine learning is an integral part of what conditional portfolio optimization is but I think taken with a pinch of sold because I feel like we need more results one thing I do feel like we shall discussing with Michael is that we they tested on very short time frames I wish there was a longer time frame to test or they test on a very specific data sets potentially you see if they're maybe given us access to some of the coding we could have done our own experiment on some sort of random um asset allocation to see how that flares up and I also feel like there's very very few implementation details I did mention earlier on that the paper doesn't actually the paper reviews a lot but it doesn't reveal much at the same time and this is quite sad because the methodologies that are being applied in the paper are groundbreaking and can actually help us move forward our portfolio optimization techniques however there's not enough coding work or not enough um there's not enough information surrounding for instance intelligent sampling to understand how exactly is the model working and Which models are actually being used to do some of this work that you claim because they did mention like I said neural networks but there was a small disclaimer or a small note at one of the pages saying that they actually use mostly used XTU boost so I feel like this I'm sorry I feel like there's very few implementation in detail regarding the actual methodologies and I think my final caviar is there needs to be further research needed to explore is applicability in different markets and that would have been done if perhaps gave us access to some of the coding work and some of the implementations on maybe on a python notebook so we can actually follow along and we can also potentially do our own experiment um they did mention that due to the fact that they have seen success with their methodology that that actually essentially make it into a product so if you're interested in trying to implement the product I think you can visit I don't know their website or something but the idea behind not giving us a code work is because they are their implementation has been so successful that they actually package it into a product which is fortunate for us researchers but nevertheless I feel like we can always go back to that skeleton framework that I'm I showed you which is what they gave us which could be something we can lean on to try and give us an idea of how everything is working more like the brain behind everything but of course there's a lot of research that will be need to be done to actually understand each component at each stage what is actually going on your EPO and I think with that I would just like to conclude and hope you enjoyed today's presentation