QuantConnect Open Algorithmic Trading Meet-Up #6
all right um i think we can get started hi everyone my name is lexie and i'm with the marketing and pr department at quant connect thank you to everyone for joining today welcome if it's your first time and welcome back if you're returning this is our sixth monthly meetup we hold them on the last thursday of the month at alternating times to accommodate our friends from around the globe the schedule for today will begin with one of our quants giving a presentation on his research into the value and momentum everywhere strategy we hope you enjoyed the presentation and if it sparks any research ideas or trading hypotheses we hope you'll join us on quantconnect.com if you're not joined already so with that um i will let derek take the wheel hello everyone thanks lexie so my name is derek melchin i've been a quantitative developer at quantconnect for 10 months (video outage) to the community forum uh directly after this presentation is done so i recommend looking for those i've organized this presentation to be sort of beginner-friendly uh there is gonna be a lot of code that we're going to go through i'll go through it as slow as i can and as clear as i can um but again reference to the files that i upload afterwards to get a better understanding um if at any point um any of these slides are confusing feel free to unmute yourself and ask a question uh but i would like to save the bulk of the discussion for after the slides are done uh the slides will only take about 25 minutes to get through um so let's get started um so this is the research paper of the strategy that we're going to be looking at today it's called value momentum everywhere one of the authors is clifford from aqr capital management the objectives of the strategy are to capture both value and momentum anomalies the cool thing about these two factors is in the research paper they show that both factors have been shown to be profitable over the past and they're both negatively correlated so when you put them together you get a nice smooth equity curve in theory we're also with the strategy minimizing selection bias and we're doing that by using commonly used factor measures to define these factor values so for instance the value factor we're going to be using the book to market ratio which is very commonly used in the literature and for momentum we're defining it as the return over the last year excluding the most recent month which is a again a very common definition for the momentum factor with this strategy um we're able to diversify across asset classes and also maintain a dollar neutral portfolios having our portfolio long and half short is there a question Member: yeah derek uh thank you for letting me ask a question first should i wait or should i ask my question during this time um if it's related directly related to the slides you can go ahead now but uh we will have a large block of time after the slides to have an open discussion Lexie: Derek are you meaning to share a presentation Derek: sorry did i not share my screen Member: yeah that's the first thing i was going to ask about Lexie: sorry i meant to jump in Derek: sorry let me just uh fix that here Derek: there how's that Member: yeah that's better perfect okay Member: and um also in your presentation will you show the mappings between momentum and the definitions that you have Derek: yeah yeah we'll be going through that for sure so this is the slide that i was showing just a moment ago this is the research paper and the objectives uh this is the roadmap so i've broke this up into four different parts we're going to build up the strategy incrementally so first we'll apply the value factor to u.s equities we'll then adjust it to have uh just the momentum factor on u.s equities and then we'll combine the factors together to see how they perform and then as a part 4 i've kind of left it as a take-home exercise it's going to be applying both of the factor values to a collection of universes instead of just us equities so part one where we're applying the value factor to the u.s equities the main idea is each month we're going to be selecting the largest most liquid u.s stocks that are above a dollar we then rank all of those securities by their book to market ratio which is our value factor i've included the formula for the book to market ratio here it's just the book value per share divided by the market value of those shares once we've ranked all of our securities we invest in all of the securities in the universe and we're going to be longing securities with a high value factor and shorting those with a relatively low value factor and we're doing this on a monthly basis so we're just going to hold from the beginning of the month until the beginning of the next month so how do we go about implementing this in quantconnect if we jump over to our ide here is kind of the high level overview of the algorithm class we've got our initialize method which is where execution starts um i've removed the boilerplate code but uh we're starting with adding a universe selection model and we are creating a custom universe selection model which i've called large liquid us equities we'll look at the definition of that on the next slide but on this slide the idea is that whenever a security enters or leaves our universe this method on securities change is called and what we're doing when securities uh come to or leave our universe is we're just creating a symbol data object and saving it into a dictionary so we can use it for later this is the definition of the large liquid us equities class in the initialize function at the very top we can see that we're specifying a minimum price of a dollar and we're setting a month to be negative one that we can know one month the month changes that we can refresh our universe the universe selection starts with the select coarse method if we're in the current month we just return the same universe that we've had that we selected before but if it's a new month what we do is we select all the securities that have fundamental data and have a price above a dollar we sort all those securities by dollar volume and we return the top 500 based on dollar volumes so the most liquid us stocks once we've got those 500 securities this select fine method is called in here we are following the research paper and we're restricting our securities to just common stock securities and we're removing any adrs reits and companies in the financial services sector once we've removed all of those securities from our universe we sort by market cap and return the largest 50. um you can increase this universe size as you like i just chose 50 so that we can have the algorithm complete in a reasonable amount of time but ideally the larger the universe with the strategy the better so that we can diversify across more securities so that's the definition of our universe selection model um what we're now going to do is take a look at the definition of this symbol data class these are the objects that we're creating for each security that's in our universe and when it comes to the value factor it's pretty straightforward uh the symbol data class just needs to return our value factor which again is the book to market ratio so we're taking the inverse of the price to book ratio that's that's the equivalent we just don't have the book to market ratio in our morning star data so that's the definition of simple data and this is kind of the overview of adding securities to the universe and removing them so this point we've got our universe defined and we've created simple data objects for each security let's just switch gears for a moment to review how we go about rebalancing our portfolio we're going to do this with tables first and then we'll look at how it's done in the code for the value factor the first step is to create a table where we have uh the book to market ratios for each security in our universe once we've got that we can then rank the value factor and we're doing it with a positive linear ranking so that the higher the value factor the higher the ranking so we can see for example symbol d has a value factor of 0.4 and it has the highest ranking in the table below
so if we move this table up here we can uh then subtract the mean ranking and this is what makes the portfolio dollar neutral the idea here is to shift all the rankings so that they center around zero we can see here that the symbol c in the above table has the middle-ranking three and in the table below it's changed to a zero now if we move this up here we can then remove the leverage so that our portfolio has half of um or is long half of the portfolio is long and half of the portfolio is short as far as the code for the rebalancing goes here it is here so in our algorithm class again at the very top during our initialize method we are creating a scheduled event and this is just going to call the rebalance function at the beginning of each month and when this method is called we're just going to be gathering the value factors ranking them ranking the securities by their value factors weighting them making them dollar neutral this is all stuff we saw and just on the table or in the tables before we scale down the portfolio to stay within the leverage constraints and at the very bottom once we've done all that and we've got our portfolio weights for each symbol we can then place the trades it's a pretty straightforward algorithm as far as performance go i did actually not perform as well as it was shown in the research paper but keep in mind that the research paper was published in 2012 um and we're testing here from 2016 to the current day so it is out of sample data and uh this back test of this chart will be included uh in the community forum post that is made after this presentation so uh interesting note is that if you switch the way that we rank so instead of having the positive linear relationship so the higher the value factor the higher the ranking if we flip it so the higher the value of the factor the lower the ranking we then essentially flip the equity curve however we're not going to exploit this finding we're going to maintain the definition that was made in the research paper so we're going to stick with this component as our value factor we'll now move on to part two so now we're going to be instead of applying the value factor to us equities we're going to be using the momentum factor the main idea here is each month we select the same universe the largest most actively traded u.s stocks that are above a dollar we then rank securities by their 12-month trailing return that's our momentum factor and we're just excluding the most recent month from that calculation that's something that's done in the research paper and commonly done in other research papers as well again we're going to invest in the entire universe long securities with a high momentum factor and shorting those with a relatively low momentum factor once we placed our trades we just hold until the next month so as far as the code goes not much changes this is our initial algorithm class with our addition of the universe selection model and the management of the symbol data objects the only difference here is that when a security leaves our universe we need to dispose of the consolidators that are attached to that security the consolidators we'll see on the next slide here what they do is they just keep the technical indicators updated this is the new definition of the symbol data class if we are using the momentum indicator for the momentum factor when we initialize it we are specifying that the momentum look back should be 12 months and we have a delay of one month so that we can exclude the most recent month we set up the consolidators and indicators the indicators we use are rate of change and delay and yeah the consolidators um what they do is they just update these indicators as new data is provided for each security after we've set up the consolidators and indicators we then warm up the indicators by using the history method and then at the bottom here we can see that we have a new property called momentum which is just returning the value of the momentum indicator we set up an is ready property to return if the momentum indicator is ready and then our dispose method is at the bottom which is removing the consolidator when the security leaves the universe so that's the definition of our symbol data class when using the momentum factor when rebalancing again not much changes we're just adjusting from value to momentum so just changing that one word there all of the rebalancing logic is the same and here's the performance of that momentum factor portfolio uh it does underperform the s&p 500 over the same back test period but it is profitable nonetheless so we'll now go on to part three this is where we're going to be combining both of the factors the value and momentum factor the main idea we select the same liquid us stocks but now we're ranking them by two factors ranking it by the book to market ratio and the 12 month return and to get both of these factors into one factor that we can use we're just going to take the mean ranking of both of those factors once we have that composite factor ranking we can then invest in the entire universe once again along the series with a high composite factor ranking ensuring those with the lower ranking and holding until the next month our algorithm code stays mostly the same with the universe selection model and the management of the symbol data objects we just need to change the symbol data class that we have both of the factors included inside so we've merged the two symbol data classes that we've seen already so just as we had with the momentum fact we're setting up the consolidators and the indicators and we're warming them up we've got both the value property and the momentum property included here so now that we've got two factors our rebalancing logic needs to be updated to accommodate for this uh let's now review how it is done with a table and then we'll look at the code so here we've got a table of five symbols in our universe and we've gathered their book to market ratios and their 12 month trailing return just as before we rank them all with a positive linear relationship once we've got the ranking for each factor we can then subtract the mean ranking of each factor this is going to make it dollar neutral so we're just shifting all these values to center around zero and then we can combine the two columns together taking the mean and that is going to give us a composite factor ranking taking into account both of those factors and just as before we remove leverage by dividing by the absolute thumb of the entire column so now we see half our portfolio is long half of it is short now with two columns in this table there's one case to note if the original rankings are ever looking like this where we have one column in increasing order in one column in decreasing order then we don't actually want to hold any of those securities so what does this look like in code here is our rebalancing function it's the same as before where we're gathering the value and momentum factor values um and then we are we've added our logic here to create the composite factor ranking and that's the code that we just saw in these slides with the tables here's the result of combining both of the factors together and now for part four the main idea is to apply the algorithm we just reviewed of using both of the factors but instead of applying it to a single universe we can apply the same strategy to multiple universes and the way this works is we just apply a portion say half of our portfolio to one asset class or one universe and we apply another portion to another asset class or universe doing this requires multiple universe collection models and the actual implementation of doing this um again i'll share it on the community forum after we wrap up here today uh one thing to note is that the book to market ratio which is our value factor that isn't included for every security on quantconnect for example these global equity etfs don't have any fundamental data when that is the case we can replace the book to market ratio with the negative five year trailing return and in the paper that i showed earlier we're able to do this replacement because the two factors closely correlate together so as far as next steps what i recommend doing is taking the algorithm code that is going to be published into the research or into the community forum uh you can then come increase diversification by adding more securities or adding more asset classes you can test other methods of computing the value and momentum factors you can also test other factors as well you can swap them in and out as you please uh and another approach which was kind of interesting those in the research paper is when you're applying the strategy to multiple universes or asset classes what you can do is instead of applying an equal portion of the portfolio to each asset class something that's interesting to look into is to scale the allocation of the portfolio based on the trailing volatility of each individual asset class the research paper showed when they did it there wasn't a huge benefit of doing so but it's something fun to try and look into and see where it takes you so that's all the slides i have for you today i'll now be happy to take any questions and we can kind of open this up to open discussion Member: derek is it okay if i ask some questions now Derek: of course Member: first i want to thank you for your presentation it is uh it's interesting this and i appreciate it especially the way that you built up starting with the paper and then you talked about how each of them you followed the same pattern and so that was more easily to follow so thanks for that so let me just start with the biggest picture i can because i'm so i've i've done some trading but not algorithmic trading so the first question i have is how did all that you did perform against the s&p state like and when i say that when someone says how does your algorithm perform against the s&p i'm understanding that to be that there was i guess it's overall it's somewhere between eight and ten percent per year is what you get the s&p and when someone says that at first is that correct and what does that mean i want to make sure the definitions are correct here Derek: yeah that sounds about right eight to ten percent is is probably about the average return for the s&p Member: so but when someone says that does that mean the top 30 stocks you just randomly pick or you pick the best performing s&p or what does it mean to say that s&p is it across all 500 or what's that meanderek: i see so um the way i do it is i just take the standard poor's index the spy and then um yeah that just tracks the index itself so you can calculate the returns just based on that index okay and the index is is an average of all 500 Derek: uh it's it's market cap weighted so uh the larger the company is the larger it's proportion in Member: oh i see okay but somewhere between eight and ten percent and now comes along your your uh algorithm based on all that with those research papers and you said it performed better than the s&p 500 but now how much better is in terms of is this like like it was 11% is this per annum or when you said it performed better what what did that mean and what percentage was that Derek: so i started sorry if i misspoke but it actually performed worse than the s&p 500. um if i just go back to the slide here so this is um the back test when we have both the value and momentum factors and typically the way that we compare the performance of the strategy is by using the sharpe ratio so i've included on the left here this one's a little bit negative actually because the return over the uh the five year back test is actually negative um but the s&p 500 sharpe ratio over this period is about 0.6 or 0.7 but yeah you can just run a quick back test on quantconnect of buy and hold the spy and that will tell you the sharpe ratio at the current time Member: okay so as an example you're just showing us a a paper that somebody did some research and what was the paper's return versus your return Derek: so the the papers return uh did outperform the s&p 500 but again that was tested on data that was in the past so that the paper was published in 2012 right so there is some decay in the strategy once a strategy is published because there's other traders that are that are using it Member: i see so then it begs the question up forgive me if i'm missing something but why would i want to use that if it performed worse Derek: so this specific implementation um looking at this back test you probably wouldn't want to trade this live if it hasn't performed well just over the last five years but the framework that is included in this strategy of having uh two factors you can define you can define them however you like but being able to rank securities by two factors um and then create a composite ranking and then allocate your your portfolio in a dollar neutral fashion that is diversified across multiple assets and asset classes that's the beneficial part of the strategy the actual performance of the specific uh implementation we have here um you know this underperforms the s&p 500 so you could adjust this strategy and that's something that i invite everybody to do for example you could adjust as we showed in the uh the further research slide here these are all ideas on how you could boost the performance of the trading strategy but the main idea to take away from this implementation is just the framework on how everything is laid out so that you can customize it yourself and begin to research your own strategies Member: so that's an interesting question to me and if i'm taking up too much time just tell me to stop this this is basically you introducing a framework now when it comes to you know the world i'm used to which is web development there are web development frameworks there's um .net has its own framework and you know um there's javascript frameworks like um for example angular or react and is there such an analogy in the trading world Derek: um i'm not too familiar with the frameworks you're referring to but i mean the entire uh lean back testing engine i guess you could you could look at as a framework like that's all we're using here right is just we're using uh the lean api and the quantconnect data to build our algorithms but by framework i just mean how we've laid out the components of the strategy of having you know the universe selection model and then creating our symbol data objects which have uh their their factor values included in them and then having a portfolio construction rebalancing logic which just updates the portfolio every month Member: that makes perfect sense because what you've said the value of this is is the framework not so much the definition of momentum right exactly Derek: yeah the idea that i'm trying to go for here is to kind of uh get you on the right path so you can do your own research further down the line you know right there wouldn't be much benefit to me coming here oh here's here's a new strategy uh run and run with it and start trading it i kind of see it as it's it's more valuable to teach somebody how to fish and kind of you know give them their first fishing rod rather than giving them a bunch of fish Member: right i that makes sense to me so the framework's there and i go in and i define my things like momentum the way i think that they will work and then i back test isn't it that all makes sense to me and but one of the things i need to back up even further derek and say i'm asking myself if i i even want to do this like to me option like and let me give you an example that um so warren buffett uh made a 10 year bet with some of the best market traders that just a simple basically a a simple uh index was very low low um uh why can't i think of the word the amount that you have to put in to be part of it low something Derek: low fees Member: low fees maybe that's it there was another word i can't think of it um very low fees and he he basically said look i'll beat your uh all your fancy algorithms and he put a million dollars on it and he won and Jared: hey nick hey nick can um can can you hear me yeah cool yeah um i i just uh last meet up you were asking a very similar train of questions and um derek and colton didn't have the answers and so i talked to the team that derek wasn't in that stand-up meeting so he missed the answer so your root question that you asked last week last stand up as well uh sorry last uh meetup as well was about algo training and whether it's profitable so it's not that simple so firstly all businesses have different goals so you've got to start with that what is your goal and so for example some investors are looking for short-term returns and others are looking for long-term returns some for example are measuring their investment in terms of sharpe ratio or the risk adjusted returns not just the percentage benchmark so for example if you've got a massive drawdown of 60 like the s&p does then you're exposed to a 60 percent loss and who knows if that would ever recover versus a lot of time algorithmic traders are looking for a better risk adjusted return at the cost of the absolute return so maybe they'll accept say five or six or seven percent per year but if that means that i'm not losing 60 then it's a win so it's all about what your goals are but then despite that and regardless of people's goals the the the absolute question of whether or not our community is making money is one that we actually measure as an internal company kpi we measure um the percentage of the community absolute net profit of the community so we want to know that our community members are profitable and that their algorithms are making money and so we actually we measure that and we publish on the dashboard inside in the company and it's about um 95 percent of the days are profitable for the community versus the broader market where that's about um 55 percent of days are profitable for the broader market so that's how we just that's how we judge our success Member: well thank you so much for sharing that so it's so good to hear that there's a measurement of a kpi so you said 95 percent of days is that what you said Jared: yes so that's a that's how we're benchmarking it is the sum of all the communities profits every day so it's an aggregate number we don't know that 95 of the community members are profitable but in aggregate the community is net to profitable Member: okay but Jared: just to just to give more space to talk more about algorithmic trading specifically um i'll be happy to answer this offline but i don't know if it's really relevant to derek's presentation so maybe we could just you know leave it there and if you'd like i'm happy to answer more of those questions Member: oh absolutely i don't want to uh i don't want to sound um confrontational or and i completely understand that you're saying what's your goal if your goal is to minimize your risk and yes i asked but and so yeah making money yep so yeah i'll take it offline and i'll i don't know how to get a hold of you or should i just do it in the community or Jared: yeah i'll send through my email in the chat yeah for sure i'm very easy to reach easy to reach Member: okay thank you great pleasure MemberL all right Derek uh thanks for presenting um i just had a quick question i may have missed it at the beginning uh you said you filtered out financials um why did you filter out financials in the universe selection Derek: the main reason for that is that's just what the um research paper did i just followed the the universe as it was defined in the research paper Derek: well yeah that's that's very interesting um maybe since you know like you said once we strategies are kind of published they tend to not perform as well it may be interesting to kind of find that maybe the strategy only works on a very specific sector and then apply it to that um yeah that's just food for further thought i guess Derek: yeah that's interesting uh one thing that's been changed with uh the universe selection in in this implementation is that in the research paper the strategy is actually applied to 90 of the total market cap so we can take you know the sum of of the market cap of all the securities and then take the largest until we're at 90 percent but that's that universe size is just extremely large um so in this implementation you know we filtered it to 500 and of coarse universe selection and then we filtered it just to 50 um in fine universe selection so you could increase the size of the universe it just kind of slows the algorithm down a little bit uh i thought 50 was kind of you know it's small enough to keep the the universe the program fast um but large enough to kind of give us a good idea on how the strategy would perform but yeah increasing the universe would be a good direction to go if you want to diversify even larger Jovad: yeah thanks Derek: for sure Member: i have a question Derek: for sure that's okay Member: thanks um i was just wondering it's just so cool you're introducing this strategy and um i was wondering what are some um other strategies out there like just to build a framework of the various strategies and if these strategies vary based on the computational complexity Derek: yeah so um a lot of the strategies that i've created for the strategy library they're all sourced just from research papers published online and there's you know tens to hundreds of thousands of different research papers that outline different strategies so you can come across you know many different types for example there's momentum as we've seen value there's arbitrage uh mean reversion um one of some of my favorite strategies to look into is using alternative data so for example on quantconnect we have tiingo news so you can get the news releases of specific securities and then you can parse the news releases with natural language processing to determine if it's positive or negative sentiment and then place your your orders that way um but yeah there's many other examples some uh big traders use weather data to kind of track different patterns that are happening throughout the world and are able to place trades that way but yeah i'd recommend you know just kind of surfing the internet and there's you'll come across many different ideas possibilities are endless Member: thank you so much and since the algorithm requires ongoing um refinement is that something we have to keep up ourselves or like personally Derek: what exactly do you mean by ongoing refinement Member: um just um continuous improvement Derek: okay yeah i mean you could uh adjust the strategy as you go as it's live trading um but there are some drawbacks to that you could make your adjustments to just be overfitting on the data that you've already seen so that's that's a little bit dangerous um another approach is a walk forward optimization which just does your adjustments um automatically so a simple example would be you know we we use the last year or the last few years and we we find out what were the best parameters for our trading strategy over the last few years and then we apply those parameters to the following year we trade for that year and then we get to the very end we do the same thing what were the best parameters over the last few years and apply that to the next year so that's just the way you can do it automatically but yeah you definitely can change it by hand if you'd like Member: thank you so much Derek: you're welcome Lexie: um also somebody asked if you could track reddit sentiment um Jovad if you're still in here didn't you just do a strategy um on reddit Jovad: yeah um i did one of the versions is posted in the community if you just go to the community forums and you look up reddit um i do have a live version of it it's basically the exact same thing but just some minor tweaks um i just haven't had the chance to post that yet it's been a while um but you know hopefully sometime this week i'll find some time to post in the community but um yeah using alternative data is very interesting it's kind of what i like to do um you know i think derek mentioned that in the strategy library and within the documentation there are some very great examples of using alternative data that quantconnect does provide um but yeah you know if you're if you're looking to learn more about how to build those strategies the way i learned was just kind of going through the community and finding strategies and playing around with them Lexie: um and derek looks like there is another question that asks about some good resources to learn more about algorithmic trading and strategies for a newcomer Derek: sure um well we have a lot of educational resources directly on the quantconnect platform that's where i'd recommend to start we have boot camp lessons which you know they give you instructions on how to build your algorithm and then they actually allow you to uh write the code yourselves just in a step-by-step manner we've got tons of uh introductory tutorials just to using python in a financial or quantitative trading environment that's a very useful tutorial to go through uh we've got probably 50 or 60 strategy library tutorials which just walk you through how to go about building your own strategy from scratch the research to production tutorials are very interesting as well it teaches you how to research your strategy in the research environment and then translate it into the back tester um so that's that's where i would start um and then there's tons of other resources online whenever i come into a problem while developing you know i just throw the error in google if i need to and then some sort of resource usually pops up that allows you to get over the error message Member: that's really exciting so does that mean um with the tools from quantconnect basically all of us even me like i could start Derek: exactly anybody can start that's that's our core mission is democratize finance and allow retail investors to have access to institutional-grade data and back testing and research tools so yeah we have tons we have tons of resources to walk you through how to actually use our software and we have an active community that is always sharing different strategies and you know helping each other fix bugs and whatnot so yeah we've got everything you need to get on the right foot Member: thank you you're welcome Lexie: does anybody have any other questions Member: no but this is very interesting Derek: thanks for coming everybody Member: wait wait wait i've got a question Derek: for sure Member: um so you said you use many journal articles right uh what journals you know provide the highest quality uh papers because what has happened in the past years that i've implemented quite a few papers and then they just don't work out like this is an example of one of them so what journals would you recommend Derek: so the ones that i usually go to i just go to ssrn.com on there you know all the or most of the research papers are actually uh free to view i don't have any premium subscription to any uh individual journal itself um but yeah my recommendation is to just look at as many different research papers as you can um but i think maybe the the real benefit of of using these research papers is to just walk you through how to build a strategy i think you'll probably end up finding your best strategies just by doing your own research and coming up with your own ideas um but i think getting uh started with the research papers is definitely a good idea and for that i use ssrn.com Member: yeah personally research papers give you like a a good idea of what is available
so for example pair trading right so if you don't know that there's a correlation between let's say the canadian dollar versus uh you know the u.s dollar and wood prices um that's new information right and you can use that to trade um so that's what i use it for Derek: that's a great point yeah get get your you know your ideas from the research papers but then you know go and do your own research and see what works for you that's Member: i was wondering sorry i was wondering if there's time for another question Derek: yeah of course Member: thank you just really excited to get started um going through all the tools on quantconnect and i was just wondering if you might be able to outline just like any major technical backgrounds that i need to get started or does uh or all the tools just like will help me and i just need to get started yeah so um i'd say the biggest tool that you need is you need to be able to learn or you need to be able to program in python um and then you need to be able to know how to navigate our documentation as soon as you've got those two skills um you know you can basically take the ideas from the documentation and write your own algorithm with it um but yeah i would i would recommend starting with the bootcamp lessons um instead of you know just diving straight into documentation the bootcamp lessons are a good way to get your feet going at the very beginning Member:thank you so much Derek: you're welcome i love the excitement Lexie: does anybody have any last questions they would like to ask derek or anybody else Member: uh i've actually got a question i'm curious if you if you're familiar with any trading strategies based around like relative computational complexity and entropy and entropy based modeling for signal detection Derek: i'm not familiar with with those concepts now okay great means there might be opportunities there thank you thank you Member: i have uh one last question here Derek: of course Member: so if you have tons of data and you're looking for all kinds of correlations just like a blind squirrel can find a nut every once in a while you'll find correlations that aren't really genuinely uh you know to because obviously there's a difference between causation and correlation how do you weed out the correlations that are bogus Derek: so the idea there is to test not only on in-sample data but reserve a portion of your data to test your model on after your model's been trained um a great example of this is to train your model you know you could train your model in all the data you have available right now and then you could just run the algorithm for the next six months and that would be unseen data and uh if the performance degrades through that out of sample period you know that perhaps the correlations that you identified earlier don't hold going forward that's me Member: thank you Derek: you're welcome Lexie: all right well i think with that um we can go ahead and wrap it up for today thanks derek for presenting with us and to all of you for joining us as a reminder we hold these meetups on the last thursday of the month we hope to see you all again next month to develop and deploy your strategies please visit quantconnect.com and to stay up to date on new events be sure to follow us on twitter at quantconnect everybody have a great day and thank you guys again so much for joining us thanks for coming everybody you
2021-03-02 12:39
