Algo Trading Week Day 5: Current trends in quant finance [Panel Discussion]

hello everyone i hope all of you can hear me and see me we are live now just want to confirm before we start that all of you can hear me can you please type in the questions panel just yes okay thank you so much for the confirmations all right so uh let us start today's session so uh welcome to algo trading week day 5 yesterday we hosted a panel discussion on sentiment analysis and alternative data we also announced our second certification program after epat that is certificate in sentiment analysis and alternative data in finance csaf to know more about this program visit us at quantumc.com csaf uh today we have another exciting panel discussion on current trends in quant finance and we will be joined by three domain experts now let me invite our first panelist david jessup from london united kingdom david is head of investment risk for emea region at columbia thread needle investments in his present role david has responsibility for overseeing the independent investment risk management process for all portfolios managed in the emea region before joining the company david was the global head of quantitative research at ubs over his 17 years at ubs his research covered many topics but in particular the he concentrated on risk analysis portfolio construction and most recently cross-asset factor investing the application of machine learning and bayesian techniques in investment management and prior to this he was head of quantitative marketing at citigroup and david started his career at morgan grenfell initially as a derivative analyst and then as a quantitative portfolio manager so david has an mma in mathematics from trinity college from cambridge uh hello david are you here with us welcome to algo trading week thank you very much very happy to be here and uh i say thank you for a very comprehensive introduction i have to say that's very impressive thank you so with that i'll also introduce our next panelist so dr devashes guava from bangalore now dr guha is director of machine learning and chair of center for research in technology business at sp gen school of global management he has more than two decades of experience working in the field of artificial intelligence and especially its applications to economics and finance he is also the founder of bangalore based company that has provided consultancy for quantitative hedge funds across the globe he has formerly been head of risk management and quantitative trading at big sky capital that is a california-based hedge fund and he was also partner at global trend capital which is another california-based hedge fund dr gufa is a graduate of iit kharagpur and has a phd from columbia university in new york welcome dr gurhar to algo trading week hello thank you very glad to be here and thank you for that great introduction okay now uh it's time to invite our third and final final panelist of the day so richard rothenberg from new york united states so richard is an executive director at uh global ai corporation which is a big data an artificial intelligence company that provides quantitative research and data-driven signals and alternative data for institutional clients including hedge funds and governments previously richard worked as a quantitative portfolio manager and a researcher at multi-billion dollar hedge funds and global investment banks including deutsche bank uh man investment and other leading institutions richard is a research affiliate at the lawrence berkeley national laboratory one of the world's largest supercomputing laboratories and as advisor at the defense advanced research project project agency that is dharpa richard is a member of task force on data for the sustainable development goals at the united nations conference on trade and development and a member of united nations science technology and innovation expert group richard holds a bachelor degree in economics and computational finance from the monterey institute of technology uh a certificate of quantitative finance from the cqf institute and a master's in management and quantitative finance from columbia university in new york city welcome richard welcome to the algo trading week thank you appreciate that all right so i think uh with that said uh let us start with our first question all right now this first question is very similar to the title of this session and that is how has the quant finance landscape evolved in recent time and are there any trends that any of you have observed so we'll go alphabetically so first leave it this question can you address this question yeah i mean i think they've been really two big things that have driven the way quant finance has changed and i might add a sort of third one to that um the first is uh the availability of data uh you know you look at i mean i'm not sort of just sort of saying the volume of data that we're producing globally i mean that's a well-known statistic although quite how much of that data is actually useful is a whole other other question um but it's also the availability of the data the fact we can get to data very very easily now it's a lot cheaper than it was um there's a lot more quality in the data in fact i mean a lot of my career has been spent cleaning data sources not actually using them and i've got to say cleaning data is by far the least exciting bit of any quant project and yet it's the one bit you have to do i think the other big trend uh has obviously been just the increasing computing power um you know mer's law is basically true um you know you go back as of 10 years and you know you look at the power of computing there's a stuff we're doing today which you just could not really do 10 years ago and i i so i think those are the two underlying sort of truths you know we have more data and more powerful computers um i think the other thing that's happening that's interesting uh to me i think there's two ends of this one is quant trading quant quant in general is expanding out of equities i think traditionally it was a very equity focused uh because again a data availability issue um and i think people are moving into credit into sort of crypto trading into a lot of other assets currency i mean currencies obviously were there all the time and then the other one which actually picking up what i'm sure richard will talk about later um and this is a different story i mean this is not a trading story is the whole area of responsible investing of um sustainable investing however you want to describe it and again there we're right at the beginning the data quality in that area is is i'm sure richard will will have a lot to say on this is horrendous you know it really isn't very good and i think that's one of the big things here at columbia thread needle it's a massive massive sort of thing for us is going much more into that area so i that's not necessary from a trading perspective but i think it is a massive thing that is going to be going on in the whole of the finance area uh over the next five ten years all right so uh uh dr debasis would you want to share your thoughts in this question yes i would add one point to something that david already touched upon and that is the expansion of quant into many many more asset classes than just equity which is where quant started of course we had commodities and currencies for a long time uh but now we have multiple asset classes where quant investing is coming into its own i myself have been looking at uh crypto but there is also uh alternative commodities you know and new uh derivatives have been introduced that have now made quant trading easier in some of these uh you know more exotic asset classes i remember talking to a friend of mine some 20 years ago and suggested to him that water would be a traded asset and he was not very clear that you know it was going to happen but he happened to know the people at in chicago and he talked to them and he came back running to me and said you were right all along water is good there is going to be a water futures so that's that's one example of an exotic uh commodity that has now become tradable and uh crypto is just the latest thing in this uh bouquet of assets that uh have become uh not just tradable but tradable algorithmically and i think that this is a trend that's going to expand even further and that's i i see that as one of the major trends that pond not only will be successful in just equity but you know point is going to take over more and more of the trading in all asset classes including the exotics that's that's the trend that i see as the clearest one i mean one of the clearest one all the others are also very clear that david mentioned and richards were probably going to talk about yes over to richard yeah hi um so i would like to echo some of the points that david and dr wu have mentioned and i would say that one of the more impactful factors has been the significant increase in the amount of alternative data out there so historically we only were using financial time series and now we're looking at credit card transactions uh satellite data sentiment data from news all over the world in multiple languages and also the ability to process this data has increased dramatically and and of course the amount of computational power needed to process this um has gone along with decrease of you know cloud computing and also the methods used to analyze this data where previously based on a number of assumptions you look at different models that are based on the normal distributions and now we're looking at the non-linear models that are more dynamic and adaptable on to to outliers and fat tails another important issue is the increase of the um as david mentioned the increase in the non-financial risk factors one of which is the environmental social and governance trends where we look at a number of issues that are outside of the balance sheet that dramatically impact companies valuation and and and this goes along this is applicable across multiple asset classes as mentioned and what i find interesting is that even beyond bonds we're looking at private equity companies you know using alternative data on illiquid assets like infrastructure and um private companies and others and so this is an interesting trend that is just going to accelerate uh and that's why i think it's very important to um you know to uh to learn more about uh these topics yeah thank you all right thank you so uh considering david initially talked about cleaning up data that no one likes i have a few questions coming up on that but before i uh start on that i would like to pick a interesting questions to and directly towards david and is it all just about machine learning now um i i we thought we'd start with the most sort of radical question when my two co-panelists basically are sort of working in machine learning um a lot of it is yes but i think you have to you know i i always come out as i always seem to come over as a slight skeptic about machine learning in finance because it's a difficult problem you know finance is a very low signal to noise area um and it's an adaptive market and it's a zero-sum game you know so if you come up with a better algorithm that's great but at the same time somebody else is therefore losing for that and then they have an incentive to beat you so i think there's a lot of things about the financial markets in terms of coming up with alpha signals it comes up coming up with performance signals um that is difficult um you know i think if you're looking at machine learning as a way of pulling out so things to do with the the sdg the sustainable development goals that's a very different question and i think that's a you know that's a fantastic area for that um i mean i find the other area that i think you know what i talked about earlier the the advent of data and computing power um has also done is allowed people to do a lot more with bayesian statistics so this idea of trying to quantify some of the uncertainty uh in in you know but also take your views into account now interestingly now there are some areas which is not one i've had terms to look into yet where machine learning invasion statistics are now combining so you know this is sort of the idea of not coming up with a point forecast but coming up with a distributional forecast um out of machine learning so i think my point with really the the the sort of sort of question or this area is i see nothing wrong with machine learning it's a great tool it's another tool and i think that's the very important point um it might not be the right tool to solve every single problem in fact it isn't the right tool to solve every single problem i can tell you that um and again i think the other problem with machine learning and i saw this many years ago and it keeps coming up is it's very hard to know when it will go wrong um you can't train a model to tell oh sorry i'll be careful uh it's harder to trade a model um to tell you when it doesn't know because you can't give it any data you can't give it the data to trade on because then it doesn't know so you sort of this weird thing of how does it know when it hasn't seen something and that's really difficult there are a lot of underlying geometrical issues with high dimensional data that make that problem very hard to solve so i'm not saying don't use machine learning please please don't take that away from what i'm saying here i'm just saying it's a useful tool it's very powerful um but it's also com you know it can be very untransparent although again people are trying to work on that so i think the other thing as well is and i'll then stop is people seem to confuse machine learning and alternative data they're always talked about often in often the same sentence and they're two different things alternative data is fantastic you know it's another way it's another source of information about the world and you might want to use machine learning to process it but they are two separate things um so uh i guess as i say i i always come over as being sort of maybe slightly skeptical i think you know i think what's been done with machine learning is fantastic uh when i was at ubs back in 2000 um we published a very successful piece of research called neural networks as they were as deep learning was then called it it's it's well sold out in the sense that back then pdfs hadn't been invented so we were doing it on paper and literally we had to reprint it it was so popular and then of course it all died because the computing power wasn't there to solve to really use the technology back then so you know the the different technologies come up different people come around um but you know as i say i think machine learning is interesting but take it as being one of the tools in your box but it's not the only tool in the box i would love to hear from my panelists given they both work for ai companies or equivalent i'd love to see what they think of questions i'll go first uh yeah i think you know a very important point that you mentioned and that is not given enough attention is that markets are a are not a control system really where it's a game it's a differential game if it's looking at your continuous time it's an end person partly cooperative partly non-cooperative game the standard paradigm in machine learning is that you have perhaps an unknown distribution that you are trying to approximate uh in the in you know whether you're talking about bayesian machine learning or frequent machine learning there is out there some kind of a unknown distribution that you are trying to approximate however the game theory problem is slightly different because the unknown distribution is not given to you by nature but it is a product of perhaps strategies mixed strategies adopted by other players in the game so what you are looking at is a sub topic if you will of machine learning where you are trying to design a mechanism which can play a game right against opponents were also trying to win so it's more like uh it's not a two-person game but you could think of it as a two-person game so it's more like trying to play a video game so kind of thing that deepmind does uh playing a video game playing go so it's in the game playing ai machine learning camp rather than the more traditional uh you know trying to you know classify images or you know the kind of thing that deep learning does very well and now deep learning does very well on in game playing also but i think that most of financial machine learning has not gone that way they have gone the other way where the the distribution is unknown and perhaps changing but it's not a product of strategic choices by other players in a game and that requires a different kind of machine learning which is what needs to be uh looked at when you are especially in certain areas where you can assume that the number of players who matter are not that high if you had atomistic players who have you know you have millions of players none of whom have any effect on the game and it's probably not a stretch to assume that there is an unknown distribution if you have maybe 10 or 20 players who are big you know investors and can actually change which is actually true in a lot of algorithmic trading you do have fairly large sized participants who can change the distribution in a short-term sense at least and that is the kind of machine learning that you need and i don't know if a lot of people are working on it if they're working on it they're probably not going to tell us so once we find out that there is a you know there's a shop somewhere that's making a lot of money using algo they're probably using something like this so that's that's sort of my take on what you said i agree with a lot of what you're saying that uh you know people confuse machine learning and alternative and big data they're two different things altogether people think that uh you know you need point forecasts what you need really is distributional forecasting so all of that i do agree with that i add to that something you which you also refer to that you know it's a game and let's treat it as a game thanks right over the return yes so um i think those those are great points and i would like to um just uh add to some points that david mentioned that um i think it's important to go back to the basics and and basically really focus on the quality of our data because i think the part the the part of extracting filtering and cleaning data is is is far more critical than what people think especially when we start looking into alternative data all this unstructured noisy data with a lot of sparsity it becomes really important to to make sure that the quality of data is good and another major problem is the issue of overfitting in machine learning models sometimes people get very excited and they think oh you know i learned this unsupervised model little network or something and let me just plug the data there and see what happens so i think it's very important to you know go back to the basics start with linear models that we can look understand as a white box understand the parameters you know look at auto sample performance um and and use it as a benchmark before we start even talking about any fancy modeling in in machine learning and i think this is very important because um i will argue that um sometimes you can think the quality of the data is more important than the um you know the complexity of the model itself why because if we go back to the basic you know if you have good data you can expect a better income but if you have bad data doesn't matter how fast your sophisticated your model is you're not going to get good results and i think i will i will start on that all right thank you so much for uh these amazing comments and this particular question about machine learning now there is a question in the audience asked by kennel so kendall asks my question is to david you mentioned that data which we get needs to be refined what are the steps involved to do so i i gotta be careful here i don't it's not necessarily refined it's it's cleaned um now this is a this is really tricky um because people talk about outliers now outliers can be true you know they take take a sequence of stock returns we know they are give or they they're sort of sort of normal but they're not they have fat tails we all know this um so simply trying to say oh that was a sort of a big move uh is really tricky um so what i'm sort of saying here is it's amazing i remember actually my wife did a project she used to work for one of the big hedge funds here in london and all they were trying to find was the days of uh ex dividends which you know what day had a company gone exxon so they'd taken data from three different sources and compared them and in some cases they had three different dates for the same dividend payment and sometimes different dividend payments for the same dividend payment now that's sort of obvious in a sense you can do that very easily it's a very low set of low volume set of data and yet it was still hard work to actually get the right numbers and get the right to get the correct data so a lot of this is you can sort of automate finding the data points that are tricky but trying to fix them is often very hard because you need to sort of generally find a second source of data to compare them to and if you're going into tick level data this becomes really difficult because you just can't look at it you know in the problem i was talking there you're looking at sort of two or four data points per company per year that's not a lot of data you can do things visually with that um with high frequency data it's just impossible so you have to have sensible ways try and think about ways of making you know checking that the data is makes sense um and it it really does depend on the data set i can't really give you a lot of um a lot of sort of straightforward rules the big one though is pl is don't just assume the data is wrong you know you get weird returns in financial series you get strange things happening um you know especially if you look at volume data volume data can have some massive spikes in it and often when you go look into it it turns out the company made a news announcement that day i'll link that to an alternative data source of news announcements and all of a sudden you can you might be able to find that um so it's tricky as i say it's really hard but what we i've tended to find sort of over the years is is finding data errors is really straightforward it's the fixing them that's that's tricky and actually i'd say richard i i know somebody was you were saying you've been doing a lot of work on sustainable data which i commented on earlier being absolutely rubbish um i'd love to have you what you found with that yes yes sure so um i mean the sustainable development data goals is actually an ex um an expanded version of the environmental social governance factors and and yeah you see a lot of sparsity and a big problem is that sometimes the units are different so for example even when we speak about carbon emissions these different and other environmental factors is different standards used by different companies and another issue is a synchronous issue you know reporting comes at different times and and this becomes a problem so in those situations this is really important to look at the data from different angles and and in some cases you have to aggregate at lower frequencies because um you you also have to deal with that a number of methods to deal with the sparsity um of the data so it becomes very challenging of course um and there are different methods that can be used to address this however having said that there are um how they say um material signals out there that have a shorter impact on companies valuation and those are interesting to to think about and there is work done previously before esg became more fashionable about political risk um and we see different methods to quantify that using keywords and some basic nlp methods and we see some of that has worked to some extent so um but just to uh come back to the original point as they mentioned it is uh it is probably one of the most sparse data sets alternative datasets out there and however i think this is an opportunity long term but in the short term it just becomes very challenging to work with yeah all right uh moving to the next question now this question is asked by alan now it's very relevant to the previous discussion that we had and it is for uh dr gufa and uh so anand asks how is gaming solution different from traditional problem solving sorry i could you repeat that question so how is gaming solution different from trading uh traditional problem solving i think it's not traditional it has to be trading problem solving i'm not sure i understand the question i will answer the question as i understand it uh trading solutions and uh you know uh modeling solutions are not necessarily the same because trading solutions depend on as i have mentioned before they're part of a game structure so all trading solutions must be thought of and tested if possible in as part of uh defining a strategy in an end person game where uh there are opposing players who are who have antagonistic and sometimes cooperative because you can form coalitions so the modeling method that is usually undertaken is as i said before not really right in when you have a number of large players if you have a large number of atomistic players then you know it probably is okay but in commodities trading in crypto trading you do have uh more of a game structure and this game structure must be kept in mind before uh before you design any trading system i hope that answers your question norman all right so moving on to the next question uh so it is to david um now we talked about machine learning so the next question i have for you is which markets are most interesting um i mean i think in a sense we we've touched on this already um i think this is going outside of equities i mean the equity markets are um you know that's really where people have been doing a lot of a lot of uh both traditional and quantitative trading um both at higher and lower frequencies so i think it's the to me the interesting areas to go into now i mean sorry i should say currency also has been sort of you know there for many many years um i think he's going into the other markets uh i i think crypto is fascinating i know very little about it um but i think there's some you know very interesting things you could do there you have interesting challenges you know none of these markets have been around for very long so how how much data can you get um and you know this is the ccpa sort of more and more of these crypto assets appearing um so i think that's a very interesting one i i did love the comment about water as a traded asset um a dreadful pun of uh i'm assuming it's fairly liquid did spring to mind whilst uh dr was talking but i i i thought i would share it um and so i think just all the all these other asset classes the asset classes which haven't been sort of mined to death um in one way or another or i think the other one would i'd stress into is markets where you can get some interesting data so if you can come up with an interesting alternative data source ideally not one you necessarily purchase but what you can create yourself um that becomes an interesting market yeah and that that's that's so there is really the data source becomes the the driver of interest um so you know if you have data that nobody else has that's brilliant and that's definitely where you you should be thinking or definitely where you should be looking um and i think this changes and i think that's that's the final point i'd make with this is is don't just focus on one asset don't just focus on one algorithm for trading with you know you look at the most successful people in any of these sort of fields they tend to run multiple models they're very good at risk control uh they're very good at giving up on models um you know when they seem to stop working so i think those would be my main points i think you know but to say moving away from your these very traditional asset classes unless you have something some sort of some form of edge very well so on a similar line we have also a follow-up question so i thought i might bring it up so from your experience can you tell us some of the markets or asset classes which are generally better outcomes with machine learning algorithms i think that varies over time i i i think that is very difficult i think the answer in the end is market you know there's a sort of markets go through phases of being more or less predictable um and i i i i use the word predictable loosely here maybe tradable um it profitably tradeable uh any market is tradable um and i think they go through phases of that and i think you see this with different signals uh you know again going back into more traditional sort of quant signals um people used to look at first people looked a lot at sort of changes in consensus earnings numbers um and then that got done to death everybody was using it the amount of sort of information in that signal really died down to zero and then interestingly it came back um because because not many people were looking at it um you know it came back as a signal that people were using so i think that's an example of how individual yeah that's a very traditional signal but it's just an example of how these things vary over time so i'm not i don't know it's probably an interesting question is there something about the market that makes it more friendly for um machine i mean i think the answer if you want a formal answer is effectively the underlying data generating process isn't changing you've got a relatively stable world going on which probably implies a lower volatility market but it could be higher volatility just but but the underlying sort of call it the statistical process that you could think of as debt generating that the market is isn't changing so the the machine has something to latch on to and to actually learn now can you quantify that i'm not sure i mean i it's an interesting it's an interesting thought is you know what properties are there about markets that make them more or less profitable um but i as i say you know a lower but not too low volatility would probably be the starting point very well all right so we are talking about markets right so now interestingly i would now know would want to talk a bit about cryptocurrency market so uh to dr gughah now what are the main obstacles faced in deploying quant strategies for crypto investing yeah i've been looking at crypto a lot and uh one the principal problem really is that traditional finance people have never taken interest in crypto the uh you know the banks and even the hedge funds mostly they stay away from crypto crypto is usually thought of as some kind of a you know computer geek video game kind of uh field um and you know the talk about uh you know the byzantine generals problem and so on that crypto comes from uh reinforces that uh stereotype and that is why i think that you know the the financial world the traditional financial world including the alternative funds have not really taken as much interest in crypto as they might have in a different scenario the other part of it is of course the extreme volatility uh if you look at you know the major cryptocurrencies bitcoin ethereum uh in the last six years some of them have been down we're talking about peak to trough drawdown of 85 to 90 percent now this is unthinkable for uh you know any kind of fund manager or any kind of retail investor who's going to sit through an 85 drawdown nobody in the uh traditional uh there are people who the crypto guys call hodu hodl which stands for hold on for dear life and apparently it was a typo someone wanted to say old type hodl and then they backfitted acronym on it but it has become a strategy so to say in the crypto world where you sit through 80 90 drawdowns and eventually come out ahead and you do come out ahead in all the drawdowns have ended and the returns the cumulative returns from crypto have been very high in all the major currencies uh all the major coins that is so i think what is what has to happen for any kind of uh financial trading ecosystem to develop in crypto that is away from the traditional crypto ecosystem uh is first of all for finance people to recognize that it's it is an alternative asset class maybe it's not a currency it's certainly not a commodity uh but it is its own asset class and eminently tradable there's a lot of liquidity and the exchanges are kind of a bit dicey right now but that probably can be solved uh but the problem that needs to be solved is that you know you're not going to be able to sell any kind of product to anybody if you uh have 80 drawdowns 90 drawdown it's just not going to happen so you need something to damp down these drawdowns you need to find a way to create a portfolio that isn't going to be down 80 percent and still have the outside at least some of the outsized gains that crypto offers and the solution as always turns out to be diversification uh however it is a slightly different kind of diversification and um i can talk about it uh you know and briefly um the idea is that if you just you know have a fixed proportion diversification sort of the markovic's variety it doesn't work because all the crypto assets tend to go down together they're very highly correlated especially in their bigger swings so you need to do something about that and what you need to do is something that i think needs to be adopted across the board not just in case of crypto but for all assets and that is that we need to give up this idea that uh assets follow you know e2 diffusion processes with fixed mean and variance when we're not fixed but at least uh non-stochastically varying mean and variance this is an in an embedded assumption in virtually all of one finance we start off by saying that let you know let the stock price be you know ds over s equal to mu dt plus sigma dw and then you know if it's a derivative problem then we have to apply it towards lemma and you know that's the standard thing that we all learn in graduate school and uh in the quant world in risk monster risk platforms end up assuming that it turns out that that is not really a very good approximation to the actual and we all know that we know that volatility is stochastic that mean changes quite a bit over time and there is a whole set of research that says that there are big swings in commodity prices in currency rates even in stock prices that can best be you know expressed as a mixture of normals driven by a hidden you know markup state but this is not standard uh theory and if you apply this theory if you say you take the basic model you have one risky asset one risk-free asset logarithmic utility what do you get if it's a ito diffusion process mu over sigma square is the constant proportion portfolio this changes it becomes mu 1 square mu 1 over sigma 1 square and mu 2 over sigma 2 square or nu i over sigma square if you have i states with fixed mean and standard deviation that means no long no buy and hold anymore that means tactical tactically shifting your portfolio every time the state changes so this is a big jump in terms of you know what you what people believe in because you may not believe in buying and holding just the market portfolio but even if you have a factor portfolio you're holding the factor portfolio you're buying and holding the back to portfolio in a fixed proportion you're not tactically moving out of value to growth or whatever and this this is something that is in the data there is considerable evidence i myself am preparing a paper on uh you know mark of state jumps in uh in the form of french factors which it seems has been overlooked and the the portfolio advice changes so if you do that if you have a portfolio that shifts it's holding according to a a hidden state that needs to be estimated from the data uh recursively then the diversification works and you get uh returns that are quite high and volatility that's much lower and drawdown that are much much lower so the diversification works as long as you don't assume that the uh that the mean and the standard deviation is fixed but you assume that the they are driven by a hidden markov process so that's i think uh this if this becomes standard methodology that people believe in and people are willing to embrace then i think not just crypto but certainly crypto because the swings are huge in crypto but in other areas also for instance uh in uh you know factory investing there's a big debate in factory investing circles whether you know tactical methods work whether you should move in and out of certain factors into other factors or bail out of certain factors which are not performing well and this basic understanding of the uh of the uh stochastic process that are driving the returns i think can make a big difference in how you approach both factor investing and crypto investing and crypto investing does work on you once you do this you get you know very very enticing shall we say uh risk return profiles out of simple diversification as long as the diversification takes into account this basic uh structure of the stochastic process driving the returns i'll stop there thanks that explains that what needs to happen in order to cryptocurrency to get into mainstream and accepted by the traditional financial markets and quants all right so i think uh the next question that i have uh here um but before that and because then it will discontinue so uh the interesting question here is by scyther he asked what are the various applications of machine learning you do daily now we can quickly answer this question and then we can move on to the next one but uh yeah david would you like to go first um we at the moment we don't do a lot we especially if we were talking um very recently about um using this actually again partly in in with the background of being in um a more fundamental uh you know investment house uh of using some sort of various machine learning techniques for sort of sex classification grouping companies together um are we wanting to use it for uh actually looking at it in terms of both forecasting carbon emissions for companies um and as part of the whole sort of move towards a lower carbon sort of type of portfolio um and also fixing some volumes i mean one of the problems of the fixed income markets is trying to get decent volume data and get back to data and you know we've been again trying to say can we use some sort of machine learning techniques to think about those sort of things so those have been some of the problems um i mean there's a lot more sort of within sort of around us as well looking at various machine learning techniques so we were chatting a different team to mine about using them for uh real estate investing so oh can you figure out which cities are this particular us team we're doing this uh which cities are the best ones to buy property in um and that's the they've got an underlying thesis about why that works but um so yeah no there's a lot of things going on within sort of columbia threadneedle uh using machine learning techniques not always not often often as part of the investment process not really as a a trading technique but much more as an input into the investment process um and so i just give you a few examples there of things either we're doing or other people within say columbia threat therapies are looking at thanks david for that so i hope that answers your questions psy so moving to the next question now this one is for richard so we talked about esc and uh uh sustainable development goals and all so what are these sustainable development goals at united station united nations conference on trade and development can you tell us more about it um yeah so the sustainable development goals it's an evolution of the esg and this is because esg is financially is essentially focus on shareholders so they might there's a matter reality framework which means it tells investors what factors or environmental have the biggest impact on on on shareholders but now we're moving from a shareholder to a stakeholder framework where it's not any more just about shareholders but it's also about society in general and how it impacts the world itself and this comes to relevance because we see a lot of systemic risk and spillovers particularly in climate where you know this building up of systemic risk can have a material impact across you know um all asset classes globally so this becomes uh critical um for global investors and even uh report by swiss re and major insurer say that um more than 30 percent of the assets globally are exposed to um climate climate risks and so this is very significant and so the sdgs is one attempt to basically refine the taxonomy of risks that we're observing so traditionally esg we focus on just primarily carbon emissions so now with the esdgs this taxonomy has expanded and we look at other factors beyond carbon emissions and also we look at governance and and other factors that are material and if you look at this from a financial perspective these can become risk factors the same way you think about a factor model you can you can think of it as a esg factor model and this can apply to each letter individually and of course this can be clustered as well because we have 17 factors and now um some part of committees and task forces and this has become a major uh um trend and an interesting um trend i see as well is the topic of sdg footprint and this is and this is close related to impact finance which is where we typically report impact based on service data basically a company says i'm doing this or a font size i'm investing on this so but this is self-reporting bias so-called green washing so how do we address these so-called fake news so one way is to look at uh again alternative data and essentially look at what other sources beyond the company itself or the investor itself are saying so for example if an investor or a multinational is invested in ecuador or south africa can we use for example natural language processing to understand what the local ngos the local news the local stakeholders are staying in their own language in the native language and uses taxonomies to categorize and quantify for example the sentiment of these data into positives and negatives to better understand the footprint and we we use some examples of united nations to uncover hidden risks that were not originally reported especially in english why because we see a trend that when you look at mainstream sources in english like financial times what's your journal you only see a subset of the news actually that are relevant and there's a lack a time lag to it so if you want to look at early warning signals you will have to look into unstructured data sources in multiple languages and so this becomes a bigger problem that we started talking at the beginning but it also represents an opportunity to better quantify these non-financial risks that are more and more material to the intangible aspect of the balance sheet which most companies is the biggest so what is your reputation another way to think of sdgs and esg is reputation so this this is where machine learning can be useful and i know structural alternative data play an important role to help investors um not just gain an edge with alpha but also have a better risk management a better process of uh selecting stocks for example and building a portfolio yeah thanks for explaining that richard thanks a lot so okay now moving on to the next question now this question is about the future right so it's combination of uh you know uh where does the future stand and uh what to expect in future or for that matter how to prepare for the future and we have also received a few questions from individuals like keith who wants to understand like what foundational knowledge would you suggest to best position oneself to take advantage of the trends that you've spoken today now so i think uh this is going to be the a very interesting question from the career point of view as well as you know forecasting the how exactly the future is going to be for trading or quantitative trading for that matter so i would want comments from all of you on this one so uh we can start off with david david can you share your thoughts on these um oh forecasting the future is always difficult to quote a phrase um [Music] i think there's a few of this a few things here i think knowing uh you know we all need to know good programming skills and that's not just problem solving programming that's actually sort of knowing how to develop software in a sensible way and make sure you you've got good tests around that and make sure if you change things you're aware of what goes on and it's these things can really catch you out you know you you develop a bunch of different software you change something and then you forget to have version control so you can't go back to a version you had before um i think also some statistical knowledge one of the things i see a lot within the whole data science area is people sort of reinventing the wheel and one of the biggest areas i've seen and has become up in in traditional finance world fairly recently has been a lot about multiple testing um and being aware of the pitfall statistical pitfalls of basically looking at the same data set multiple times because if you look at it long enough if you to quote a phrase and i can't remember from whom if you torture the data long enough it will tell you something um if you torture a time series long enough you'll find a trading strategy that works um but so being aware of that is is very well and these are all there already um i think keeping abreast of of how new asset classes are developing i mean crypto as i say is the obvious one at the moment um but i think the so to say going back to the comment about water earlier i think other fundamental things like that could become tradable could become things people want to trade globally carbon is obviously a sort of been around for a little while with that so i think none of that it's very hard to sort of say is there anything absolutely new here i'm sure there will be um and one thing we i have been looking at on and off has been quantum computing um which is trust me if you try and figure out quantum computing you you will get confused um and i could there i will quote richard feynman who said you know if you actually think you understand quantum anything to do with quantum you don't um but quantum computing could completely uh change everything i mean if if you get true quantum computing the whole of the uh encryption algorithm behind bitcoin is crackable instantly uh there are other cryptocurrencies uh uh crypto assets out there which don't have quantum crackable uh encryption um but to see that i think is one thing and a few a few very big hedge funds are investing quite a lot into the quantum area um so that's one i think to be aware of um there's nothing there yet there's nothing practical really coming out of that yet but it is there um yeah to be honest i think the impact of quantum computing if you want the honest truth will be much bigger in material sciences and in in pharmac and pharmaceuticals because it'll allow massive sort of automated development of drugs but and materials but it is there um so there's a bunch of things there one is sort of basically in terms of in terms of the sort of people listening to this is do you have these basic stuff there get the basics right you know um and then on top of that be aware of these this sort of interesting stuff uh that that's going on out there but um i will warn you if you start looking into quantum computing it is a rabbit hole into which you can vanish quite quickly um and get yourself both both very very sort of entertained and very enlightened so i will pass on to my other panelists dr google what do you apart from quantum computing which just certainly uh you know promises to be a very big development in the future the other you know major development in the ai front is the development of very large neural networks and gpt3 which is you know created quite a stir uh gpt iii is a essentially a nlp natural language processing uh deep learning network which has i forget how many trillions of uh parameters uh or maybe it's billions and there is gpt-4 that's coming soon that will be even bigger and the hope among certain ai people and i'm not sure i count myself among them is that these very large models will point us to a way to general artificial intelligence that other people are saying that no that's not the way to go common sense must be built in somehow and we need to go beyond uh you know we went from deductive inference to inductive inference and we must incorporate abductive inference in order to uh have truly common sense uh general purpose intelligence but i think that you know the the lesson of uh the development of ai has been so far at least is that increases in the capacity of hardware and software's formed everything else deep neural networks were written off in the 1990s and even in the 2000s because they didn't perform very well and then suddenly with in 2012 they started performing very well to the extent now that you know you've practically given up on all other kinds of machine learning and deep learning is going to take over the world it's going to be the you know the final uh iteration in all of this so uh one thing is very clear that you know the increase in the capacity of hardware and software that we have seen you know no matter what people have expect it hasn't stopped and it doesn't look like it's going to stop anytime soon we're always being told that we're running against running up against some kind of physical limit but we haven't gone there yet and if quantum computing turns out to be practical which it may uh you know in the next five to ten years then of course you know all bets are off you know we have past new capacities so i think that the big thing that's going to happen is that continuing expansion in hardware and software is going to make general purpose trading robots possible things that can extract data from the surroundings including social media and other media have enough sense to to make sense of it and take trading decisions based not on just numerical machine learning but uh you know image understanding language understanding understanding of uh you know the the semantics of uh you know media media compendia and take a decision that is uh and you know completely outperform human traders in the future the way chess programs completely outperform even the world champion and go programs completely outperform the greatest go player of all time and i think that that is going to be the future whether that comes in five years or ten years is something that i have absolutely no idea but uh it is going to be there and it's not too far in the future maybe 5 to 10 years that's my guess this is going to happen and you know people that haven't worked with gpt3 for instance they will find it quite fascinating i think what this uh piece of software can do you give it a prompt and it can write a whole paragraph it can write uh you know james bond stories in the style of ernest hemingway it can write harry potter stories in the style of pg booth house so it can do a lot of things that we didn't think a a large neural network can do even if they are very large and very deep uh but it can do it it can do it and you know the capacities are increasing every day so just the expansion of hardware and software capacity may lead to uh general purpose trading machines that are as we say in ai superhuman in their performance by superhuman we mean way ahead of any human performance that that's going to be my uh prediction for the day uh as uh david has already pointed out uh neil sport has already told us or was it max planck that prediction is difficult especially of the future richard yes um so i will i will uh echo some other points mentioned earlier and i will mention um i i i think it's the future is basically a convergence of these points i mentioned earlier and interestingly at the lawrence berkeley national lab uh there's been some experiments of benchmarking for example a supercomputer versus a quantum computer and there's also an interesting because it becomes in some specific specific problems uh low dimensionality you have a performance of quantum of problems that will take years or millions of years of a computer they will be solved um in a much faster time frame and so one of the challenges is basically that's been addressed now is the development of an operating system for the quantum computers and this is in progress today and there's been some interesting use cases in finance even for example for portfolio construction and as as we evolve to enable our city future is not just quantum itself but it's a combination of today we have cpu and gpus in the future it's going to be a combination of high performance competing and quantum working side to side and when when we start to incorporate more data and more models i think the future lies in the convergence of these high performance computing slash quantum and ai and all these non-linear models that combine with large amounts of data at a working at higher speeds that will basically make many functions obsolete um across the human space and we're talking about many jobs so it is good to you know uh you know prepare and and start to um you know accelerate the incorporation of um new tools and techniques as they become available because i think this process is going to grow exponentially fast and i don't think it's 50 years i think it's more like what dr woops has said in five to 10 years we should expect to see a radical transformation across many functions and industries and and we should be ready yeah thank you thank you so much uh uh our pants thank you thank you so much to our panelists uh for uh taking out time and explaining these deep insights i mean these are very helpful especially for all the aspiring quants out there this would have been improved a very good points to start their journey on and i hope this will help them a lot and so again a big thank you to all of you for taking out your valuable time and joining us here and yeah that's about it and we are going to conclude this session now for audience those of you whose questions are not picked or not answered today please make sure that you share those questions at the end of the survey with us and we'll make sure that all of your questions are going to get addressed with that said thank you everyone and thank you david dr guha and richard for your time thank you thank you very much yeah you take care and have a good day ahead thank you all right for the audience uh i would like to share that tomorrow we are going to start a bit early so we will be joined by dr thomas stark tomorrow at 2 30 p.m indian time and uh uh this is going to be a very interesting session it is going to be a q a on machine learning and trading and uh do not miss it uh and i'll catch you tomorrow in that session thank you till then have a good day bye you

Channel: QuantInsti Quantitative Learning

Views: 550

Rating: 5 out of 5

Keywords: learn algorithmic trading, algo trading course, Algo Trading Week, QuantInsti, Algo Trading Week 2021, sentiment analysis trading, alternative data, sentiment analysis

Id: 7K9W6Xrybbs

Length: 67min 50sec (4070 seconds)

Published: Tue Sep 28 2021