131: Morgan Slade – Strategy Development, Powered by Machine Learning

Chat with Traders is brought to you by Trade the Pool. Did you know that every decade the market reinvents itself? Online brokers opened the doors, mobile apps made trading seamless, and commission free trading erased barriers. Now a new era has begun. Meet, trade the pool, limited risk trading. And now you also have unlimited time to reach the profit target. From now on, your trading risk is capped. and your trading opportunities are limitless.

Trade the pool funds home-based stock traders with up to$200,000 in buying power. That means you can trade larger positions and scale your strategies without risking your own savings. It's time to trade with more capital, making it truly worth your time and effort. Ready to trade the pool? Click the link in the description and join the stock trading revolution today.

Are you ready to get serious about trading? Then join Tasty Trade, Investopedia's best platform for options trading in twenty twenty six. Options, futures, and more. Tasty Trade has everything you trade all in one platform. Get low commissions, including zero commissions on stocks. So you can keep more of what you earn. Trade smarter with advanced charting tools, a pre-built strategy selector, risk analysis tools, and more features. Visit Tastytrade.com slash.

Tasty Trade Inc. is a registered broker dealer and member of FINRA, NFA, and SIPC. This is the first one. Podcast.

How you doing everyone? Welcome to episode 131. Now you'll recall I had Andy Kirschner on the podcast a few episodes back. Towards the end of that episode, Andy briefly mentioned a cloud-based algo development platform and fund, CloudQuant, which is a subsidiary of Kirschner Trading Group. I mention this because with me on this episode is Morgan Slade, who is the CEO of CloudQuant and has been since 2016.

Prior Morgan's career as a trader and portfolio manager goes back about twenty years and during this time he's worked at many prominent firms and funds. Just to name a few, Melbourne Ridgefield, Merrill Lynch, Citadel, and Austin Trading, amongst others. During our chat, Morgan explains why he feels as though the common approach to strategy development can be counterintuitive and

and he gives an alternative three-step formula. We also spend a fair amount of time discussing how machine learning fits into a trader's toolbox. If you are new to machine learning, you'll be fine. Morgan does a great job of explaining things in a way that's easy to follow. Alright, intro done. Here's the conversation I had with Morgan Slade.

All right, excellent. Well Morgan, just give us a quick rundown on your background. Tell us a little bit about your history, your backstory uh in this business. I know you've been uh floating around for a while, so um yeah, please share. Sure. Um I started out uh studying engineering actually. um as an undergrad and graduate student uh at MIT and there were a l a lot of a lot of people that um worked for O'Connor and Associates, which was an options trading shop here in Chicago.

uh that I saw leave the engineering field and and come to Chicago and start trading. And um as I was finishing up school, you know, that opportunity kind of wound its its way down and Swiss Bank purchased them. But uh I found my way to uh to New York City and started working for uh a very old CTA um that been around for about thirty years called Milburn Richfield. And uh their idea was to train

quantitative traders uh first by training them as traders. And so my first job was actually um as an execution trader for a large hedge fund. in the New York City area. I spent uh the better part of the last twenty years or so using um quantitative trading strategies and various uh shapes, you know, both on the buy and sell side. and have been using, you know, things like machine learning and and techniques like that for most of that time.

and uh have had the opportunity to trade um most asset classes, um FX and equities and and some options and uh and futures of course.

I'm interested to know why that that first place you worked at, what was their idea or their thought process behind Getting quantitative traders to start out as, I guess, more uh discretionary traders before actually coming across to the quantitative side. What what was the idea there? Well, there's a big jump in in terms of understanding how to get your trading algorithms to to generate the returns that you think you can based on backdus.

A lot of this comes down to what I call trade expression and understanding transaction costs and market impact. When you trade for a firm like that, they had a two billion dollar um book that was levered up seven to eight times. So um in the end they're trading almost fifteen billion dollars. uh face value in in terms of futures. So as a trader for them, you learn to execute very large transactions. Um, you know, you a small trade might be, you know,$25 million trade.

And sometimes your bigger trades could be a couple hundred million dollar trade in a currency or um you know, if you're trading futures, you know, these um these funds sometimes can be trading position limits. in a lot of the futures contracts that they trade. Trading that size clearly has a market impact and you have the ability to see your impact on the market and just kind of internally builds a a you know, a sense for

how big you can trade something and and how much you're likely to move the market. And figuring out how to measure the capacity of a strategy is one of the hardest things to do. And so I think that's that's something that they had in mind when they looked at that as first of all, you need to understand the rules of the market and how it works and how to enter trades. And um Save yourself from chasing ghosts. And they also felt like by sitting on the trading desk you generate a lot of new ideas.

And so rather than just take what you know from graduate school, they felt like adding that together with a bunch of real life experiences on a trading desk. would probably create somebody who was set up to come up with a lot of good trading ideas and trading strategies. Right. And I think it just might be interesting to hear a little bit more about how you would actually go about executing

a hundred million dollar trade. Like how do you actually put that position on? So I mean, do you work that over a couple of days? Um, what markets were you trading when you were trading that size? And Uh yeah, just speak to us about that a little more'cause I I think that'd be uh why, obviously not relatable to most people listening to this, but um I think it would be interesting to hear how you actually work a position like that into the market and also

Is it an outright position? Like how you actually structuring that trade? Uh yeah, these were these were out outright positions. Um uh so you you typically have um Uh a directional bet that you're placing. And we obviously traded options as well, but most of that was to just gain leverage in the directional bets we were placing. Um the trades themselves typically were traded through, you know, large broker dealers that had

been instrumental in raising capital for us and had uh you know a an execution relationship with us uh through some sort of managed account. Um and so we would have to um take some portion of each trade and trade it through the appropriate counterparty. Because of that structure, it it was sometimes challenging to to keep things, you know, kind of under wraps and and and keep them quiet, um, in some of the, you know, over the counter markets. There was definitely, you know

at times a bit of a chuckle on the end of the line when um you called the last guy on your list. Um of course the lists are randomized and and and fair, but um you know the last transaction on the list was usually somebody who knew

kind of what you were up to and they just needed to know what size trade you needed to do with them. So um that's kind of my experience, you know, trading FX, uh, you know, uh with institutional counterparties. Um the um the next step would be to kind of enter into you know the forward swap We're not taking um delivery of the uh of the currency, so we would swap it out to a forward date.

usually the IMM futures expiration date, but sometimes other custom dates. Um on the futures trading side, we usually had brokers on the floor that would be executing for us and we'd, you know, many times have a direct line where, you know, we would pick it up and and they would ring and uh we would stay on the phone forum for sometimes hours at a time, kind of working these orders. Um and uh, you know, it would generate a lot of uh

A lot of commissions for them, but they also worked very hard for us. You know, just uh trying to get that much volume through the market was sometimes difficult. So I mean there were times where you definitely saw I remember trading crude oil once and

I kind of felt like I personally, you know, moved it a buck or two with the size that I was doing given the volume that we had to do that day. So um and that's, you know, those are valuable kind of insights to have as you go back to the modeling side.

to kind of have a sense of scale. Um it eliminates a lot of questions that you have and it eliminates a lot of things that you might try to do that you realize aren't practical. So so Where you actually Like obviously when you want to buy a hundred million dollars of whatever that might be

that's broken down into a lot of smaller trades. Where are those smaller trades being executed? Like I I know some people in the past who have come on this podcast have spoken about uh VWAP volume weighted average price as like a institutional benchmark. I mean I don't know if V if VWAP is more of a a uh was necessarily used, um, however many years ago this was, but um i is there a point where institutions like to try and fill their orders or is it sort of all over the place?

Well yeah, unfortunately, uh, Aaron, I think you're you're dating me. Um you know, back when I was doing this, you know, people didn't really use VWAP that much. Um and it wasn't actually until um You know, I I joined Merrill Lynch a few years later that uh I really got involved in Kinda helping to write some of the the first VWAP strategies. But um you know, for futures traders, um algorithmic execution was not something very commonplace at all back then.

And I I don't really even think it was available. And so, you know, the the traders really played a large role in breaking the the orders up into smaller pieces and

trying to get the best ex execution they could, you know, for for the end and investor. Okay. And just going back to something else, you said you were talking about when you were putting on these big trades that you had to work with like an institutional counterparty. So just help us understand why that's the case. Like why can you not just put that order into the the open market?

Well, for for over the counter stuff, many times there would be some c contractual relationship with uh a bank that say raised capital for the fund, um they would, you know, they would be able to um ask that we execute, you know, or at least usually execute the um the orders through their trading desks so that they get some benefit from that. Um as kind of part of the deal. On the futures trading side for listed products, obviously you're executing on the exchange, you don't have a choice.

And so um many times they might have their own, you know, floor presence, but a lot of times um we'd go through independent brokers on the floor.

So we had a lot more latitude in terms of who we who we dealt with. Well let's move this conversation a little bit. I'd really like to talk to you, Morgan, about developing strategies. So we had a call uh was it about a week or so ago prior to uh you know speaking now uh and recording this one of the things you'd mentioned to me is that

I guess the quote unquote normal way of developing strategies to you seems kind of backwards. So I'd be interested to hear a bit more about your process, uh or I guess an alternative process to Approaching strategy development. Sure. Yeah. Happy to go into that. So one of the things that is is most frustrating as a as a trader and as a quantitative trader is, you know, you put a lot of effort into coming up with ideas, trying to figure out what to plug into your model.

And coming up with some risk rules and how do you cut your losses and how do you take your gains. And and then at the very end of the process you get a little bit of feedback in the form of an equity curve or or or some sort of sharp ratio or some sort of rate of return. And it seems like you went through an awful lot of steps to try and kind of get that that metric, that that measure of whether you actually did anything right.

And uh most of the time unfortunately we don't we don't get it right the first time or the second time or even the nth time.

It takes a lot of practice and it takes a lot of a lot of, you know, blood, sweat and tears to just kind of find something that that actually works. The markets are pretty efficient. And there's a lot of smart people out there. And and so After, you know, kind of going through that process for many years, I started thinking about, well, if I start out and I I mess up the signaling part of my strategy. then everything else is probably gonna fail too.

And so I wish there was just some way to to break'em up into separate steps. And and I started thinking about it and there really is. You know, if you break it up into a prediction step um or a signaling step where you come up with some signal that tells you to buy or sell. Um and then you view the second step as maybe um, okay, how big a trade do I want to put on? So we'll call that uh portfolio construction or or risk allocation step.

Um and then a third step is um, you know, uh how do I enter my trades into the market? We talked a lot about how we did it, you know, twenty years ago when I was a futures trader. The world has changed a lot since then. I've actually spent the last, you know, twelve years doing high frequency market making, so I understand

how uh to enter orders into the market um now as well. And um and so the trade expression step is a s is a third step and really all three of those steps contribute to your P and L. But if you can measure each step along the way.

as a milestone to see whether you're getting you're headed in the right direction. Um it makes a lot more sense in terms of uh You know, not wasting your time on portfolio construction, not wasting your time on trade expression until you have a really good signal or a really good alpha.

Okay, well let's flesh out each one of those steps a little further. So you said the first step would be I guess we could call it the signal step. So What's involved during that step? Well, for a lot of people nowadays, um there's a lot of data scientists out there, people aren't learning machine learning. And so they might look at that step as a a data science problem.

Where they wanna do some supervised learning. So what that means is they basically label a bunch of examples of their ideal trade, what they would have done if they had the ability to do Monday morning quarterbacking. And then they take a bunch of things that they think might have predicted and been helpful in predicting that ideal trade. And then they do some machine learning to find a rule set based on the inputs.

as to how to most accurately predict the perfect trades that they gave as examples. And that's basically what supervised learning and machine learning is all about is identifying what you would do in a perfect world. and searching for things that help you predict how to do that in the future. Other people might take, you know, technical indicators And just play around with mixing them up together. and different rule sets and different things that, you know, maybe you think a product trades

in a correlated manner with another product or a co integrated ma manner with another product. So you might take, you know, one leg of a pair of s of socks that are highly correlated and And you might use one to signal the other. Um there are lots of different things that constitute a signal and everybody has their own definition. But at the end of the day you're trying to predict something.

And you're trying to figure out what your win rate is. And so I would focus on, you know, what is my win rate? And Um I would also focus on the symmetry of or or what the ratio of your average profitable outcome.

to your average, you know, losing outcome is. And um you can do some reading on uh the Kelly criterion if you you want to look at, you know, some gambling theory and how that might inform your signaling process. But at the end of the day if you have something that pays out, you know, even odds, then The win rate has to be higher than fifty percent in order for you to make money.

But if you have something, you know, as an example, something that pays two to one odds, say your your winners are twice as big as your average loser, then you can afford to have a win rate of just over thirty three and a third percent. Most people don't really think about it that way, but if you do and you you use a little bit of kind of gambling um background or or theory uh to to look at your signal. You can realize that um

that, you know, it's not just a function of how how many trades there are winners, but also the chafe for the distribution of winners and losers is also very important. So you want things that have what's called a positive skewness, things that have bigger winners than losers, typically.

Um and so you could look at uh Melbourne Richfield where I first worked and you know they had some surprisingly c profitable strategies that had win rates in the twenty five to thirty five percent range because their odds of uh you know the the the odds ratio for their trades, their average profit could be five times larger than their average loss.

And so I think that's a a dot that I connected, you know, after a few years and I think for a lot of new traders is is kinda troubling to think that uh you can have something that wins less than half the time and it's still okay.

Right, so as we're still on the first step here, which we've labelled as the signal step. Is this signal, is this just the entry, or is this essentially the rules for your your full strategy in terms of where you enter and where you exit? Like where where does the signal kind of stop? I like to make it just the entry. Um I think it keeps it clean. And allows you to kind of fine tune the exit later on. Okay. So how do you determine how good the signal is at this point?

At this point you're looking at the trade win rate, you're looking at the um odds ratio or the average profit to loss ratio. But is that not dependent on where you exit the trade? It could be. Sometimes you you say, Oh, I'm gonna keep it simple and I'm gonna look at something from you know from close to close or from open to close. And so Each one of those time horizons that you want to try out becomes a different step one that you you need to explore.

And so you you really particularly when you're doing the the machine learning, you certainly have in mind, you know, when the ideal trade started and when the ideal trade stopped. And so that's really defining your entry and exit conditions. Okay. And that one day hold or open to close, is that something you use quite frequently in your own strategies?

Yeah, I mean we use lots of different time horizons but but very common ones are one day or open to close or two day or five day, you know, holding periods to try and capture kind of medium frequency opportunities.

Okay, so let's say you have a signal which you're happy with. Let's move on to the next step. So the risk allocation step. So how would we move forward from here? Well at that point you wanna try and figure out how the best to weight your trades. Um a lot of people are trading baskets and so if you have a basket strategy then you might start out with

uh an equal weighted or naive allocation as a starting point. And then what you want to do is um if you're gonna use something like a mean variance optimizer, to uh wager trades than um or or some other type of allocation scheme that you cook up. Um, you wanna compare that against that benchmark to see if it actually improves.

um your your strategy. Um there are some diagnostics um uh that you can that you can get into, uh which is probably too much for this call, but um You know, there's um uh a body of work in active portfolio management that talk about looking at um uh the TC coefficient, um which measures whether your weights actually um are proportional to your alphas in your portfolio or your expected returns in your portfolio.

Sometimes when you over constrain things or you have too many rules that you're that you're adding to your um to your model so it's too complex, when you try and do these optimizations, you'll actually get very suboptimal answers. And so there are some diagnostic tools that you can use to tell if that's happening and to kind of protect against it. But for a simple strategy where you're trading single stocks and things like that,

You're really looking for something that doesn't make any major mistakes and doesn't ruin your alpha basically. So simpler things tend to work better than not. And in large part your risk allocation decision has a lot to do with your capacity. And um one way to figure out your capacity is by by testing your strategy at different sizes over time and and trying to get a sense for when the performance is best.

There's, you know, different size accounts. Some people have retail accounts, um most people do. And so they might be looking at um uh capacity a little differently in terms of what kind of capital they have. But um larger hedge funds, you know, they have other capacity problems where they're they're too they're too big for the market and they they need liquidities. So

Um lots of different problems depending on who you are will crop up in the second step. One mistake that I think is very common is when people start out with a mediocre signal, many times they'll try and make up for it in their back test by adding complex stop loss and trailing stop rules. that essentially curve fit that particular back test and make it look better, but aren't actually predicting anything.

And aren't robusts at all. So that when they go out of sample, these same rules that made them look awesome in their back tests. um end up causing them to be doing really strange things that are suboptimal and uh they just don't work and and and out of sample, you know, performance isn't what they expected it to be at all. So those are kind of some of the pitfalls for step two.

Okay, so a couple things. As you mentioned stop losses just there. How do you treat stop losses in your strategies and how do you actually think about where to actually place them? Because In some of the strategy development I've done, which is of course nowhere near the level which you're applying at, but

I've found stop losses to actually decrease performance in many cases. My strategies obviously have a very clear exit rule, so I'm not gonna hold a position and just hope that it comes back one day, but you know, th th there's a clear exit rule, but that e that rule is generally the same rule that applies for when I take my profit. How do you feel about stop losses and and how have you actually found them to be effective in in s developing strategies?

Well, sometimes for mean reverting strategies I found that they've been helpful. But as I've applied, you know, uh a greater portion of my time uh to developing machine learning based strategies, um I've realized that, you know, what you don't really we what you really want is uh another machine learning signal or another s another more complex rule set that's maybe m multi-dimensional, not just looking at your PL.

Uh to tell you when your trade's not working and when the initial you know hypothesis you had when you put your trade on based on new information or or some sort of circumstance that you see, you know, set of uh price action you see in the market or or or something that you can quantify.

um tells you that your initial forecast was wrong and it's time to get out and you know and wait for another opportunity to kind of take another nick, another bite at the apple. Um so stop losses, I think you're right, they do.

decrease performance. They can make your sharp ratio slightly better, but they're very hard to get right. And they don't they aren't very robust most of the time. Um, you know, having traded a lot of different things, I can tell you that there's very few circumstances where Just a pure stop loss is is the best answer you can come up with.

Now, I don't want anyone listening to sorta take that the wrong way. So I'm not saying, you know, you shouldn't be there you should be trading without stop losses and just be cautious of your risk. I totally agree with you, Aaron. That's uh You you've got to use stop losses so that you live to fight another day. But um I think that um you know, we can come up with better answers than just um, you know, P and L based stop losses.

And on this point, how do you actually measure how good your risk allocation is? Like is there any sort of gauge to to help you decide whether you should move on to the next step or whether you need to I guess put a bit more work into how you're allocating risk. Well I I like trading, Aaron, so I'm usually in a hurry to to start trading even if I don't have the best risk allocation scheme.

And I find that until I really know my capacity for my strategy, which I can't really figure out until I've I've traded, although we have we do have some tools on CloudQuant that uh Allow us to use high frequency data to try and infer those types of things. But um generally the final answer and the final verdict comes from actually trading something. So I would I typically would choose um a fairly simple uh risk. Set of risk rules.

and take, you know, the strategy into the third step and start uh doing forward testing and and and doing actual live trading and collecting data on, you know, what the impact is.

Right. Well let's tackle that third step. So if I remember correctly, the third step was uh trade expression, I think it was. That's right. So what exactly is trade expression referring to? Well, the the world has changed a lot in the last, you know, ten or fifteen years in terms of of execution. uh because algorithms are so pervasive, um they've automated a lot of the execution tasks on the trading desks and and so what we see now are you know a lot of the

the buy buy side and sell side traders um are using algorithms to do their job on a daily basis. And um and of course they're they're very productive and they get things done well. And so You know, we we leverage a lot of those tools. Um I've spent um a lot of time, you know, building autonomous algorithms for market making in the um the equity and futures space.

And uh there are a set of tools that aren't really very useful for anyone else who's not a, you know, a high frequency market maker, but there's a set of tools um that use a lot of machine learning and things like that. That allow you to kind of come up with a fair value bid ask for um for elec electronically traded markets that have order books. And um and so, you know, you know in practice we use that. Um we have

A couple of different types of traders that we work with. We have internal traders that uh that are, you know, typically have a a master's or PhD in engineering. And then we have um crowd researchers we work with who They may not even be experts on market microstructure and may not understand a lot about it. So we actually, you know, do a lot of that work for them and use our um trading expertise to help them get good trade expression on their strategies.

And so when they're doing backtests on our system on CloudQuant, they they have a menu of algorithms that they can choose from that we have pre configured for'em. And ideally, you know, Aaron, I think um a lot of people add the most value in the first two steps. And I think the third step is something that um is is best left to somebody who's, you know, got a lot of market making experience.

Or if you don't have access to'em, you know, send it to an algorithm uh that does a VWAP or a TWAP. Um and if you if you can't do any of those things, then another thing you can do is you can do c something called A B testing where you You in parallel try, you know, to randomly um uh express your trade in a couple of different ways and measure which of the results is best and then select that as your um standard way of doing things.

And so that's done by, you know, by the professionals who do trade expression every day, who do order execution every day. They do a lot of that kind of testing with their order flow where they randomly um you know select some of the order flow to do

as an example, and another to do a TWAP, and another to do an implementation shortfall algorithm. And they're gonna try and compare which of those made the most money for for themselves or their client. So this third step of of trade expression, is this something which is perhaps Maybe more important to traders who are trading much larger size who can't just put their entire position on with one click.

It is. And it's also more important for people who are doing shorter time horizon strategies where they're edge per share, which I define as, you know, dollar profit per share traded or edge per contract in the future space, is if their edge per share or edge per contract is is quite low. chances are they need to spend more time on trade expression to make sure that they actually capture those theoretical profits.

or that theoretical edge. If you're making fifty cents or a dollar a share, chances are you're holding your trades for quite a long time and you're not as worried about um whether you got in at the very best price. So a lot of times y it'll be self determined where you spend a lot of your time. Um if you're if you're looking for really juicy stuff.

you're probably not gonna worry about it too much. If you're looking for things that have a really small edge, really skinny edge, then you're gonna spend a lot of time doing A B testing and experimenting with different trait expressions. Are you ready to get serious about trading? Then join Tasty Trade, Investopedia's best platform for options trading in 2026.

Stocks, options, futures, and more. Tasty Trade has everything you trade all in one platform. Get low commissions, including zero commission on stocks. so you can keep more of what you earn. Tasty Trade is packed with advanced charting tools, backtesting, a pre-built strategy selector, risk analysis tools, and more features to help you trade smarter. See equities and derivatives with high trading volumes, dividends, upcoming earnings reports, and more with their pre built watches.

Or create a custom watch list to keep an eye on the companies and sectors that matter to you. Manage your positions with speed and precision using active trader mode, one-click trading, and smart order tracking. Plus, TastyTrades Stellar Trade Desk team offers live support during trading hours if you need it. Visit Tastytrade.com slash chat for more info. Tastytrade Inc. is a registered broker dealer and member of FINRA, NFA and SIPC.

Yeah. Have you ever watched a stock explode and thought, if only I had the capital? Or sat on the sidelines because your account balance felt too small to matter. Good news. With Trade the Pool's limited risk platform, you don't need millions or even thousands to start trading the U.S. stock market. Bypass the PDT and tap into over 12,000 U.S. listed equities. From penny stocks to big caps, ETFs, even the newest IPOs, and short anything you like, with zero locate or hard to borrow fees.

Start your evaluation, get funded with up to$200,000 in buying power so you can go big without risking your own savings. And now you can also have unlimited time to reach the profit target. It's a game changer. Not ready to trade yet? Trade the pool offers a free demo and educational resources. Practice on live data, master the platform, and build confidence risk-free before you even pay a cent. Click the link in the show notes to start trading with Trade the Pools Capital.

This might be a good point to ask you. There's there's one thing I wanted to ask and you kind of I guess uh touched on it there. Have you found users who come from a data science background approach trading strategies in a much different way to those with actual trading experience? And if so, how? The pure data science background, uh, folks, they they tend to have um some ideas as to which kind of tools they'd like to use.

But you know, it's you know, it's like they've got a hammer and th they think the entire world is is a nail and they they're not really sure exactly how to apply it or which data sets to take a look at. So I found that they they pick it up quite quickly, but they they use a they could use a lot of mentoring and coaching.

And um and so we engage people, you know, um, through our platform to to help them kind of get up the learning curve. Um and as people kind of stand out and as they look like they're having progress, As data scientists, we actually develop a a relationship with them, um and and ask them if there's ways that we can help them.

Uh sometimes I'll give them kind of almost homework or you know, books to read and papers that they might read that are kind of relevant to the problem they're trying to solve and will work with them to help them grow. From uh the experience trading perspective. Um a lot of the challenges have to do with you know, many experienced traders have strategies they want to re implement. And so but being a little older like myself, uh, many times, you know, the computer science skills aren't there.

They might have great ideas and they might be almost um able to kind of get them coded up but maybe need a little bit of help. And so we we do spend time, you know, helping people. uh coach them on how to how to design their algorithm to plug into the cloud quant interface um uh that will allow them to basically, you know, kind of flip a switch and and w we can get it funded and get it into production.

So those are kind of the two different experiences we have. Um the more experienced people, um tend to know kind of what they want to do. Um, but I I've found that many of them are open to to new alternative data sets and things like that as well. They're just different conversations. Uh people have different uh different levels of foundational skills to work with and um But uh, you know, the one thing they have in common is they all like to trade. So Alright, well let's

talk more about machine learning. You you've brought it up a few times, so I think it might be cool if we could uh actually dig into that a little deeper um and maybe talk some specifics. I mean, I guess just so we don't lose anyone here and anyone who might be new to I guess trading from more of a data scientist's perspective.

What is machine learning just at its highest level? Machine learning is really just kind of an automated approach to identifying the best rule set for um identifying something you said that you cared about. And uh in this case it's usually, you know, I wanna know when I should have bought and when I should have sold. We actually um it's kinda funny

Traders have been doing machine learning but manually for a really long time. It's um it's it's just an iterative process that where you you try to basically grow your P N L to be the best P N L you can find. with the advance of, you know, cloud computing and things like Python programming language.

and, you know, a lot more data. You know, people have begun to share software that allows people to do machine learning pretty readily without really writing very many lines of code. I mean you can You can probably do um uh you know, supervised learning on a data set and um and build a decision tree or build a uh you know build a random forest model um using maybe twenty or thirty lines of code.

and all the other code has been written by some somebody else who who, you know, was a domain expert and um and shared it as open source on the internet. So we're really in kind of a golden age for for doing research in that uh alternative data sets, cloud computing, uh machine learning, and all these things have come together all at the same time.

And um and so uh what we found is there's there's a a lot of people out there um who are crowd researchers who do machine learning and they're kind of leveraging these three things. to do a lot of research. Now, how do we do it or how is a simple form of what we do? Um Well, you know, we're we're gonna go through and um we're gonna say, okay, I wanna build a um a decision tree based model which is similar to the same rules you came up with with your um technical trading strategies.

um based on what the RSI is doing or what stochastic is doing or when their crossover moving average is doing this or that. Um those rules can be arrived at using a machine learning algorithm. Um if you take the same things you were looking at and you feed them into an algorithm with examples of what the right thing to do was. it will very exhaustively, in most cases, find the best rule set that you could have also arrived at on your own if you did it manually, but it would have taken you

10,000 million times longer and a lot of a lot of sleepless nights and a lot of time. And so the way I look at it is, you know, somebody asked me the other day, with all these machine learning models out there trading with each other. Aren't the markets going to become unstable? And my answer was not at all. The types of models that machine learning algorithms come up with.

are very similar or if not identical to the same models we've been coming up with for twenty or thirty years. It's just that the research process takes much less time. It's maybe a thousand times faster. And so you can be, you know, much more productive. We see that, you know, in real life. We have people we hire that come on board and start, you know, ramping up strategies within a couple of weeks. We have crowd researchers who are using the platform who

you know, have strategies they develop in, you know, one to two months and they look quite good. So um I think that it's a a great time to be um a quantitative trader and a great time to be a trader and a great time to be a researcher for for these reasons. So

Now I know you have a a story y you can probably share with us about how you took an existing strategy or someone who you work with um took an existing strategy. and actually improved it using machine learning. So the strategy was already profitable, was already making money. Uh but then you I guess optimized, if that's the right word, uh, the strategy using machine learning algorithms and actually

uh got the strategy to perform better. Can you walk us through that example and sort of how that all took place? Sure. Um we were adding to we have a quantitative fund that we operate um as part of CloudQuant and we've been adding to our bench of portfolio managers.

I found uh, you know, somebody I felt was, you know, really creative and and and quite special. He had developed a um Uh a strategy based on um analytics data that uh he had found somewhere on the internet, uh kind of an alternative data set, if you will. And he was using it to forecast um earning surprises and disappointments, which is kind of a favorite for a lot of equity traders because there's a lot of volatility and

And um you can win big and you can also lose big. He had a strategy that had been making about twenty percent under leverage returns over three or four years. And I was impressed with the fact that he was a CPA. He didn't have any formal

formal trading experience, uh, other than running this this fund he was running. He had kind of taught himself how to build this fund and he'd raised some money from friends and family and And um he knew uh, you know, fundamental analysis inside and out because he was a CPA. But other than that, you know, he really had zero programming experience at all. He had never even picked up a programming book in his entire life. He did everything in Excel.

And so one of the things I wanted to do was to show that it was possible for, you know, somebody with his background who was obviously very bright. uh to kinda to kinda grow personally and to pick up this new skill set quickly. And so, you know, I I kiddingly, you know, told them, you know, all right, we're we're gonna

we're gonna have to, you know, get you up to speed on Python because that's the language for machine learning. And um he said, Well where will I start? And I said, um, well uh you know my my daughter Ava is eight and she has a Python programming book.

that she uses, I'll I'll borrow her book and I'll I'll get it for you. And he thought that was that was pretty funny. Um but uh he he picked it up really within a couple of weeks. He was able to use CloudQuant um well enough to start doing back tests. And within two months He had replicated his strategy and he was he was starting to do um, you know, simple machine learning to try and improve the signals.

And within about three months, he had come up with um an overlay to his strategy that allowed him to avoid most of the uh bad trades that he was getting and really increased the sharp ratio to where it w went from around one and a half to about three, three and a half. And Um that same guy has now hired uh several quants to work for him who have a machine learning background. Um so he has one quant working for him and he's you know gonna be adding another

um here pretty soon. And um and he is um uh in the process of kind of growing his book and he has uh now five strategies that he's been funded um on and probably a total of about uh ten to fifteen million dollars in risk has been allocated to his strategies. Um and he's been here, I think uh he'll be coming up on just over six months. So that's how somebody who may feel like they don't, you know, really have

uh programming background can kinda pick up uh and acquire new skills and learn machine learning, which he didn't know at all, uh, very rapidly on some of the new tools that are out there like Cloud Quant. So

So you said there is strategy went from a somewhere around a a one point five sharp to three three and a half, which is a massive increase. I mean a three and a half sharp ratio is very impressive. What actually changed? Like you used machine learning to improve this strategy. What actually changed about the strategy? Well, because he started using machine learning rather than having to do kind of that manual like trial and error type approach.

He was able to bring in all sorts of data that he didn't even know. He knew it sorta mattered. He instinctively knew it should matter. But he didn't know how to plug it into a set of rules. And so It just wasn't clear to him how to make it part of his model. So by using, you know, this random forest model he was able to label all the ideal trades

that he was wanting to make and feed in lots and lots of features that he felt like should be relevant. Pick out the ten or fifteen that actually, according to the machine learning algorithm, were relevant. and run a strategy that made a lot of sense to him and it was quite transparent. I mean he could actually look at what the rules are. He could even write them down in English and say, all right, well, when these five things happen, then I want to buy. And when these

You know, nine things happen over here. I wanna sell. Um and so it's nice to use decision trees because they're transparent. Um and if you if you use a bunch of decision trees together, that's called a random forest. And it's more robust for some reasons which I don't want to get into, but um there are some books that I could refer you to if uh If the listeners wanna, you know, read about uh machine learning, they can get uh Python machine learning from Sebastian Roshka.

And if they're looking for um for trading ideas and and places to start, quantitative trading by Ernest Chan is a very good starting point. Okay, so is this a fair comment to make? So This gentleman who you're referring to, he had these other data sources which he felt should be relevant.

He fed those through a machine learning algorithm. which then determined which of those data sources had any predictive power and sort of discarded the ones which didn't. I mean, is that a a fair comment? Yeah, that's exactly you know what you do. So uh in particular with the machine learning what you want to do is you know you have an initial set of data that you wanna do that feature selection on.

And then what you do after that is you say, Okay, so those are the pieces of information I'm gonna use. And you do what's called a walk forward and you basically from that point onwards using any more information to change which inputs you have and you just start doing your predictions. And so, you know, as long as you use a big enough data set to train with, we usually use about thirty percent of that to train.

And then we leave, you know, seventy percent of that as as kind of a walk forward out of sample. When you're using machine learning algorithms to optimize an existing strategy, you know, just in this example Are there any other measures you need to take to try to avoid curve fitting? Like you said right there, you you split the data between thirty percent in sample, seventy out. Are there any other measures which are necessary to take when you're dabbling in machine learning?

Yeah, I mean one of the rules that I I try and use um is that, you know, for every new feature that we feed into the model, every new kind of piece of information we feed in. Whether it be a technical indicator or some sort of output from another model or or or whether it's, you know, some sort of piece of data off the internet that we get from some alternative data source. Every one of those, I really want to see like a thousand independent predictions.

with good statistics out of sample uh before I'm comfortable that I've I've got enough out of sample data to know I haven't curve fit. Because the power to do machine learning is also the power to curve fit. And you have to be very disciplined about drawing a a tight confidence interval around your answer if you don't have enough data supporting it. And that data's gotta be completely independent of your your training process.

Now just before you used the term features, now this is a common term that comes up uh when talking about machine learning. Can you just explain what features actually refers to? So features is synonymous with data inputs. It could be if I want to build a machine learning model based purely on technical indicators, eight one feature might be the twenty day moving average. Um, another might be the distance to the twenty day moving average.

A third might be, you know, the distance between a twenty day moving average and a thirty day moving average. Um those all three of those things would be just independent features that may or may not get used by the model.

Um but what we mean is inputs to the model. Now, a buddy of mine is much more advanced on machine learning than what I am. I mean I haven't really even got into it uh too much, but I told him I was going to be interviewing you and um I asked if uh there was some things which he might like me to ask.

or or would be he thought might be worthwhile asking. Um, just because we've we've talked about uh machine learning a little bit on the podcast previously and I kinda wanna explore some new ground. I mean we definitely have already done that, but just to take things a little further How do you go about creating features? Uh is this process automated or is are all the features created manually?

I I think in the long run we'd like it to be automated at present. I think most people are are um spending a lot of their time and their creative energies coming up with better features. Um in a way the signal um the prediction process or the signal process uh the first step um can be broken up into some substeps and and kind of the the middle of it has been automated by these machine learning tools.

And so as quantitative traders, uh as creative people, you know, we get involved at the very beginning at uh engineering the features and then we get involved at the very end at analyzing the results. And, you know, really the software and the and the computers kinda do the work in between for us very efficiently. And the ability to kind of iterate on that that workflow very quickly, very rapidly.

is is really um how people make make a lot of money in this business nowadays is by being able to rinse and repeat that over and over again. Um but we um in the long run aspire to use, you know, some of the more advanced machine learning techniques like deep learning to, you know, automate the feature creation process. Um uh and that's definitely something we're very interested in. I I think, you know, for the most part, I don't know of anybody doing that. So

And also another question I guess on a on a similar path here. How do you narrow down the search space to settle on the actual features you feed through your machine learning algorithm? Well there's there's everybody has their own kind of pet way of doing it, but I'll tell you one that's um easy to understand and and makes a lot of sense to me. if you were to build um a a single decision tree, which is really just uh a simple, you know, one of the most basic machine learning algorithms.

what you what you do is you come up with a set of rules that most efficiently classifies, you know, the good and the bad trades the way you'd like them to be classified. And it does something, you know, called min uh entropy minimization to do that. Um and that's in the end, it's just a math problem um that somebody else solved for us and we get the benefit of. Um and so as long as you prepared your data in the standard format and feed it in, it will spit out.

uh a decision tree that is just like any other decision tree. It tells you how to how to classify something. Um if you look at the things used by the tree and you count how many times one of the inputs was used. in creating your tree, you get some sense of the importance of those inputs. And so one of the methods of figuring out which features to use is to create um a a bunch of decision trees and count how many times each feature was used.

And then rank them from most frequent to least frequent and take the top, you know, X percent of those features based on frequency of use. All right, well just before we move off of machine learning here.

You know, for for str some traders who maybe have a little bit of programming experience or are st are keen to start learning, when they are at a point where they feel comfortable taking on machine learning, if that's something they wish to go into. Where's a good place to start? I mean I know you mentioned a couple books before. Is there anything I guess uh beyond books like uh any I guess uh like exercises they could try to do. Uh do you have any suggestions on where they could start?

Yeah, I mean I I think that, you know, one of the things that you want to do is is, you know, find some data sets. You can go to you can go to Quandle. Um we certainly um um have a lot of people who who use our data sets, uh CloudQuant. Um um internally and on our farm team to to do that type of work. Um And so you wanna you wanna find data sets that hopefully nobody is looking at, maybe things that haven't been picked over completely. Um and you wanna create your kind of ideal trades.

And you wanna, you know, alongside them put all of your your features. And then, you know, you want to basically, you know, get the book that I mentioned, uh, the Python Machine Learning and you know, try logistic regression or or try um try to do a decision tree or an ensemble of decision trees called random forest. Um all of those I think will be pretty straightforward to to set up and will give you a sense of what it's like to use a machine learning algorithm um you know pretty quickly.

All right, Morgan. Well, like I mentioned earlier, we have spoken previously we spoke uh about a week ago. One of the things uh which you also mentioned during that call, uh, which just sort of came up, uh when we were talking about sharp ratios. You said it's pretty straightforward to be able to get a strategy which has a sharp ratio of greater than three or maybe even a sharp ratio of four. How do you go about this?

And what are the characteristics of a strategy which can achieve a sharp ratio of three or four? Because not every sort of th that's a specific sort of strategy, isn't it? Yes, it is. And I I I think part of it, you know, w we we uh we have to provide some context. We have a little bit of a an advantage um with our you know, our access to Cloud Quant and the the high frequency simulation capabilities.

Um I think it it should be said that, you know, provided um you have access to these types of tools and you have access to interesting data sets. I feel like most of our people that we've we've brought in and most of the people we work with um on the crowd research side uh through our research partnerships um with CloudQuant.

have been pretty pretty successful at coming up with with good strategies. Um so we have, you know, probably uh at any given time, you know, thousands of of users, you know, coming to our free website and and starting, you know, to do uh quantitative research, you know, every month and

And we have we have a lot of people that have, you know, come up with strategies and and ask for funding. And so um you know, I guess what I'm saying is on the basis of kind of my anecdotal experience working with crowd researchers, and the new people we've brought on board. Um we've had good success with people um being able to come up with high sharp ratio strategies.

Um now, you know, a lot of times they don't all have the capacity you'd like. You know, they can't trade large amounts of money. And that's that's really ultimately the holy grail is to find the

you know, the fifty or a hundred million dollar strategy that also has a sharp ratio of four or five and, you know, has an unlevered return of twenty to thirty percent. Those are truly hard to find. But there are people finding them. And as as we add more data sets and add more tools to make it faster for people to complete step one of the model design process, we're finding that people are, you know, increasingly becoming more efficient at doing that.

Okay. Well I mean I think probably most people who are listening to this right now would be happy to have a a strategy which has a sharp ratio of three or four and be able to run a hundred thousand dollars through it. So I mean we're y you know what I mean? Like

Uh what are the characteristics of a strategy though which can achieve a sharp ratio of three or four? Because, you know, it's obviously gotta be a very short-term strategy. It's gotta trade very frequently. Does it have to be I guess an HFT strategy. No, it doesn't. And so what the way I look at it is I look at kind of what I call the Kelly edge. I look at the fair value win rate given the average profit to loss of the ideal trade that we're trying to capture.

Um and I look at the actual trade win rate and I look at the spread between the two. And I look at how many independent trades can be entered into on a daily basis. Uh and I can then calculate essentially the daily win rate. And so these types of strategies typically win, you know, sixty to seventy five percent of all days. But they may only uh win on, you know, fifty one to fifty eight percent of all trades.

If that makes any sense. But because you're placing lots of bets, the aggregate power of your your edge over the fair value win rate. plays out over a large number of trades, just like somebody who's counting you know, who who's counting cards in blackjack, um, who who knows their odds. Um, if you play enough hands, then you ultimately beat the house.

And the same thing applies here is you don't necessarily need to always make money. If you can find something that has a certain amount of statistical edge, where you have lots of opportunities to trade lots of stocks. Um and this may be the playground for institutional traders rather than more retail traders. But I've seen I've seen, you know, retail traders who have have um, you know, used our platform and

and, you know, come to us with some some good ideas. And so, um, while we don't facilitate we don't actually facilitate people, you know, trading um through CloudQuant, we we end up funding their strategies and running it ourselves. and paying them a share of the performance. But um but there are people who, um whether they come from uh uh a large investment bank or whether they are a really smart engineer or whether they just have always loved trading.

Um, they come up with things and sometimes they're basket trading strategies that have the opportunity to place lots of bets and take advantage of small statistical edge. So I don't know if this is a dumb question, but how many trades would it need to take uh would you want to take on a any given day on average?

Well, I'd like to make as many as possible that um I think fall in the category of having positive edge, which may seem like a a cop out answer, but that we really cast a w we we cast a wide net. We try and trade as many different stocks as we can. Anything that we feel like we have an edge in. All right Morgan, well let's talk a little bit about CloudQuant um and then we'll probably wrap this up. So

Just so that anyone listening to this um knows a little bit more actually about CloudQuant as a platform, what data is available to use as you've you've kinda hinted throughout this that there's I don't know if high frequency data is the right word and different types of data too. Can you tell us a little bit about that? Sure. Well um w we do actually have have the high frequency data. So uh CloudQuant's the world's first free um cloud based um backtesting platform.

Um where we provide any user who who cares to register on on uh cloudquant.com access to high frequency trade and quote data uh for all US equities and ETFs. Um we also Also provide, you know, kind of the the number crunching, the the clock cycles, if you will, um for free. And um we have earnings calendars. Um we're gonna be rolling out uh news sentiment data. And um uh we have some fundamental data and we have a number of other data sets in the pipeline that we're adding.

um that will give access to all the crowd researchers. And the way people use that um is, you know, we are an incubator for quantitative strategies. I think a lot of people have been challenged with raising capital.

for their strategies and kind of taking it to the next level beyond what they can fund with their own personal account. And um we notice that with the confluence of crowd researchers or just independent thinking people who are interested in the markets, um, which we we kind of call cloud researchers, but uh crowd researchers.

Uh and then cloud computing and alternative data sets and machine learning. All four of those things have kind of come together. A few years ago, um Andy Kirshner who was on your um on your show, you know, a few weeks ago He had this idea that this was gonna happen and he started building a technology platform to allow

um really anybody to have access to institutional quality research tools. After you know, five or six years um of working on this, we've finally gotten to where it's it's ready for uh the crowd researchers to start using it. We've been using it internally to manage um about a hundred and hundred, hundred and fifty million um and we're we're growing that number um very rapidly now that we've added to our team of internal quants.

And we we provide this platform for free so that uh people can uh push a button that says fund my strategy and um we will take a look at it, compare it against our existing strategies and our um years of experience. and try and figure out if we think it's something that has staying power and um allocate, you know, somewhere between a quarter million and and five million dollars to to strategies that uh we think have capacity to handle that.

We have other people that have bigger strategies that uh would get larger allocations, but um that's kind of our target area for what we're shooting for initially. And uh we have uh probably about a dozen to two dozen people who are in the process of

going through that allocation process right now. Can you just expand a little more on how you determine which strategies actually get an allocation? So as we've been trading quantitatively with CloudQuant internally for um a little over five years and Personally, I've I've been trading quantitatively for probably twenty-five years. Um, you know, we have uh a significant amount of experience with not only how our platform behaves in real life, but um how simulations in general um

perform in real life. And so we um we do a lot of um analysis um that's proprietary. to help kind of figure out um if somebody has, you know, curve fed or or done something that's um not gonna lead to a robust, you know, um live performance. Um so a lot of that unfortunately I can't kind of go into, but uh we do have a a methodology for analyzing things and we had a lot of um of success stories internally that we can compare things to.

Um we can compare you know the sharp ratios that we see from um the internal PMs um and and look to to those um comparisons as a way to select crowd research strategies. Um and one thing I want to mention is um the crowd researchers um, own all of their strategies and it's completely private to them.

up until the point they they decide that they would like an allocation, would like to share in the profits, then we essentially have a a small group of individuals that partner up with them and help them with the trade expression part of their algorithm. Um and so those people obviously would work kind of hand in glove with them and kind of perfect um their strategy for for prime time. But uh for the most part, um most users come to our platform and

We really can't even know what they're doing, um, because of how it's set up. It's very private. There's a very strict privacy policy and uh nobody has access to any of that uh information. And so the only way we would see um their activity is if they're

showing up on our leaderboard as having a high sharp ratio or a great rate of return, then we might um, you know, kind of give them a pat on the back and see if they are interested in an allocation. Okay. Yeah. No, I'm glad you mentioned that. It was something I was gonna ask you about actually.

Um, are you able to share what the profit split is? I'm sure some people listening would be interested to know what that is or does it vary on a case by case uh basis? Well it it's actually pretty amazing. It's um it's It's really quite generous if you look at, you know, kind of the hedge fund business and and how uh challenging things have become. Um obviously, you know, fees are dropping, you know, very quickly and

But expenses are going up and uh some management fees are going down. Expenses going up is not the the place you wanna be. We uh view, you know, CloudQuant as the first opportunity to generate alpha at scale using those four key ingredients that I mentioned earlier. Um and so we we kind of view um the crowd researchers as partners in this this business. Uh they're a crucial part of um the research group.

Uh, we actually, as of last month, um, realized um that we were doing more research with crowd researchers on our platform than we were with. um internal researchers on the platform. We have nine we have eight uh internal portfolio managers. They actually in aggregate with all their quant trading teams do less research using the same tools as the crowd researchers than the crowd researchers do.

in aggregate as of last month. Um expect that trend to continue because um we have people in eighty different countries doing research actively on the platform and we are expecting that that those numbers are gonna gonna grow. So, you know, I think at the end of the day, we pay half of what you'd normally get as an incentive fee. So twenty percent incentive fee is what a hedge fund typically gets. We pay ten percent to the crowd researcher because we value them so highly.

Um, and we also create opportunities for them to increase that percentage even further. uh by working closely with us um on uh furthering their research and um and by achieving high you know performance you know metrics. Sorry, can you just repeat that? It's ten percent profit split to the trader? On the baseline, that's right. Of net profit, yeah. Okay. Right. Very good.

All right Morgan, well let's uh leave it at that. Um where's the best place for listeners to go if they wanna find out a bit more about yourself or and also about CloudQuant? Well I'm not as interesting as CloudQuant, but uh refer'em to cloudquant.com. Um And, you know, we have uh they're welcome to follow us on Twitter, LinkedIn, um, and uh, you know, we're constantly posting new updates on on alternative data sets that we're adding and new features that we're adding to the platform.

Uh we really are there for them and so we appreciate their feedback and um would like to uh get as many people to um to you know take advantage of this opportunity to Essentially be our business partners um and come up with trading strategies and hopefully um all of us can make make lots of money. That's the goal. What's uh what's the Twitter handle? Uh it's uh CloudPoint.

At CloudQuant. Okay, cool. All right, Morgan. Well I just want to say thank you very much for doing this. I appreciate you taking the time. It's been uh it's been good speaking with you. Thank you, Aaron. I've really appreciated it. And um you have a great day. You've reached the end of this episode of Chat with Traders, but rest assured there are more. way soon.