Up next in 10
Dr. Tom Starke - From Physics PHD to Quant Trading Virtuoso - Ep. 8
369 views
Dec 11, 2024
Have you ever wondered what happens when a physics PhD collides with the high-stakes world of algorithmic trading? Join us as Tom, a former academic turned quant trading guru, shares his extraordinary pivot into finance. ✉️ Learn the secrets of legendary traders and investors: https://bit.ly/4aCna7i 🔗 Follow along: https://analyzingalpha.com/episode-008-dr-tom-starke ⏱️ Video Chapters 0:00 Algorithmic Trading and Trading Strategies 17:51 AI Trading Strategies and Portfolio Management 32:32 Algorithmic Trading 37:24 Financial Markets and Human Ingenuity 41:43 Quantopian's Return and YouTube Channel 49:08 Discussion on Interesting Papers and Value 🏷️ Tags #trading #algotrading #quantitativefinance ---
View Video Transcript
0:00
hi Tom I can't wait to speak with you today I want to know how you got into algorithmic trading because I know that
0:05
you have a background in in physics is that correct yes that's right so so uh well
0:13
thanks F first of all uh for having me and um yeah really appreciate uh to be
0:18
on your podcast and yeah it's true I've got a background in physics I did a physics PhD of all things um and of
0:27
course being a physicist is not an easy life these days especially if you're on academic and so um you know I was
0:35
interest always interested in in many different fields and so so I really was
0:42
more more into casting my web a bit wider and and just seeing what's out
0:47
there in the world and and using what I've learned uh during my time in physics to apply to other maybe a bit
0:54
more real world problems so I I I guess that's that's that was my
1:00
main intention with this I think I'm just not genius enough to be a physicist
1:05
in in that regard What attracted you to trading um well yeah it's an interesting
1:12
one so so I basically worked in engineering uh for for quite a while and
1:18
on on many different problems from from jet engines to semiconductors um and I
1:25
always worked a lot with computer simulations and so as so I also used to
1:31
do really quite large scale computer simulations on distributed servers literally hundreds and thousands
1:38
of them and at some point I thought well if if I have all this computing power and and I have the understanding maybe I
1:46
can just uh build something that that automatically trades the stock market
1:51
and um makes me money while I'm lying on the beach and and at the time really
1:56
that was really before retail Traders had any access to systematic automated
2:03
trading systems so it was like a crazy idea but I got quite obsessed with that
2:09
I thought oh maybe I'm I'm one of the first people that ever even thinks about that and uh yeah I got into it I had no
2:17
idea about Finance or trading when I started so I basically I'm I'm I'm I'm
2:24
entirely self-taught you know and I I absorbed I don't know unbelievable
2:29
amounts of of literature on trading and and everything uh Beyond and
2:35
finance and I I I guess now I'm I'm I call myself reasonably confident H
2:41
competent in in that field you know with that being said you haven't really gone from the the discretionary route right
2:48
you know I know we had a preall previously where you said you've actually never traded maybe from a
2:54
discretionary chart perspective things like that you've taken a completely different approach can you kind of tell
2:59
me a little bit about that yes yes it's it's it's an interesting one actually because I read a whole lot of books
3:04
about discretionary trading um and you know how you interpret charts and so on and and quite frankly I didn't
3:13
understand any of it I looked at this and and I just thought that doesn't make any sense to me you know I'm obviously
3:21
I'm coming from a background where interpretation of data is is is important you in in engineering and
3:28
physics and I look at this and I'm I really didn't understand uh uh the
3:34
way Traders were thinking so what I basically concluded after reading maybe
3:39
10 books or so on that subject I I just don't know I I don't understand it so I need to do something
3:46
else and um you know I I thought I basically
3:51
started straight away just just just punching time series into my computer
3:57
into python analyz ing them looking how you analyze them what you can do with it
4:03
predictive modeling all that sort of stuff and that made me feel a lot more comfortable and it's it's it's
4:11
interesting because I know there's a lot of discretionary traders that are incredibly good at this and they're making money and to me it always seemed
4:17
a little bit more like a black art something like astrology also
4:24
uh and I guess you know some people are incredibly good at that but I'm definitely
4:30
uh more drawn towards really hard uh testable
4:36
fundamental uh things so so I ended up doing it that way and
4:43
also I I looked into into executing manually and I found that extremely
4:49
tedious and I didn't really want to spend a lot of time on the computer so I also straight away looked into what what
4:56
are the ways to execute in a in an automated f fashion and and what which
5:01
was good in a way because by by trial and error and learning a lot of things I actually learned quite a bit of computer
5:08
science on the side as well which helped me a lot uh to to further on you know to
5:14
to really get a broad view of of the whole subject um how about some of your early uh I guess trading strategies uh
5:21
you know what what did they look like obviously you're one of the first ones uh using big data so to speak in
5:27
creating your trading strategies you know what did you find early on and it did it work out for you um well I I
5:34
actually I literally started with a moving average crossover um so so at the
5:40
time there really wasn't a lot of information out there on on this subject
5:47
so so there was really just two or three bloggers that that really talked about
5:53
this and then I joined them um as a blogger as well in the hope that that
5:59
maybe some people have come and and have discussions with me uh about about these
6:05
things and so I really just started with extremely simple things and and look
6:12
like how well do they work and and you know I went basically through the whole
6:17
sort of traditional technical indicator uh world and and looked at all these and
6:25
and and found you know it's it's it sort of works but it's not great and then you
6:31
know slowly started combining that with more uh statistical methods uh
6:38
tools and and it just really it just really went from from really simplistic
6:44
to more complex the interesting thing is that as you go that there is a sort of
6:50
inflection point where suddenly you start to everything becomes more simplistic again but not simplistic in
6:58
the way where you started uh as basically as the as the fool that doesn't know where to go you you
7:04
suddenly have much more deep insight and and uh the Simplicity comes from your
7:10
from your understanding of all the things that don't work I suppose and so yeah so so I I went from
7:18
more complex again to to probably more simple but but not simple in in in
7:24
stupidly simple but more that the sort of educated side of simple I suppose so that's great so I guess in regards to
7:30
today what what are you doing that obviously isn't too simple but you know it's fundamental in nature of the
7:36
markets um yeah so so I mean a lot of a lot of what I do really involves because
7:42
I build strategies that can handle very big volumes uh and and you know large
7:49
large amounts of capital and so there's a few challenges in that and and the
7:54
challenges are mostly not uh coming up with eight trading strategy that that is
8:01
that rules them all but it's more about managing the risk appropriately um
8:09
managing at the portfolio um and and also understanding how execution or what
8:17
what execution does how that how that skews the uh the portfolio performance
8:24
and then of course scaling up and down so I would say like when you know the
8:30
underlying trading strategies that that that are generally used that they're not not particularly complex but the
8:37
complexity comes with the implementation of it you know as I said making sure the
8:42
risk is appropriately managed the portfolio positioning is is really well
8:47
done all these things they are the things that are in my opinion at least
8:53
uh far more important than coming up with uh some some combination of of
8:59
whatever indicators or so um that's almost the easy part but but and and
9:06
that's the quick part then the more difficult part comes into adopting um a strategy that they want to
9:14
trade uh to a specific uh I guess uh risk uh Benchmark
9:20
and and or risk you know risk assignment uh uh to to the to the capital that that
9:27
you have and into other constraints that that you may have uh uh trade you know
9:34
trading other people's money and you bring up a a great Point obviously a portfolio uh which I think that's you
9:40
know from uh my understanding and and discussions with other people that's you know where you want to head to
9:46
eventually as a Traders as a portfolio of of uncorrelated strategies for multiple reasons but um where would you
9:52
say uh for a new Trader um you know what direction would you point them into if
9:58
they were going to start trading okay so so let me um let me perhaps point out
10:03
two things first of all um Harry marovitz famous Nobel Prize winner said
10:10
diversification is perhaps the only free lunch in finance and coming from a Nobel
10:16
Prize winner I think he wasn't that too wrong with this so whether you are a
10:24
small Trader or or a big gun you know trading a couple of billion dollars whatever
10:30
it is still it is still interesting and important to understand that that diversification can help extremely uh in
10:38
in the performance of anyone's trading and the other thing is that um
10:44
there there is a book which is I call it always a little bit the Bible it's it's
10:49
a book by grinold and Khan called active portfolio management and they have in
10:55
there an equation which is called the Golden Rule of asset management and and what it says is IR which is the
11:01
information ratio or also the sharp ratio or risk adjusted return equals the information
11:07
coefficient uh times the square root of breadth now that sounds a bit like a handful but what this means is that your
11:14
risk adjusted returns are are sort of consist of two components one is the
11:20
information coefficient which is really the skill that you have as a Trader and the breath that you have or you could
11:27
almost say this is a proxy not not exactly the same but but a proxy for how
11:32
Diversified you are and um and so diversification is not just like
11:38
different assets it's different asset classes different time frames uh
11:44
different strategies all this sort of Jaz so so there's you know there's a lot to diversification
11:50
but what's interesting about this is people say like you mentioned Warren Buffett they are incredibly skilled in
11:57
in picking the right stocks or at least I think they are and so their information coefficient is very high but
12:03
their breath is really low but for you know a lot of people including myself
12:09
where where my skill isn't particularly uh great uh we need to in order to get
12:17
better uh performance in terms of risk adjusted returns we need to increase the breadth of our strategy so that's why uh
12:25
that's why this is important and of course that only scales with a square root so so we're not you know it's not
12:31
linear but nevertheless I think even if you're starting out it's worth a
12:37
consideration that rather than trying to find uh the one strategy that rules them
12:44
all you may consider actually um you know not not building or not trying to
12:51
find that that amazing strategy but but actually combining a few strategies and
12:56
getting better performance that way I I think I I think this is probably something that I
13:02
learned not too early and I wish I had known this earlier that's funny that you mentioned that did you try coming up
13:09
with initially a Holy Grail strategy of course of course I did and and it's it's
13:17
quite a sobering experience you I I it's actually quite interesting um and and you know I've
13:24
I've built this this system that basically generates uh uh hundreds of thousands or millions of of strategies
13:31
randomly and you know they're all like you know just just trading strategies and what what you see is that
13:39
that or you would expect that when you do this randomly there would be some single strategies that that actually
13:45
perform really really well and of course some of them do okay but what's fascinating is that there is
13:54
some sort of a limit of how good the performance can be uh when you have a strategy that is
14:01
purely based on historical price data
14:06
so in in a in a way when when you look at historical price data the the out of
14:11
sample predictions that you can do at least for my gut feeling there is actually a limit of of how or what you
14:18
can achieve and obviously this limit changes over time um but it's it's definitely visible in in
14:25
what I've seen so far so my my feeling is of course in Sample uh you can build
14:32
strategies that that return unbelievable amounts but out of sample I I I do think there there is
14:40
really a limit of of how well a strategy can perform and I'm not sure exactly what
14:45
that is yet that's probably for the academics to work out now that's really interesting yeah I I still remember you
14:53
know my first trading strategy where it ended up getting so complex that by the end of the by the end of it I didn't
14:59
even remember what I was doing right um then I've I've come to learn that you know a diversified set of strategies is
15:06
is for multiple reasons the best way to go but but that's really interesting the point that you bring up um about those
15:12
random strategies um what do you think the cause of that do you think it's because the the the market participants
15:18
learn and you know then the the strategy sort of comes discounted in in the price
15:24
or due to history or what's your what's your thoughts well well I mean first of all the amount of information that you
15:30
can extract from historical uh performance is obviously limited
15:35
and and you know that there's clearly events that are not related uh to the
15:41
historical performance of your price data which play quite a significant role in the movement of stock market or
15:48
whatever asset prices in general so you know even even let's just use the
15:53
example of insider trading uh if there's some insider trading happening this cannot be
15:59
included in in past prices so whatever you extract from there has got a a very
16:06
strong uh limitation uh and and so so
16:11
that will always mean that that what the the amount of of
16:17
information that you can extract will always be suboptimal it will never be perfect and so the the key is to
16:25
question how can I extract extract in enough information U from those series
16:31
to build a strategy that is good enough for my needs and often you know and as
16:39
what you said Rings very true for me as well I started uh building strategies
16:44
more and more complex and and realize well this is this is really not the way to go um because what happens is yes
16:52
they work well in back tests but but the uh the predictive power of them the out
16:57
of sample information that that carries through to to a market yet unknown is
17:03
quite limited and it turns out that often comparatively not not extremely
17:10
simple but comparatively simple things can be hugely effective rather than uh
17:17
trying to make things too complex um the complexity is really it's
17:22
it's an interesting bias in our mind we we see those historical price charts and
17:27
we say well if we add something here and we add a moving average filter there and and some other filter here oh that must
17:34
be you know we get the perfect thing but but this is just not quite how it works
17:39
and I mean anyone who's done a bit of Quant analysis obviously knows that quite well yeah that's funny I feel like
17:45
it's almost a right of passage you have to discover that that eventually um but
17:51
no that's great so I guess in regards to a uh you know a strategy what what do you can what what do you consider um you
17:59
know a strategy whenever you're you're thinking about you know randomly generating all of those strategies well so so first of all um for me I don't
18:07
think in terms of of entries and exits um so so this is something that that I
18:13
had to let go reasonably early on because I mean and I'm not saying it's
18:19
wrong but but it's just not the way I think the way I think is in terms of
18:25
rebalancing portfolio positions so my Strate is usually that portfolios and
18:30
then these positions change so there isn't really an entry and exit as such
18:37
you know you always have some sort of position on and then it scales up and it scales down
18:43
um and and that uh that scaling is is driven by a few factors uh uh one of
18:50
them is obviously a predictive model of some sort so so so this is where the creativity
18:58
obviously comes in uh from from a Trader to to find interesting uh predictive
19:04
models that um you know have at least a little bit of of Alpha in them to to
19:10
make them worthwhile and then of course you can combine them and do whatever you want with them but then uh the second
19:18
part that I think about is how do I then take those uh predictive signals and tie
19:25
them into my portfol man my portfolio so so
19:30
so basically these predictive signals then inform how I position myself in the
19:37
market with with each different product that I trade so so I need to think about
19:43
you know it's it's effectively position management how do these signals then
19:49
come into into uh uh building my portfolio but of course portfolio
19:54
building isn't just about the predictive signals it's also about how much risk uh do you want to take
20:00
with a single position and so on so so basically the the second step is is
20:07
really thinking about portfolio construction and portfolio management and some a lot of people have heard of
20:14
the mean variance portfolio which is sort of the the most basic uh uh idea in
20:21
portfolio management which I think also comes from Harry marovitz which I mentioned earlier uh but then there's
20:28
there's obviously more sophisticated ways to do that uh you know you can use
20:33
Kelly which in some edge cases actually converges to to a mean variance
20:41
portfolio but you can also use something uh like deep learning and um um I've
20:47
actually put out a new course recently on quantra or it's actually just coming out now in in end of January about using
20:56
uh deep learning now I was always a little bit skeptical about deep learning in finance but what what you find is
21:02
that if you set it up correctly it can actually perform or better out of sample
21:08
uh than your standard uh mean variance uh portfolio stuff now now that's the
21:15
second part but then the third part is execution so execution is something that
21:20
is often quite neglected by Traders uh they go well we cross the spread uh we
21:26
get filled uh but but we lose quite a lot of of money with that when we cross
21:31
the spread because you know you're you're you know you're paying for for the privilege to execute in that
21:38
moment and therefore if if you actually think about execution systems uh especially
21:45
when you trade quite large amounts of capital just just not crossing the
21:51
spread all the time really can save you millions of dollars per year but even um
21:57
for for a a retail Trader when you when you have a strategy
22:03
that seems to perform well but but it's very um you know the the profits per
22:08
trade aren't that great having a good execution where you don't cross the
22:14
spread all the time um but but you maybe you're quoting also uh that can really
22:19
help to improve the performance of the strategy so these these three components
22:24
I guess are the main ones that that I think about when I build um when I build a new strategy that's
22:31
fantastic yeah I'm I'm excited about for your new course I uh the first course that I took of yours uh a while ago was
22:39
on the neural networks and um reinforcement learning I tried forever
22:44
to build a reinforcement learning strategy yeah which I thought your course was fantastic by the way um that
22:51
was my first introduction the challenge that I realized after after doing it of
22:56
course of course um was that even and again this is when a portfolio comes into play um even when you come up with
23:03
a strategy that works you have no idea why it works and I just don't know if you have this staying power to keep it
23:10
uh keep it on so I think um from a from that perspective what's your thoughts about where AI is taking trading and you
23:18
know what's the room for discretionary or algorithmic traders to start using some of these AI tools well well first
23:25
of all I think AI at least in finan is a little bit overhyped I'm always very
23:31
skeptical about anything new and anything complex for that matter the way icai is really purely as
23:39
a as a tool as a means to an end now it's interesting because you see a lot
23:44
of the young data scientists that come onto the scene and they're claiming oh I
23:50
built this AI tool that can beat the market by a huge amount and all this
23:55
every single paper that that claims that I've seen so so far I I could very easily uh debunk um um it doesn't take
24:04
much um but that is not to say that that AI isn't something worthwhile it
24:10
definitely does something interesting which is it also takes into account
24:16
nonlinear um correlations in in in the in the price data so so ultimately what
24:23
what you need to the way you need to think about is if you want to build Trading I is purely on Price is
24:30
that you you're really just working with correlations within the data and you the
24:37
the traditional way to do it with with most indicators and so on is fairly linear and then and then um some
24:43
indicators and then AI tools uh they actually extract nonlinear uh
24:49
correlations and and so the way I see it is really it's uh in some ways I always
24:56
call this a more GL glorified regression engine and I'm definitely not overly
25:02
excited about it but I'm I'm happy to use the tools if they if they help me
25:09
but as I said you got to be really careful about this to say well what I
25:14
really want is a trading strategy that is based on reinforcement learning but you don't really have anything else it's
25:22
probably not right you you need to first come up with some ideas of what it might be and say well well how can
25:28
reinforcement learning help me to support my investment thesis and and and
25:35
my goals and then apply it that way rather than starting with reinforcement
25:40
learning which is what a lot of people do they they just get so excited about those AI tools and they start with them
25:48
and in my opinion at least that's not a good approach I think the good the best approach is to
25:54
start with a solid idea of of of some description and then pick the tools uh
26:02
that that that you uh that you want to achieve that um and so so that way uh it
26:09
is much more robust uh it is much easier to achieve what you want rather than
26:14
just just getting excited about the tool it's a bit like you know someone gives you a hammer and a chisel and you say
26:21
wow wow I've got a hammer in a chisel now I sculpt the most beautiful sculptures and that's just not how it
26:28
works you know you need to First have have a concept of what you want to
26:33
sculpt and and and and and some sort of artistic um ability be and and then the
26:39
hammer and the Chisel is just a tool you know it's not it's not it's not what actually makes the sculpture
26:46
work I'm not sure that makes sense but I hope from my experience it makes perfect
26:52
sense right you know everyone wants to just create this black box that prints them tons of money but um you know just
26:58
it just doesn't work that way but but there are some some use cases the use case for portfolio management is pretty
27:04
strong but another another place where those tools can be very handy is
27:10
in execution um and and so so there's quite a few problems in execution where
27:16
reinforcement learning but I have to say not deep reinforcement learning but standard reinforcement learning is is
27:23
very routinely used and it's it's relatively easy to see that this could
27:29
be extended to deep reinforcement learning the only issue with that is that often execution relies on on low
27:36
latency and and and one of the reasons why deep reinforcement learning hasn't been used so much in execution yet is
27:42
because it's comparatively slow and you know it really uh it it really destroys
27:49
your your latency so once once we get faster machines and and faster
27:55
calculations I'm pretty sure that that will uh creep in as slowly as well
28:01
that's great to great to know in regards to your back testing you know what um platforms or systems are using are you
28:07
just using pandas or um you know what are you doing there so I build all my
28:12
back tests from scratch uh in in in Python I don't use uh platforms
28:18
myself um it's it's a practice that I started reasonably early on um because
28:24
what it does is it really forces you to always reconsider consider what you're actually doing every time and every time
28:30
you build a back test you get a little bit better than than what you built
28:35
before so so you get a real uh routine in in in building back tests
28:42
quickly and and you can build all the assumptions that you want to make in yourself one of the issues that I found
28:49
with platforms quite often is that there is inherent assumptions in the platform which are not necessarily always
28:58
clear and obvious and the code is is often not open source and so it's it's
29:04
actually very difficult to see in the end whether your profits and losses come
29:10
from your strategy that you tested or they come from inherent assumptions in
29:15
those platforms now of course they are very convenient but um I never felt like
29:21
uh the convenient route is is the one for me so so I I definitely uh build back tests myself
29:29
and and one word uh to back tests one of the standard practices that I use is I
29:36
always very cleanly separate the signal generation so the predictive part of the
29:42
modeling the portfolio uh sizing and the execution from each other so
29:50
so so I usually not mix them up in my back tests so once once I have my predict predictive signals then then I
29:57
move on to to the portfolio construction and then once I have that I look at the
30:03
execution but um I I try to I try to
30:08
really have a clean separation and not not a mix up because it just gets too complicated and confusing and my my
30:14
little brain struggles to hold that all in one place so I rather uh I rather do
30:21
it step by step and it it makes it a lot easier to to understand um and and and
30:27
to deal with and you can also see well okay what is the p&l that is generated
30:33
by my signals what's the p&l that is actually generated by doing my uh
30:39
portfolio management and and and and the same with execution you can really see where the
30:45
source of your performance comes from rather than mixing it all up and then and then having no clue and then you can
30:52
tweak so many buttons at the same time that it's it's almost impossible possible to find a really good sweet
30:59
spot in that yeah that's great I and I I agree with that advice too I started uh
31:05
with a platform and if I could do it all over again I would definitely go with uh
31:10
you know just Python and Panda simply because um I think you gain you gain
31:15
another skill set that's more Broad and there's just so many ways to to use it um but then that brings me to the
31:22
question what about uh execution um do you then take that out your the algorithms and move it to a different
31:27
system or how do you go about doing that just because of the risk of you know coding errors when moving it to another
31:34
platform well it obviously depends like when when I'm when I'm putting my my retail head on um probably a good way to
31:42
do it is to have some sort of system that that generates signals and does the
31:48
portfolio sizing and so on and then um you you probably want to build something
31:54
like a separate system that handles the ex execution and then have them communicate so so in in in you know in
32:02
in today's computers you know there's nice uh uh packages and tools like like
32:09
zero mq also that have different programs communicate with each other for example and so what you can do is you
32:16
can actually build some sort of execution system and all all you need to do is just Supply that with uh Market
32:24
signals and then it it it handles everything else for you and you don't need to worry about it so much anymore
32:31
once it's done and um I think it's a really good way of doing it and and you know a lot of the
32:37
Brokers that today they they also have execution algorithms that often give you
32:44
a better a better uh executed price than than you would normally be able to
32:51
achieve when you just cross the spread so if you're not yet uh want to build your own execution algorithm and and
32:57
deal with that a good way to do it is to use someone else's and and yeah a lot of
33:03
Brokers like I mean I'm not affiliated but but for example like just just what I know IB they have a whole Suite of
33:11
execution algorithms for free uh for anyone to use and you just go and pick
33:16
what what you think uh suits your needs uh the most and I think it's a it's a
33:23
good way of doing it of course when people start really early on I also tell
33:28
them maybe it's actually a good idea to start with manual execution first to
33:34
actually submit your trades by hand just to see what happens because there's a whole lot of issues that that can arise
33:42
from uh automated execution and the last thing you want is to have an error in
33:47
your code and then suddenly you have a million orders in the market and and you don't know how to close them out again
33:54
something like that I mean most risk engines will prevent that but there's definitely been these cases uh uh where
34:02
this was so and and and you know firms got uh got completely obliterated I mean
34:08
night capital is a famous one but there's plenty of others where the same case same thing happened as well and it
34:15
can happen in retail too so you know um now the other thing
34:20
is using a platform is is a bit bit more straightforward and and I'm not saying
34:26
you shouldn't use plat forms I think platforms have a really really important place and and they make it a lot easier
34:32
and sometimes I look at my my code and you know I'm I'm I I write decent code
34:37
but it's probably not amazing and I think man I should just use a platform this is just so what a mess um so you
34:47
know if if you if you uh can do that obviously the platform should have a back testing uh uh facility that and
34:54
then you can maybe compare your your your initial simulations with the back
34:59
tests in in the platform and and see whether they agree which is actually
35:05
more tricky than it sounds um but you know that there's obviously ways so so I
35:11
wouldn't say oh don't use a platform but yeah there there's alternative ways
35:17
and and I guess for me also I love learning so so so learning all these alternative ways is really cool because
35:23
you learn a lot of skills with it you know computer science skills and and algorithm skills and all that which are
35:31
extremely handy in the long run speaking of Technology you know what advances in
35:36
you know machine learning crypto or anything like that do you think um you know that that could be interesting in
35:41
the future for for algorithmic Traders good question uh it's it's really
35:47
difficult to to know that um because as as to technology changes the market
35:54
changes and so there's always always this ongoing uh it's it's almost an
35:59
evolutionary process of predator and prey right you know we have a I suppose
36:05
the Predators which is the the the the people with the technology and the prey
36:10
is the market and the other people um and so it's really difficult to know what's
36:17
going to be um interesting going forward because when you actually look at
36:23
evolution is fascinating you you see that that uh in in in as far as I
36:29
understand it and I'm definitely not very good in sort of biology all this but the way I understand it is we we
36:36
evolve to more and more complex life forms up to a point where complexity actually is becomes destructive and then
36:45
the whole uh the whole system starts to collapse and and reset in a in a much
36:50
more simple uh system which then evolves again and and so there's this constant
36:57
um this this this constant cycle of of increasing complexity and then
37:02
collapsing complexity so I I can easily see that
37:08
that this could happen in the market as well that that at some point the complexity becomes so
37:13
huge that in it it actually somewhat collapses and and and resets itself to
37:20
start from a less complex stage again and and because of that I I really
37:27
wouldn't want to even start to predict what is going to happen in in the future
37:32
in in financial markets or with tools I mean there's obviously interesting stuff like Quantum
37:38
Computing and uh you know the the the really deep machine learning algorithms
37:43
but you know there's there's many you there's many aspects of that that I'm
37:49
very skeptical about and I don't know if you heard this but only a couple of days ago Google researchers came up with a
37:56
paper that actually shows that even those really really uh deep
38:01
models um have not the ability to to think outside of the box as you will so
38:08
so they're only really reasoning on what they know but they're not able to come
38:14
up with things that they haven't been trained on at all not not even simple
38:19
things and I think this is really interesting because yeah you've got all these complex models but at the end of
38:26
the day they only know what they know and the Ingenuity of humans is that they can come up with stuff you know outside
38:33
of the box quite a lot and and this is what makes what really pushes things
38:38
forward in in reality and so yeah tools are all great and they're going to be
38:44
more shiny and more fantastic but I think the human mind will still be the
38:49
ultimate defining um uh property and and the recent Google paper again showed
38:57
that you know we're we're not actually that much closer to this general intelligence so I don't necessarily hold
39:04
my breath for that to happen anytime soon now that that's great I'm I uh
39:09
would love to check out that paper yeah it's interesting because I think about this a little bit too uh you know stying
39:16
a little bit from from Trading but it's you know how unique are we as humans you know when you really look at how the
39:21
mind works and neural networks work it's very interesting and I have a young daughter and I just remember when she
39:27
was learning uh you know just how to feed herself and walk it was a lot of repetition I'm like this is you know
39:34
this is back propagation this you know and I'm just thinking like hey are we really really that unique I I think I
39:40
think there is a lot more to the to the human uh story that that we we think
39:47
I've seen once a paper where where researchers actually looked at uh a brain connectivity and they they they
39:54
used MRI skins or something some technique like this to to look at how how messages get sent in the brain and
40:01
it was very chaotic but what they then did is they they actually mapped the space uh to a seven I think and an 11
40:10
dimensional space so they mapped all the signals instead of like the three dimensions in the brain to some seven
40:16
and 11 dimensional uh uh spaces which which is some mathematical procedure I
40:23
don't understand but it it turned out that suddenly when
40:28
they did that everything made perfect sense everything was very regular and that is not something that a
40:35
neural network can actually do yet it's it's we're far from that so the neural
40:41
networks that we use in the computers they're not necessarily a higher dimensional or they don't act as
40:49
sort of higher dimensional entities as far as I understand so I I think there is still a lot more to to discover in
40:56
the in the human brain that isn't there you know we're not quite there yet with our computer technology tell me a little
41:02
bit about you know what you're doing currently I know that you've got AA triaa quants uh you're doing a lot of
41:08
Great Courses on on quantra can you give us a little bit about that uh sure so so yeah triaa qu that's my uh my
41:16
consultancy uh company so I provide consultancy for financial institutions
41:22
and so on but but the reason I've also um Focus quite a bit um on um on
41:31
educational stuff so so I built a few courses and as you mentioned recently I built a course uh for quantra on on
41:39
using deep learning uh for portfolio management but also uh I don't know if
41:45
you heard that but quantopian has come back up again they're they're back online and and I don't know if you no
41:51
this but I used to work quite a bit uh with quantopian before uh they
41:57
disappeared and and so now they're back and they they have they start all the the the
42:03
CEO Force started a new platform and so I was always very uh I always loved
42:10
quantopian and I loved working with them and I ran workshops for them and I'm actually really happy to be back so if
42:16
anyone's interested check out quantopian there's some you know it's still small
42:23
relatively small but there's some amazing people on there and it's going to be it's going to be really
42:28
interesting as well you know um and it it provides you know DD space they
42:34
provide very different um I guess very different themes uh in in in
42:42
the Quant Finance world so quantra or or or Quant in or or quantopian or so that
42:48
you know and you know I'm always I'm always happy to be involved because at
42:53
the end of the day I'm learning so much from that as well so so it's
42:59
it's really worthwhile spending the time now I'm also um I'm also building um um
43:04
um quantitative or Alpha models as I said this is one of my my jobs for for
43:10
for a large uh financial institution so so uh that's that's a little bit outside
43:17
of of the other things but uh yeah that that's actually an enormous amount of
43:22
fun too and I think when you provide education it's always good to also walk to walk and not just talk to talk what
43:29
what I see is that that a lot of the people um online they you know they talk
43:35
a lot about this but they they not necessarily actually do the trading or or do the the the research themselves
43:42
and I mean my my my I would say my my workshops they're not necessarily very
43:49
slick or smooth but but you know I try to provide a different Edge for someone who actually works in that field
43:55
actively uh so you know it is a place for everything in the world and um yeah I'm having a lot of
44:04
fun with it that's yeah that's wonderful and no I am super glad uh quantopian is
44:09
back that was kind of where I got my start and there was so many great people that you know I've lost touch with that
44:15
were on there and uh hopefully you know it'll make make a comeback um and I also
44:20
saw you've got um a lot of videos that are they're coming out recently yeah yeah so so was quite a bit on YouTube
44:27
for free um so so the idea was I uh a friend of mine he he and he's he's a
44:35
good friend he wanted to learn um more about quantitative trading he's got a programming background I did a did
44:42
extensive you know programming studying programming but he didn't know anything
44:47
about it and I said well I'll you know let let's just sit down and I'll I'll teach you some and then I thought well
44:53
why don't we why don't we record it and just just um and just post it on on
44:58
YouTube so it's it's not it's not again it's not super slick but but what what it does is it it
45:05
really uh um I really take him through basically quantitative trading
45:12
from absolute zero and step by step basically teach him stuff that that you
45:18
really need to know and I think if you uh if you could deal with the fact that it's it's myself and and and and Josh
45:26
uh being you know to two nerds uh talking about this uh I think there's a
45:32
lot of really interesting information in there and yeah check it out if if you
45:37
want I think it's it's has actually a lot of really cool Snippets in there to to learn and especially if you start a
45:44
fresh and you want to go down the more I I would say quote unquote the hard way
45:49
uh that's probably good way good place to start nowadays it's wonderful to have YouTube and be able to have great
45:54
resources like that so so thank you from all the the beginners that are going to try that out sure sure I mean it's it's
46:01
it's actually a lot of fun to do this as well because it makes me often question
46:06
um you know what I what do I know and and go back to the real Basics and and
46:13
it it's it's it's always it's always a two-way learning in some sense and and that's that's why partly why I'm
46:21
doing it because even even if this is not stuff that that is completely new to
46:27
me I'm still learning new things while I'm while I'm doing it I find that teaching and or writing about something
46:33
verifies your thoughts right for for sure absolutely all right Tom as we uh
46:38
wrap up our interview can you share maybe one alpha drop or some unique Insight or maybe a controversial Insight
46:44
that uh that you feel that maybe isn't widespread or widely known okay okay
46:49
yeah let's let's do this so so um this is actually really really
46:54
interesting um I don't know if you heard of this idea in machine learning that came out only
47:00
in the last two years I think called double descent so it's basically some
47:05
people wrote papers on the fact that when you have a AI model or machine
47:10
learning model as you increase your your features or the number of your
47:17
features your uh your arrow will basically uh uh drop first of all but
47:25
then as the number of features increases the number of data points that you have so so the data size of the data set the
47:33
arrow will go up again so so so you will basically you know like like if if if
47:38
there's if there's too many features you know your model won't be as good you know you because because you're just
47:45
adding stuff that that isn't really relevant but what's interesting is if you keep
47:50
going and you actually increase the number of features Beyond uh the size of
47:56
your data set so actually like really big ones uh the arrow will actually drop
48:03
again and so that's that's a very fascinating thing and it's been it's
48:08
been also shown in finance recently that that that these large models with like a large number of Fairly nondescript um
48:16
input features can actually outperform uh smaller models and so there's a lot
48:23
of uh there's a lot of theoretical ideas behind why that could
48:29
be the case and so on and ultimately I think no one fully understands why this
48:35
is happening yet but for those of you who are uh into Ai and machine learning
48:42
it's extremely counterintuitive to any sort of conventional wisdom that I've
48:48
I've known and and most people but um yeah it seems like this is the new uh
48:54
this is the new interesting thing uh to to look at and to work on so I'm
49:00
actually doing a a discussion of a paper on that h on the quantopian platform on
49:05
the 26th of January I think so what what we're doing is we're actually just picking papers that are interesting and
49:12
then we have a a live discussion around those uh so I I think it's a yeah it's
49:19
pretty fascinating and uh definitely worth for anyone who's in that field uh
49:25
worth looking at that's great but speaking of great discussions um you know I really appreciate your time today
49:31
I think there's a ton of value definitely thanks for having me have a have a really lovely uh uh New Year and
49:38
uh we'll definitely stay in
49:54
touch
#Brokerages & Day Trading
#Commodities & Futures Trading
#Investing