 nikol
|
|
Total Posts: 1273 |
Joined: Jun 2005 |
|
|
@doomanx
Thanks for confirmation of the direction. I do not claim to be the "inventor" of this stuff. Just "educated reader" with encyclopedic views )) Totally agree with DRL / ANN approach (almost London-"DLR" which b.t.w. runs without drivers for a long time).
@bullero
What I have written is far from mathematics. Consider it just a little "pepper and bay leaf" to give a flavor.
There is also another approach, where you can create models describing certain intermediary functions (like filling probability) and then plug it into your Bellman-eq as "expectation" which is deterministic because stochasticity is convoluted within E[.], so the dimensionality reduces.
Something like this. |
... What is a man
If his chief good and market of his time
Be but to sleep and feed? (c) |
|
|
 |
 bullero
|
|
Total Posts: 75 |
Joined: Feb 2018 |
|
|
@nikol, I think you missed my point. What I was suggesting is that approximation is _the_ way to go in practice. It even has a new sexy name: Deep Galerkin Method. You solve the trading model using neural net at a set of random points. |
|
|
 |
 nikol
|
|
Total Posts: 1273 |
Joined: Jun 2005 |
|
|
Indeed, I missed it. Thank you for reminding about Galerkin. "Deep Galerkin" is quite impressive.
PS. Somehow, Hartree-Fock method popped up: https://en.wikipedia.org/wiki/Hartree%E2%80%93Fock_method
|
... What is a man
If his chief good and market of his time
Be but to sleep and feed? (c) |
|
|
 |
 gaj
|
|
Total Posts: 116 |
Joined: Apr 2018 |
|
|
I could be wrong, my experience is nowhere near everyone else in this thread. But I've worked with teams with large %ADV in major markets, and I've never seen any of the fancy stuff mentioned in this thread being used in production. The correlation between mathematical sophistication and trading success seems to be quite low. In many cases it may even be negative. |
|
|
 |
.jpg) doomanx
|
|
Total Posts: 94 |
Joined: Jul 2018 |
|
|
@gaj I really think it depends what kind of game you're playing. If you're virtu-esque where you're just back-to-backing everything at insanely high frequency holding no risk sure, you don't need to model a whole lot. If you're actively working an inventory the problem becomes a bit more complicated. Not to say you need HJB or RL or any of these things, but they are tools that have use cases. |
did you use VWAP or triple-reinforced GAN execution?
|
|
|
 |
 rickyvic
|
|
Total Posts: 222 |
Joined: Jul 2013 |
|
|
Relatively off topic.... Do you guys know or have references on the optimal message rate. It seems to me there is a relationship between a high cancel replace rate, latency and expected return/risk horizon. Any high level ideas appreciated too. |
"amicus Plato sed magis amica Veritas" |
|
 |