Our trading system. An informal overview of the work so far.
hi this is david with the enzo trading system and uh today a different video i would like to do a review of the work we've done so far uh this is more introspective i guess it's more for those that may want to embark on a similar endeavor in the future and about the obstacles the things we found out along the way and they may be useful to others uh yeah the reason for sharing is uh partly publicity and also because i think sharing helps our helps me definitely to put things in perspective as as they... tell them out as it often happens so this is our website uh we started in uh uh february 2018 and we didn't have an application back then we had the general idea that we wanted to do a trading system and we actually called it a bot uh which i think is a bit of a misnomer because uh what we're really doing is uh quantitative analysis uh so the bot is more like the acting uh software but uh the software that needs to act on a on a useful model one that has proven to be effective and that's kind of the hardest part apart in many ways so what we do we actually develop a training system the models but also the runtime so let's see i have a few slides here to help me along the way um yeah we started 2018 from scratch personally i had no clue what uh i mean i had an idea what trading was uh i had no personal experience but i knew people that tried to get into the field and left disappointed even smart people but i think uh the real issues really that one needs to be open to it put a lot of work into it before seeing any results so it's a lot of sacrifice in terms of time and and resources so the initial idea is like you know buy low cell high it looked very simple uh what do i mean like uh yeah i mean honestly i was so clueless uh you see chart you see something like this and then you say yeah okay i just kind of want to buy around here and sell around here it doesn't matter maybe i maybe maybe i buy here in a sell here maybe it's not so efficient but still so of course i mean that looks simple enough but it's never like that uh what really happens in in truth is that you don't really get a nice pattern like this it's more like yeah maybe you you get it for a while and then what really happens is that it starts to go really down so you think uh yeah you know i you know i buy here and sell here i buy here and then when do you sell now you got into a loss and you probably lost more than you made up to this point so i mean this is obvious stuff but if you've never seen it like me uh in 2018 it seemed you know it's intuitive you can see fractals you can see patterns and you realize that you can use them and you can use them but the problem is that uh then there is the exception that breaks everything else so in a nutshell that's really what makes it uh so difficult so it was uh you know the obvious decision to jump into machine learning uh not because it's a buzzword i honestly didn't even think about machine learning what i thought about is uh uh yeah let's develop a system that offers some multiple solutions and then in real time it will apply the best possible solution and this is actually on our first and only white paper uh here there's uh some some graphics that i put together myself at the time and yeah you know the you know normally what one does uh without a trading model uh you have to kind of decide the market if you have a script you then you set up a script and then you hope that it will work instead uh ideally you have kind of a multiple realities running at the same time and then you pick from the best one of course the the big problem here is that you pick from the multiple realities which you know they just did the one the best one that just did really well but is it going to do well next and what does it mean next and what does it mean just now uh so do you base your your cost function which is what tells you how valuable an algorithm is uh do you visit on the last week last few days in the last month you base it on one year but is it going to translate well in the market that is going to come so it's really uh all decisions that are made in the past in a sense you have to guess what is a useful portion of the past that will translate into the future and it's not just the useful portion of the past the usual full portion of the past that works with your algorithms which probably by definition they already they don't do profit anyway so i mean in an ideal world you have a system that builds the algorithm itself but that would be really intelligent and uh then what you end up doing is creating a system that uh builds a huge uh overfit which means like yeah really tuned for the past but really doesn't work well for the future so this i mean the idea basic idea uh works but to make it work it's a whole different uh game but yeah we started with that the basic idea and i personally started uh working on the strategies i learned about the the technical indicators which is what everybody talks about is your stochastics which means like a range of minimum and maximal within a certain range in the past once again relying on the past or uh the line fit you know the market is doing like this so you you kind of extrapolate uh direction a linear direction needless to say these are all things that everybody thinks about and therefore they don't work maybe you make a linear extrapolation maybe you have two linear extrapolations maybe you build a machine learning system that will try all possible uh extrapolation based on a different series of historical uh data and then again you always have this problem that you constantly not really finding the right parameters uh that will translate it to the future there are there are patterns to be discovered but they're not that simple that's kind of uh the problem and they're not forever so there's a lot of work to find those patterns and it's not direct necessarily drug human work you may have an intuition but then you have to put into the machine and then figure out if it works and maybe you have a range of parameters and then you figure out you the a to the machine you figure out whether uh that's reliable if you have to find a fixed parameter if you have to find a range of parameters that that you select in real time so just just uh this is just a infinite way of possible possibly doing this stuff and so then i start testing every possible strategy everything you can think of today you know somebody comes i know anybody who will uh from time to time will come and tell me you know did you try this did you try that yes i tried it uh i tried it maybe not specifically that thing but in most cases i try to and um and the tricky part is that some stuff may actually work but there is many ways of trying it you know if you try like if it is and green candles then you buy and if it is N candles then you sell maybe in that specific context doesn't work uh maybe there is something else that you have to do with it so in many ways sometimes you can almost make everything work as long as you put other things together and then and um but in general you know everybody you know anything that anyone can think uh of uh i've been pretty deep into it and i tried uh just about uh anything and other stuff you can just you just can't try because it would take you know it would get you into a rabbit hole and i could easily spend one year just about anything there's only some idea you can get some white paper of somebody who says oh use kalman filter so use this kind of pattern matching and it works for us yeah in the paper but the stuff that really works i think in this world is more secretive so i come from computer graphics where papers are published and even there you have to figure out uh if that stuff really works in practice if it works for you uh what's the catch and that's like more an open world you go to the siggraph everybody's happy to present their paper you go to the game developer conference when it comes to finance of course it's more like uh closed secrets because um i mean anything can can make you money almost instantaneously if you figure out something it's not that simple but there are some things that you just don't want to tell them around so it is kind of exciting in a way but at the same time it's tricky because you you just uh um you know walk in the dark you don't really know if you're really close to something that works if you give up on something or some approach or uh and then maybe you work on that it happened it happened to me now i have we only use custom indicators today 2021 and uh there are many ways it's an original idea that we came up with like at the beginning but then it didn't seem like it worked and didn't seem that it worked and i spent months and years trying to make it work and then eventually uh with a certain approach it worked so some stuff you just throw it away you never look back some stuff sometimes you look back it's like okay now it works for some reason so yeah uh it's maddening but that's kind of how it works and so yeah we try all sort of uh custom indicators historical pattern match any kind of polynomials uh fractal approaches you know multiple uh levels of uh you know self-similarity at a different level of frequencies it's just a this joke and yeah and then you know as i mentioned machine learning uh the trick about finding the best cost function which means how do you evaluate an algorithm so that you select that algorithm from from now on and then do you evaluate it you know to work for the next month for the next year for the next few days and uh and then you know your evaluation may have bugs uh you you never realize that maybe the idea the basic idea was there but it didn't quite work a common one is uh oops so cost function uh you want to you deal with uh many times the obvious thing is evaluating your profit so how do you calculate your profit because it's very um you you generally you kind of divide for every uh position you divide the price of which you sold and the price which you bought and that relationship will give you if it is higher than one it's a positive if it is less than one is a negative but uh a few times i bumped into this issue that basically i would sum uh those values but you you know those are uh not meant to be summed they should be multiplied together that's how you you accumulate something that is uh fractions uh so and that's why actually uh it is suggested to use uh logarithms and put as much as possible into logarithms because then you can sum them it's just in general it's not like it's some special voodoo technique is just uh many times you'll find yourself you want to sum things out so if you have a logarithms of of the fractions of the sum of the profit then you can sum the profits and uh but if you're summing a profit that are not logarithms then you have a skewed uh cost function and uh and the trick is that many times it will it seems like it works because you can tune your algorithms around the cost function it's you you don't really ever have a fixed point so you can say okay this is my cost function i rely on this then i'll develop the algorithm on that and then somehow the algorithm works on the tested data and then it doesn't work on the day that you didn't test them perhaps you bump into cost functions oh this was broken now you fix the cost function now your algorithm maybe doesn't work anymore because it was based on on a wrong premise so it's really important to i mean yeah you say i i implement machine learning but any bug can throw you off another thing that we tried is our genetic algorithms this one is a pretty intensive on the cpu and on the ram i think yeah if you have a lot of a cpu to spare but um i mean genetic algorithms frankly because it was one of the simpler things to implement if you want to use neural networks maybe it's better but even then you open up all the different kinds of can of worms but um yeah the problem is that it's kind of a idealistic approach where genetic algorithms it's basically a way to find a solution for a very complex problem meaning that you build your algorithm so that is a very complex algorithm that has a lot of parameters uh possibly you put all the possible indicators and then you tell it you built kind of a virtual machine that builds some uh op-code and and parameters to use any possible combination of indicators and then uh what and then you have this very complex problem that you put into genetic algorithms what that what it does is that because you couldn't possibly test all the combinations you will kind of refine the combinations in a way and discard some of them maybe it will discard the ideal combination but nonetheless so you will kind of find a good enough solution but even then it's very intensive i mean actually this machine is 32 gigabytes and a few times i had the windows uh just crashing on me um too much ram too much cpu and and then you you end up building something that really overfits um it's not something to discard completely uh but um it's kind of uh overkill and it will uh easily lead one to have an overfit solution which is definitely what you don't want to have if you want to make money and not just show some pretty numbers and back tests so and then now we have some early success with the simple approach which is i called uh the climb uh which means basically let's see if i if i have one that i can show it's a reactive algorithm that it will well i don't have it here let's see one of the markets please it's a reactive algorithm that will jump on the opportunity uh which is uh instantaneous so whereas other algorithms with the technical indicators will try to uh look back like days uh maybe months and kind of guess what's happening what's going to happen in the next few hours and this one really basically jumps i see the spike of volume and it decides that's a very high volume and then and then maybe there is some some pattern that detects uh which i won't necessarily reveal right now and then it will jump in and buy instantaneously and actually oops and then sell rather quickly if you can manage to have not too many false positives uh this is kind of a very good bet this is actually our most reliable approach so far but uh the thing is these opportunities are quite rare and uh and these patterns they they tend to vanish so from time to time i wouldn't know if this one works on year down the road and in fact since we started uh this approach i had to review a few times i think now it's more robust we actually apply to all the algorithms [markets] but um it's something that you have to maintain but i think this is closer to a more realistic solution it's less uh voodoo is less about uh you know looking back a lot of in the past and trying to fight this part so it's really yeah there's a lot of volume just jump in and uh and jump out uh it's not so easy but it's we had some early success uh it was based on a very high frequency data and it was really heavy to test and then uh they worked well in 2018 and 19 then start they stopped working you never know if it's gonna work in the next few months then we dropped it and then i picked it up again and now it works and it's kind of an example of things that uh work really well or maybe don't work and then then you forget about it and then you pick it back up and then eventually it works it happened a few times and if i've known before you know but you cannot know you just have to live through it and trying to guess and um yeah initially we started working on multiple market eventually we focus on bitcoin just because the back testing it's if you have to test just one market is one thing if you have to test multiple markets it's more work you have to optimize your back test and also you have to have some sort of portfolio management because it's not enough to say oh yeah i could do well in this market i couldn't do well on this one but you know because i can do well in this market then you know i'll put it online then how much do you allocate to each market that's kind of the big key and that was kind of the thing how do you make a back test that deals with multiple markets uh you kind of have to build infrastructure to test for it and this is something that eventually we implemented only uh this year but back then we decided it was better to focus just on bitcoin it was probably for the best uh so and in 2018 also we started the internal fund uh with a few selected uh friends so 2019 i don't know what the what the hell we did it was like more of the same uh uh the algorithms a lot of uh existential crisis uh uh can we do it fully automated uh you know you start tweaking you start making trades by hand some trades by hand you don't trust the system uh you have to decide uh can you really improve the system so that you can trust it or do you just have to wait for the for the current algorithm to uh do its work that's that's always that that big question um the solution is like yet to continuously develop and uh develop mostly under improve the testing because that the better the testing the the more uh reassurance you have that what you have right now it's something that uh it's workable um yeah we did develop all our major custom indicators as i mentioned like whether is an indicator that tells there is going to be a pump uh or some other kind of uh indicator the individual patterns and historical matching uh but there was like just a lot of work figuring out what works what doesn't work and sometimes you put together some indicators and the back tests will tell you yeah this this is a good algorithm uh because you mix these uh indicators together but what you're really doing is it's kind of uh yeah again overfitting maybe in the end the other indicator all it did it was just a save you from some major loss and a few trades or or you just somehow i influenced that kind of a butterfly effect and influenced you to the algorithm to get a really good trade the you know it was jump 10 20 percent the mid year the year and so you have to figure out what's the real usefulness of some indicators sometimes you put them together and you seem like they work and then in fact they're just building randomness maybe you just have to use separate indicators on several algorithms uh yeah whatever yeah i mentioned this already with you know we tried uh genetic algorithms super algorithms where you know ideally you build uh you have this uh virtual machine approach with the code and the op-codes and kind of it's it's an interesting idea you know kind of like a truly uh genetic approach where you the code builds itself is a self-modifying code in a way uh but is it more like for research and because it takes so much uh computing power it's better to cut to the chase and just could uh use a little bit more brain and know so much about this stuff you know maybe it will be more interesting research to to do one day when you know i can kind of retire so we also explored with the forex markets uh but uh yeah it looked very hard but honestly we didn't have a good algorithms anyway so i didn't know very much what i was doing uh and forex market it was just uh possibly even worse than uh trading uh bitcoin the key for markets uh forex markets and even in the crypto today is really to be able to leverage multiple markets and really jump on the one that uh that works best at any given time which of course is not that easy but it's better than focusing just on you know japanese yen us-dollars or something like that and uh back then we didn't have a system to deal with multiple markets at the same time so um we explored it and i figured someday and then uh yeah we still had the internal fund and it was going just going down i don't know what bitcoin did to 2019 but uh yeah the fund wasn't doing so well but it didn't go bust so how much i'm with time i don't know so uh 2020 finally something concrete uh we finally uh developed a trading client that others could use so that enticed uh that implied uh using having a proper uh graphical user interface having a connection to the server setting up the signal server meant even more infrastructure infrastructure up to this point was mostly to data gathering and uh the build machine all the basic stuff that one needs to for uh for development but now it's like it was really important to have a signal server because if you if you're trading for yourself locally you just uh don't really need a server you're just the clients running and generating signal and doing the trades if you have to generate the signals and let others do the trades then it becomes a kind of a more of a network issue yes that's uh yeah it's a whole can of worms and uh 2020 was the first year with a set of trades that we did ourselves and other uh users did and we could match and uh it was fairly profitable uh so 120 percent but with the maximum down 25 percent uh which uh seemed tolerable seemed good but uh once you start playing with leverage it's definitely too much so we had uh you know we noticed months of drawdown but we ended up in profit uh we learned a lesson yeah you have to be patient and that's what we always say and we continued to work on algorithms and testing uh finally we implemented the sharpe ratio and uh i think at this point we finally put monte carlo validation but with sharpe ratio as an industry standard monte carlo validation it's it's more a generic term you can do in many ways but honestly uh sometimes they they're reliable sometimes they're not reliable if you have a few trades these can go really haywire i mean i have algorithms that work really well and the monte carlos are crappy these days because they generate a few trades so we should generate more trades but if you generate more trades then then you you know you end up having crappy trades so uh there are ways to improve these things like uh sharpe ratio for example we base it on the trades i think and then maybe better to sectioning uh instead of basing on the trades based on the split the trade uh depending on the if it is spanning multiple days maybe or multiple hours split for the hours and then do the subdivide and uh and calculate the sharpe ratio maybe i don't know but um the sharpe ratio idea is an industry standard how you really calculated it's a whole different thing and uh it's a bit tricky because you want to tell people oh i have this sharpe ratio but uh honestly i can just tweak one thing and it will be better or worse so it's more like for internal use but you definitely don't wanna i for myself i'm gonna rely either on this and not even so much on this uh when it comes to really evaluate the algorithms profit and uh the relation of profit to the drawdowns it's the best evaluation so far and um and then the internal fund finally it was back to its original glory so making up the losses and uh and actually uh back in profit so yeah 2021 uh tricky year we started feeling uh quite confident we started using leverage and and we started considering more markets than bitcoin we've seen a lot of returns up to may and then with a big drop it was interesting because a big drop was actually a success big drops are not a problem the problem is like what happens afterwards uh the the volatility of the market and that it's kind of both a negative slope and highly volatile market that's really what kills you and um and yeah uh we were kind of another thing we really didn't have too many markets only up to five markets trading so but there wasn't really any uh true diversification because once bitcoin starts to move a certain way everything else moves that way so there was no diversification there was high leverage and uh the behavior was actually kind of as expected uh because we did have uh we didn't know that we were going to have uh possibly 25 percent uh 30 percent maximum drawdown and with leverage becomes like sixty seventy if you have 5x maybe even eighty uh maximum draw down and uh if you see in the test okay fine then you get out of it but in practice so it's quite debilitating and uh yeah that was a kind of a lesson and that as a reaction we did a major work on position sizing and the focus really on reducing maximum drawdown um this is something that we knew but uh kind of you postpone it uh and then you realize even if it's not the end of the world because we survived the problem is uh maybe you're okay but people they uh a lot of users they may jump into the market with our system uh when it's doing really well and then they get the full brunt of the maximum down which is which is all the way down without even ever had the any profit prior so our internal fund is still kind of a profit even though i mean we went thousand percent you know back 50 percent that's i mean that's a big loss but it's a relative loss we're still in profit if you come in uh at the top and then you just get the the losses just to start with this uh quite stressful and we definitely don't want to have uh you know negative publicity you know it's the ideal fund this one that doesn't give you so much volatility in the fund and um and then we also have the major refactoring of the build this is more of an internal thing but how you structure the software and uh make it more maintainable switch into libraries that are more standard for connectivity you know uh so this is this kind of infrastructure that is necessary that was kind of postponed because up to this point i was mostly dealing with the runtime by myself and finally i got l polls on this side of things and uh this was you know great and necessary and uh and then we have the support for bybit uh not only no longer binance only uh it's very important because finally don't have to rely on a single exchange and and if we have to implement another exchange we can we can do it once you implement the second exchange the third one is not so hard and then uh and now we're focusing more on generic not genetic algorithms meaning that instead of focusing so much on bitcoin uh really trying to make the best possible on just on bitcoin or ethereum or just a few selected markets uh which is a bit of a fool's errand because yeah you put a lot of effort in there but uh you it's more likely that you end up overfitting and then you stress a lot when that when you know if you're only on bitcoin and you miss some big jump uh positive jump then that's that's all you have to demonstrate at that given time and uh and uh it's basically 100 failure or it's it's more binary but if you have an ensemble of markets that you're dealing with you kind of a little bit of a chance here a little bit of a chance there and this is definitely the way forward and actually i'd rather even also not just uh focus on crypto itself the ideal uh the final solution would be that uh we are on multiple markets we're forex we're in stock market uh i don't really care i mean yeah uh crypto is great but uh at the idealistic level but as far as uh uh trading making money out of it i just wanna go whatever market uh whatever and assemble a market so uh works best so now with uh see what major lessons uh did we learn definitely i learned yeah impossible importance of test infrastructure this was um kind of something that i didn't value as much at the beginning because it's very easy to say okay you know this back test looks great uh but really you have to put just a lot of effort to try to negate uh whatever positive comes from a back test i said this before but it's worth stressing again it's definitely trying to find holes in your in your algorithms and trying to make uh tests as reliable as possible and one option for that is like if you have a lot of market historical market data then to that and crypto we don't really have that much uh so another thing is as is this algorithm this technique work on multiple uh markets yeah they're all similar but they're a bit different it's it's kind of you can consider maybe ethereum and ripple another and other markets as a variation of uh bitcoin in fact a testing system that some uh uh push for is that of generating uh random mutations of the certain market if i only have a bitcoin then maybe i generate random mutation and then i also perform the back tests on those mutations of the market and on the other side you know you can consider ethereum or ripple or bnb or whatever as mutations of the bitcoin so this value into either building a system that builds random mutations of given market or just build an algorithm though it works as well as possible in a variety of markets and um yeah testing uh test the risk as well always uh deal with the risk uh worst case scenario also deal with the fact if you're doing leverage consider uh that if it is tolerable or not and this is also very important it's not technical but uh very key importance of a communication with the whether you're calling users, investors depending on the kind of product your but still if you have any sort of uh if you generate any sort of signals that other app to use uh you really have to be sure that uh people understand what they're getting into so some some concept they require a lot of time to understand and you can't really give a simple explanation some concepts i i mean i've been on this for uh you know three years and a half four years and uh some things eventually they click in your head but it takes a long time to and this is my this has been my life uh uh for most of three years and somebody that jumps in and you you try to say oh this is the risk uh don't worry about this but really i mean you can't i cannot give you a guarantee but statistically this statistically that and uh and people can easily get scared so you kind of uh to anticipate the expectations but to the general level understand what people generally expect uh if you know big drawdowns and the lengthy drawdowns is something that nobody likes and then it depends on the you know some people did they they're more obsessed on a daily basis you know and this is not uh a dig because that's why i'm also obsessed on a daily basis but uh is not something healthy i have to do it because i'm i developed this i'm kind of responsible for it but uh the better thing would be to just uh be patient and let the system work but uh it's easier said than done so and it depends on the people some people do more or less emotional they're open to more or less risk and that they don't really understand necessarily what kind of risk they're getting into meaning risk it's usually i mean it's not like i don't i never seen permanent loss i don't don't envision that although it is possible but uh the temporary loss uh the stress that you have to go through for uh you know seeing months of of the drawdown can be quite stressful and uh and what can happen the people that you know they may pull out uh thankfully it's must never happen but uh uh if somebody pulls out after a drop that's kind of the worst possible case because uh basically you concretize the the you solidify the loss and uh you basically remove a chance of making that backup and i actually heard the statistic i think it was uh joel greenblatt a famous uh uh hedge fund manager that uh he mentioned statistics by which uh most uh hedge funds they're profitable but most uh investors even sophisticated investors in hedge funds they're not profitable or a good portion of them and that's because they do play with adding more and removing uh the from the fund from their own fund and that in itself is a way of uh trading so uh yeah as i said it's easier said than done but this is very important importance of our communication and uh if you cannot communicate something then err on the safe side it's always better yeah the problem the problem is that you want to be on the safe side but the other uh face of the coin is that uh if you tell somebody you know take a very little risk don't use leverage only use 2x something like that and then when you don't beat the market it's like okay now the bitcoin has gone up uh all the cryptos going up but we didn't go up as much and so then you have to say yeah we didn't beat the market because we took less risk oh i wish i took more risk but uh that's a bigger discussion and um yeah i'm learning as i go but uh yeah definitely communication is uh very important that's very stressful for people and for us because you really want to prove and you really want to give as much as possible as quick as possible but it takes time but uh you know that's the name of the game you know you wanna if you don't want to get the heat don't get in the kitchen we're in the kitchen i'll take the heat not as much heat you know it's not so bad but yeah that's uh um yeah i hope something useful came from this maybe it was just a way for me to to rehash things but um yeah until next time you know if you have any questions put it in the put it in the comments whatever uh see you
2021-10-22 06:19
