• C++ Programming for Financial Engineering
    Highly recommended by thousands of MFE students. Covers essential C++ topics with applications to financial engineering. Learn more Join!
    Python for Finance with Intro to Data Science
    Gain practical understanding of Python to read, understand, and write professional Python code for your first day on the job. Learn more Join!
    An Intuition-Based Options Primer for FE
    Ideal for entry level positions interviews and graduate studies, specializing in options trading arbitrage and options valuation models. Learn more Join!

Deep Learning in Quant Trading

Hello,

I'm a deep learning specialist currently working at Tier 2 (maybe tier 1, but not FAANG) tech company. I think I have a pretty impressive background in deep learning and have created some state of the art algos in very competitive areas (such as text to speech). I also have experience building out great deep learning teams that lead to an acquihire. Also, I have a masters degree in pure math with a specialization in stochastic PDEs.

My specialty within deep learning is sequential models (such as TCNs, transformers, etc.) as well as graph neural networks. I should also add that I'm super motivated to learn about the field and like to devote myself to hard quantitative problems. I think that aligns better with quant culture as opposed to the 40hr/week tech culture.

What are my prospects of joining a quant trading firm to work on deep learning? I'd be interested in pioneering that work/team if it's not available at the firm, or focus on improving existing ML solutions by replacing them with deep learning, or taking a fundamentally different approach with deep learning (as is usually the case with graph neural networks).

I don't have any experience in quant trading, so I'd greatly appreciate any insight on this thinking or connections to the right people.

Thanks in advance
 
So, quick background: I work at a prop trading firm that has a HFT group and another group that works with Fixed Income (middle frequency/low frequency (muni bonds)); My work as a QR on a daily basis involves creating models using AI (Genetic algorithms, gradient boosting, adam optimization and so forth).

I would say that if you're going to a prop trading firm/quant trading firm with the intention of doing "pioneering" deep learning work, you're choosing the wrong industry. The primary goal of every prop trading firm/quant trading firm, even the ones that use AI, is to solve a purely finance problem (maximizing portfolio pnl while minimizing risk) using whatever the best tool for that is. Deep learning is definitely not the end-all-be-all solution for any of the problems we've faced. In fact, we're definitely one of the top three market makers for each of the products we trade, and AFAIK, we don't use deep learning at all.

Moreover, during my interview for my current firm (which admittedly is a very tech heavy firm with many ex-Googlers/Facebook employees/top MIT/Cornell grads), it was repeatedly emphasized to me that if I'm looking to always be using the latest in Machine Learning I'm coming to the wrong industry. At least with hedge funds/prop trading firms your main goal is to make money while minimizing risk while also making the most efficient use of your researchers. If creating a NN to price a bond, for example, will take an extra months work for maybe a 2% improvement in overall PnL (and increasing the risk of a weird price that results in a bad trade) then your boss will likely avoid it. That's just been my experience with "deep learning" in the trading industry, so others may have different opinions. I think it's fairly rare (outside of places like DE Shaw) to have roles dedicated to certain algorithms/tools and (this might be controversial) is probably indicative that the company doesn't actually know what they're talking about (source: I worked at Goldman)
 
Last edited:
Actually a brief anecdote that might help explain things better, but the most common algorithms I've seen at some of the more profitable places (goldman for credit trading, my firm, and a couple of others I have friends at) are linear regressions, genetic algorithms, logistic regressions, gradient boosting trees (very popular in lieu of NN at some shops), and kalman filters (I believe Goldman's credit trading team, which is the largest and I think most profitable market maker in Corporate Bonds/credit uses this extensively) ; Best quality a QR can have, and that I look for when interviewing candidates, is a wider toolkit so they're aware of what to use when, rather than specialization in a particular algorithm
 
So, quick background: I work at a prop trading firm that has a HFT group and another group that works with Fixed Income (middle frequency/low frequency (muni bonds)); My work as a QR on a daily basis involves creating models using AI (Genetic algorithms, gradient boosting, adam optimization and so forth).

I would say that if you're going to a prop trading firm/quant trading firm with the intention of doing "pioneering" deep learning work, you're choosing the wrong industry. The primary goal of every prop trading firm/quant trading firm, even the ones that use AI, is to solve a purely finance problem (maximizing portfolio pnl while minimizing risk) using whatever the best tool for that is. Deep learning is definitely not the end-all-be-all solution for any of the problems we've faced. In fact, we're definitely one of the top three market makers for each of the products we trade, and AFAIK, we don't use deep learning at all.

Moreover, during my interview for my current firm (which admittedly is a very tech heavy firm with many ex-Googlers/Facebook employees/top MIT/Cornell grads), it was repeatedly emphasized to me that if I'm looking to always be using the latest in Machine Learning I'm coming to the wrong industry. At least with hedge funds/prop trading firms your main goal is to make money while minimizing risk while also making the most efficient use of your researchers. If creating a NN to price a bond, for example, will take an extra months work for maybe a 2% improvement in overall PnL (and increasing the risk of a weird price that results in a bad trade) then your boss will likely avoid it. That's just been my experience with "deep learning" in the trading industry, so others may have different opinions. I think it's fairly rare (outside of places like DE Shaw) to have roles dedicated to certain algorithms/tools and (this might be controversial) is probably indicative that the company doesn't actually know what they're talking about (source: I worked at Goldman)
what do you mean by "is probably indicative that the company doesn't actually know what they're talking about"?
 
Meant to reply to this- agreed with binomial-torrent. At Goldman they hired a managing director for "Machine Learning", but the lack of proper communication between him and the various teams he could work with, and the lack of proper resources meant that whatever work he did was ultimately fairly slow, or ultimately not useful (building a neural net to price a bond when the fixed income team isn't going to use it or already has their own modeling they're happier with).
 

Daniel Duffy

C++ author, trainer
I
Meant to reply to this- agreed with binomial-torrent. At Goldman they hired a managing director for "Machine Learning", but the lack of proper communication between him and the various teams he could work with, and the lack of proper resources meant that whatever work he did was ultimately fairly slow, or ultimately not useful (building a neural net to price a bond when the fixed income team isn't going to use it or already has their own modeling they're happier with).
Before software goes into production, you need a prototype to show its feasibility. We are getting to the stage whereby the unfounded hype has died down a bit (e.g. NL 10,000 times faster than trad PDE.. ridiculous and irresponsible claims).
Example: Rough Heston ML is [7,17] times slower than traditional ANNs. Sobering Is ANN and variations the way of the future? do they always converge??



We should compare apples and apples

 
Last edited:
throwing money at the problem by hiring people with state of the art credentials for fancy roles thinking that that’ll automatically resolve it
That may be plausible if the roles of preparing data sources and designing the strategies, and the roles of fitting and tunning models are totally separated. The companies don't want their data scientists to know everything about the strategies, for example, what they are fitting and what kind of raw data are used.
 
That may be plausible if the roles of preparing data sources and designing the strategies, and the roles of fitting and tunning models are totally separated. The companies don't want their data scientists to know everything about the strategies, for example, what they are fitting and what kind of raw data are used.
And as a researcher, I think this type of environment is fairly unfulfilling. Lacking full context on what you’re doing and how it plays together with the rest of your strategies kinda neuters your work in my opinion. At that point why bother doing deep learning in finance if your job is just fitting models and tuning?
 
Top