Maggette
|
|
| Total Posts: 1350 |
| Joined: Jun 2007 |
| |
|
Him
no idea how I ended up there while strolling through the internet during lunch break.
https://www.xtxmarkets.com/career/quantitative-researcher/
Under "Hiring"
"Our interview process seeks to gain signal in the following topics:
Modern machine learning techniques, especially deep learning"
Is that fishing for young researchers? Or is there an actual chance they are using DL?
I mean they could. But.... well you know...unlikely. |
Ich kam hierher und sah dich und deine Leute lächeln,
und sagte mir: Maggette, scheiss auf den small talk,
lass lieber deine Fäuste sprechen...
|
|
|
 |
ronin
|
|
| Total Posts: 708 |
| Joined: May 2006 |
| |
|
Didn't the deep hedging guy move to XTX? If he did, that's your explanation.
Maybe they think it's not a gimmick. Or maybe they know it's a gimmick, but they want to run with it. Who knows.
|
"There is a SIX am?" -- Arthur |
|
 |
ETwode
|
|
| Total Posts: 7 |
| Joined: May 2017 |
| |
|
"a research cluster of ... six thousand A/V100 GPUs (and growing fast)"
Could certainly be marketing but if ^ is the case I imagine they're serious? Similar scale as Tesla (https://blogs.nvidia.com/blog/2021/06/22/tesla-av-training-supercomputer-nvidia-a100-gpus/) |
|
|
|
|
sloppy
|
|
| Total Posts: 13 |
| Joined: May 2011 |
| |
|
| @Magette: Why do you think it's unlikely? |
|
|
|
prikolno
|
|
| Total Posts: 97 |
| Joined: Jul 2018 |
| |
|
Not a gimmick. I've seen DL outperform other models like gradient boosted trees, RFs, decision jungles etc. by a substantial margin. I've only done this in environments with sub-15 mic signal computation paths, but I have former colleagues with similar success in low-frequency environments with textual features and 4+ years live trading track record.
I prefer developing better monetization and researcher productivity tools than signals, though. I feel you get better bang for your buck than eking out model improvements for conventional asset classes these days. |
|
|
|
|
sloppy
|
|
| Total Posts: 13 |
| Joined: May 2011 |
| |
|
@prikolno:
you're saying that you grow P&L more by "monetization and researcher productivity tools" than by improving the model or signals? What exactly are these tools doing if they're not improving the model? |
|
|
|
ronin
|
|
| Total Posts: 708 |
| Joined: May 2006 |
| |
|
> Not a gimmick.
Is that the Yuri paper? |
"There is a SIX am?" -- Arthur |
|
|
|
prikolno
|
|
| Total Posts: 97 |
| Joined: Jul 2018 |
| |
|
@ronin: ?
@sloppy: So many things? For instance:
- Create more features or meta-features.
- Create more monetization rules or strategies.
- Improve simulation.
- Trade more tickers.
- Trade more markets.
- Reduce probability of software crashes in production. => More time to do other things.
- Simplify debugging of strategy or signal implementations. => More time to do other things. Less corner cases where signal values are updated incorrectly.
- Simplify configurations or meta-configurations. => Less operator error during production.
- Faster roll-outs.
- Faster re-parameterizations.
- Faster optimization of execution trajectories.
- Lesser "downtime" between market sessions or during corner cases like partial trading days or around halts.
- (Cleaner) interfaces to access venue-specific fields or functions, e.g IOC indicators, user-defined instruments, MPIDs, RFQs. |
|
|
|
Maggette
|
|
| Total Posts: 1350 |
| Joined: Jun 2007 |
| |
|
@sloppy maybe I shouldn't have an opinion on that, since I actually have no experience in market making. I do have some experience in strategies and apply ML for a living (outside of finance and in Energy Trading) for almost 10 years though.
In my experience features in the financial world are rather shallow. You wouldn't need a very deep net to learn these representation.
As an wild oversimplification: IMHO ANNs shine when you want to learn a very complex function from lots of data with an ok signal to noise ratio.
IMHO finance is about learning a simple function from a noisy data set that probably isn't that big (relatively speaking, since markets are instationary it is not entirely clear what you would do with years of (maybe outdated) order book data).
Just my 2c. Take them with a grain of salt. Prikolno is probably practical experience from the market making world and I would defer to that.
|
Ich kam hierher und sah dich und deine Leute lächeln,
und sagte mir: Maggette, scheiss auf den small talk,
lass lieber deine Fäuste sprechen...
|
|
|
|
sloppy
|
|
| Total Posts: 13 |
| Joined: May 2011 |
| |
|
@prikolno:
Thanks, much of that makes sense to me. I guess you mean different things by "feature" and "signal", whereas I took them to be the same in this context: the X's that get pumped into your ML black box.
"Execution trajectory" = how you slice up a trade decision into orders? kind of like a vwap?
"Monetization rule" = rule for turning ML prediction into a trade?
|
|
|
|
sloppy
|
|
| Total Posts: 13 |
| Joined: May 2011 |
| |
|
@Maggette:
Got it. I thought maybe you had some experience that convinced you that DL is nonsense for trading. But I agree with your general assessment: financial data is intrinsically noisy and non-stationary. Those characteristics seem out of step with the DL success story cases. |
|
|
|
|
prikolno
|
|
| Total Posts: 97 |
| Joined: Jul 2018 |
| |
|
@sloppy
I think most folks distinguish "feature", "monetization", "rule" and "signal" in the specific way I do. I think it's mostly because of the C* lineage in many Chicago firms. There's nothing wrong with drawing the boundary anywhere and calling it whatever your team likes, though. :)
Somewhat. A few large quant HFs and execution brokers have a similar problem because they're moving a lot of size, and the standard procedure is to solve a massive optimization and output the whole trajectory (in time dimension) of trade decisions across every symbol in the portfolio. This not only has to jointly take into account multiple model outputs, strategy/portfolio-level constraints, but also the different business constraints across multiple funds and clients, and also "experimentation costs". Even state of the art solvers with the best infrastructure take seconds to run.
Rules can be something simple like, "always cancel an order if it is at the top of the book alone", "always cancel the residual of a marketable order that posts". I specifically qualified "monetization rule" so as to emphasize the point on rules that aren't conditioned on the model(s) like these, since you asked, but you can also have model-driven rules like "always cancel orders if abs(signal) is larger than ...". |
|
|
|
