Bridging the gap between academic MM and general alpha trader

[Lebowski]

The academic papers to me, the semi classical trained reader, solve “given a LOB, a current inventory, and a constraint, what are the optimal quotes?”



What I would like to discuss is how to approach the problem of getting smarter in practice. In practice you observe more than just the LOB and your inventory. There are correlated assets, if you’re quoting options maybe you want to observe underlying LOB imbalance, etc. How does an option market maker incorporate these adjustments programmatically without ending up with one incomprehensible “Uber strategy” big ball of mud? I strive for something more maintainable and testable—optimistically where PnL could be attributed to specific pieces of code, not just “we added a feature to the firms ouija board that allows us to pull our quotes in the calls when there’s selling in the underlying. It sits in the same class as Avellaneda‘s life work.”



Now let’s say you have all the above nailed but you have some view about vol or the underlying besides just “stay flat and back to back everything.” So now you must either:

1. Naively trade those strategies in separate books and perhaps even consume your own toxicity (EDIT: now that I think about it I think Self Matching Prevention on CME st least precludes this).

2. Find some way to combine them. Maybe you can tune some parameters that get you your desired exposure and pay less in t-costs



I know I’m as usual asking for the holy grail, but given all we talk about here nowadays is random industrial products from China maybe someone still interested in HFT will indulge me.



EDIT: and just to be clear I’m not looking for SOLID principles. I’m just looking for the right way to think about separating a generalized quoting policy (something from some phremch paper or if you’re like me just mimicking everyone else) and “everything else” (alpha signals, etc). I can figure out how to put the code in different classes, what I’m wrestling with is how to incorporate them into the generalized quoting policy in a way that’s clean and extendible.



Another EDIT: Maybe a first cut at it would be as simple as having a quoting policy and then have a linear combination of the other observables spit out an offset to the price/quantity suggested by the generalized model.





Thanks, all

[bridgeprod]

At Chengdu XinZhu Road and Bridge Corporation, we offer best in market making railway/bridge functional parts for large waterway gap.



At present, XinZhu's coating line has gained certification for 3 spraying-3 drying technology full automatic anti-corrosion coating that greatly reduces environment toxicity.



XinZhu's HZSG series steel mixing plant strive for maintainable straddle type monorail transit beam products. Visit our company site for a quote on our inventory of SOLID tension pot rubber bearing .

[hzkaiyue]

Do you need firm chemical fiber board and flat TPU sheet that is clean and extendable. we are the Hangzhou Kaiyue New Materials Co, Ltd located in Hangzhou north west, Zhejiang province, it established in 1989. Kaiyue mainly produces linear ball bearing in high temperature hot mud process.and develop specific pieces of fiber insole board which used mostly in book, shoes lining, cases, bags,furniture and Auto industry.



We also have got the certification of ISO9001 and OHSAS18001 industrial certification system.Kaiyue's theme is generalized no toxicity in our current inventary.

[nikol]

> the problem of getting smarter in practice

> I’m just looking for the right way to think



Are you asking about "how to develop intuition in MM/direction trading"?



--

Perhaps some structure is needed. You ask about:

- interrelation/interaction between LOB of underlying and LOBs of derivatives

- same but between correlated assets

right?



Or you have more items to cover?

[bullero]

What if you relax the assumption of zero inventory to be an arbitrary quantity. Then you run two models on top of each other. First layer determines appropriate possibly non-zero "allocation" to all assets and feeds them to the naive MM layer which just works as a quote engine trying to set the quotes such that the inventory drifts to and then around the target. Of course you can determine the whole model to incorporate information from multiple assets but its likely to be more difficult than what I describe here. What this almost-naive way enables you to do is to change and mix the models on the fly depending on how the assets you trade tends to trade. Essentially this enables you to stat arb while you make the spread, for example.



The picture below illustrates better the workflow. Here the S = price and Q_hat = target position. The AM stands for allocation model and AT for auto-trader. q_j is current inventory. Others are quite self explanatory.

[img]/User%20Files/12291/feedback_control_picture.PNG[/img]

[ronin]

> maybe you want to observe underlying LOB imbalance, etc.



There is a short answer, and the long answer.



The long answer:



Based on the notion that the LOB generates a short term price prediciton. So the underlying LOB generates a short term delta prediction, and the option LOB generates a short term option price prediction. Then you can back out the short term vol surface prediction. And that is something you can trade against your vol surface model, what ever it is.



The short answer is that that is all nice, but crossing the spread in options is expensive and the only intraday alpha there is comes from queue priority. Speed trumps smarts.



So you are left with the second part of your question, which is trading against your medium to long term vol model. And then you are effectively asking what is the best way to execute a basket of options.



So yes, the two strategies are effectively separate. You can let the market making strategy pass any unwanted inventory to the positioning strategy at mid, free of exchange fees. If that sort of thing matters for your positioning strategy, great. And if the market making strategy knows it can get out on one side cheaper than the other, that can bias its quoting to be more aggressive on that side.



But you are effectively running two strategies plus an internalizer.

[Jurassic]

Is Avellanda's work the standard way to create a market making strategy? I always thought it looked quite impractical.

[bullero]

With Avellaneda's work you are referring to stochastic control?



Edit: Btw, why do you think its not practical?

[jiangsuwarrior]

Avellaneda's is 2 generations behind .D*** build the complete way of doing it at *S,then called O****** then upgrade this by come up with a faster local approx to the solution of the HJB PDEs, at ***, Then he went to T** where the same idea is applicable towards capacity optimization problem. Still trash.



We, Jiangsu Zhaojia Materials Technology Co., Ltd has a better model than T**. In 1999,our company was built specializing in the production of melamine super plasticizer and dispersible emulsoid powder. Our founders from Peking and Tsinghua met when they were graduate school student under Shing-Tung Yau.



oday our team of 35 quants make markets in over 80markets. We are 1st biggest in almost all Europe and Asia equity, 7th biggest in US equity, 3rd biggest in CME and ICE, 4th biggest in Eurex, and last month we launch SPOTFX take away Citi market share they drop 1st to 4th.We also havae 375,000 tons annual production capacity of melamine.



We are hiring ASIC engineer and enigneer good with Xilinx device. contact me sww76jsieccn

[Jurassic]

With Avellaneda's work you are referring to stochastic control? Yes



Edit: Btw, why do you think its not practical? I literally cannot remember, probably to do with strong and unrealistic assumptions