OneChronos, a proposed trading venue, wants to use AI to match orders
bloomberg.com
bloomberg.com
In short, we're applying combinatorial auctions to financial markets. Combinatorial auctions are generally speaking computationally intractable. Our art pertains to structuring the problem in a way that's a good fit for finance, while remaining tractable. The classical approach to solving combinatorial auctions for indivisible goods on the "hard" cases (ones without a totally unimodular constraint matrix) generally involves encoding the problem as an integer program, at which point you're faced with solving an NP-complete combinatorial optimization problem. Needless to say, that doesn't lend itself well to financial market speeds or scales. As such, we've taken a different approach based around guided search, similar to AlphaGo.
The Bloomberg article doesn't go into this, but the other half of the model involves inverting the traditional market event loop. As it stands, market participants receive market data and fills from geographically discrepant trading venues, action it via their models, and update their resting orders accordingly. We invert that model by putting points-of-presence equipped with highly precise network timestamping infrastructure at all of the major financial data centers. We use said infrastructure to timestamp snippets of code (our bidding language) that market participants send to us, as well as market data from other exchanges. At discrete, random intervals on the order of every ~100ms we resequence data in accordance with its origin timestamp (buffering by max time-of-flight) and feed it into our client's code for them to act upon and enter orders into our combinatorial auction. The net effect is an environment that's agnostic to geography, immune to latency arbitrage, and free of the traditional exchange race conditions. That's a huge win for everyone currently stuck in the prisoner's dilemma outcome that results from continuously reinvesting in what's now commoditized low latency trading technology.
Combinatorial auctions aren't a new concept, nor are Lamport clocks/distributed systems primitives based on timestamps, nor is the notion of using code as a proxy bidder (ad exchanges). Our art is around making combinatorial auctions practical for finance, both from an API standpoint, and from a computational tractability one. There's a huge impedance mismatch between what market participants want to do from an economic standpoint (trade on a portfolio level/express contingencies/force atomicity across actions) and what they can currently do with traditional limit order book/auction based venues (simulating the aforementioned by cancel/replacing as quickly as possible). Fixing this will be a big win, and broaden market participation.
Seems like you just need really fast optimization to collapse the problem space. A robust, shallow, transparent approach seems both safer and easier to sell.
I've found that customers are suspicious about "AI" positioning by startups. Plus, any arguably "AI" approach would be a bad fit to the problem (opaque, slow, no discernable advantage), why take the risk of putting the "AI" label on the project?
Sound so like they're doing some sort of approximation (using "AI"), which is the best you can hope for.
> What's with the "AI" positioning?
It's an implementation detail that’s abstracted away from end users, but also one of the enablers/answers to the "why now" question on a business level. This article happened to focus heavily on that, probably because the other aspects (distributed systems and mechanism design) aren't broadly relatable.
> Seems like you just need really fast optimization...A robust, shallow, transparent approach seems both safer and easier to sell.
With you on that. Academically/literature wise our approach is AI + ML given that it has priors, and doesn't rely on oblivious search techniques such as branch and bound or LP rounding for end-to-end solving. The article states that there's a "black box" involved, which is incorrect. When we talk to regulators and market participants we're very clear on how objective and deterministic the process is, and the bounds. We've structured the problem in a way that uses the economics of a double auction as the floor, meaning that any match we return is equivalent to or better than what would have resulted from matching against a conventional limit order book (CDA) or batch auction. That, combined with the timestamping aspect, results in something that's more fair, transparent, and auditable than the existing market structure.
From parent:
> There is no collapsing it
This is correct for truthful approximations and most formulations of combinatorial VCG. We don't fall into that camp, but we're still left with something that is worst-case exponential.
> ...sort of approximation
In many cases we can find the true optimal, and the cases where we can't become adversarial inputs for the AI and are still subject to the bounds described above. This has to do with the empirical runtime versus worst-case runtime of an algorithm, which is a fascinating field in its own right.
It seems that market incumbents wouldn't have a compelling reason to use OneChronos other than a sudden sense of altruism. Given that these incumbents make up much of the trading volume in existing exchanges, what will OneChronos do to garner their business?
Short term market impact results from supply and demand imbalance. As such, multi unit blocks are either an asset or a liability, depending on how you're positioned. Block venues are largely predicated on natural crosses, which are somewhat rare, given correlation across the street. Some market makers will quote block size for conditional orders, or against individual clients, but due to the limitations posed by double auction (mainly the inability to ensure atomicity between getting a hedge off and the outright), it's not done on a firm, multi dealer basis. Combinatorial auctions allow for multilateral block market making on a firm venue, which solves a ton of agency problems and improves liquidity.
Similarly, spreads on anything except for the most liquid financial products are generally superadditive when viewed from the taker's standpoint, given that you you'll need to either cross spreads or accept legging risk to execute.
The question most continuous limit order books face is "What is NBBO right now?" That question can't be answered thanks to the speed of light being finite. But because of the way our auctions allow time for orders and market data to arrive at the matching engine, we can more thoughtfully construct a consistent view of "What was NBBO 1 millisecond ago, at the moment we struck our auction?" This is a much easier question to answer, and also has the effect of smoothing out network jitter, data gaps, etc.
Also, the auctions are combinatorial but are computed across linear components (stock symbols). NBBO applies to symbols individually, so it can be taken into account as a standard constraint on each symbol.
I last looked about 2006 and I remember one problem with 11 participants on both sides requiring 2^38 steps to find a solution (make the set of bundles for the buyers that clear the auction at the best available price) within a not very strong bound from optimal. The details are vague as having sat down and done the sums I thought - well that's just not a goer is it - and dropped the issue like a hot spikey thing.
So - many challenges.
Isn't the premise of the company to use "AI" to find a solution? I'm going to go out on a limb and assume AI here means deep learning. I took both an AI and ML course during college where the FAQ was "what's the difference?". At least from the way the courses were presented, the AI class focused on traversing these exponential search spaces. ML on the other hand searched for local maxima / approximations. Obviously, you seem more familiar with combinatorial auctions. I'd love to understand where an ML approach fits in here.