Launch HN: OneChronos (YC S16) – Combinatorial auctions market for US equities

That it is!

> What is the one sentence describing who would use this and why though?

Today's market structure costs institutional investors (and by extension households) at least a trillion dollars annually (and Smart Markets hold the potential to eliminate that loss).

We very much agree that the potential is even more significant for markets where opacity and price discovery is less efficient than US equities. We chose US equities as a beachhead given the mature regulatory framework and the extent of market fragmentation. Other asset classes and geographies are immediate next steps.

The optimization procedure (which includes bid evaluation) is ~30ms. We cycle bound (under a formal model of computation via function application and graph reduction) computation of bidders to ensure that everyone shares an identical and deterministic resource cap.

> Do you use a database of some sort?

Not as part of the real-time trading system, which operates as a CP fail-stop distributed system model checked for safety and liveness by TLA+ and system tested by Jepsen.

> How do you to handle settlement?

Regular way (T+2 settlement with a 3rd party clearing BD)

> How do you handle ingest?

We use a constellation of GPS synchronized Stratum 1 clocks and proprietary network timestamping software + hardware to ensure that we process orders entered by the auction call time regardless of what physical host we receive the order on. We do the same for market data broadcast from other trading venues across data centers and geographies. We stream both market data and orders to a central point for processing. Every node in our distributed system that processes orders or “away venue” market data broadcasts a “Gateway Call Announcement (GCA)” message at auction call time to downstream compute nodes that run the auction. Auction solver nodes get to work after receiving GCA messages from the hosts they expect to hear from.

Say you want to invest in the health + tech space, but there's some risk that covid ends and gyms come roaring back. So then you want to minimize the risk by putting some money in the gym industry as well -- getting an entire package of peloton, apple, and 24hourfitness stock is actually worth more to you than the individual stocks on their own.

Substitutes in capital markets are ubiquitous. There might be hundreds of candidate hedges in the example above, but given how trading workflows are, there's no way to communicate that amongst market participants. A market maker has no way of knowing if someone wants to buy SNAP (and thus potentially has market-moving information about it) or if they're using it as a hedge for a short position and would substitute something that the market maker wants to gross down on (and offer a more aggressive price because of that). As such, market making is a game of pricing under risk and uncertainty. Combinatorial auctions eliminate much of the uncertainty.

Initially we're US equities only. Stay tuned for other asset classes and geographies. Spot vs derivatives is a core use case that we want to do as soon as we can (only national exchanges can do listed derivatives trades, so it's a big lift).

> there's no strict long-running orderbook here, right

The default good-till behavior is one auction cycle (100ms Poisson random back-to-back).

> Are you using FIX for your protocol

Yes! And, we have a formal model of our FIX spec and self cert flow that makes onboarding us way easier of a process than what's typical [1]

> Are you using FIX for your protocol and where are your servers geographically located?

Initially, just Equinix NY5. Longer term, we plan on PoPing at most financial data centers. We use a constellation of GPS synchronized Stratum 1 clocks and proprietary network timestamping software + hardware to ensure that we process orders entered by the auction call time regardless of what physical host we receive the order on. We do the same for market data broadcast from other trading venues across data centers and geographies. We stream both market data and orders to a central point for processing. Every node in our distributed system that processes orders or “away venue” market data broadcasts a “Gateway Call Announcement (GCA)” message at auction call time to downstream compute nodes that run the auction.

[1]: https://www.onechronos.com/docs/fix/fix-42/

Yes. Some brokers are incorporating us into their algo suite, others are offering us as a direct route, and most are doing both.

> Do you work with some big broker-dealers on integration?

Yes! We're excited to be launching with many of the household names, and most have plans to connect by early H2. We'll be updating our website with a list of launch partners in the coming weeks as part of our full launch announcement (the HN fam is hearing it first).

> I know you started out with equities, but bond portfolio transitions are a much bigger pain - any plans there? Or issuance, i.e., mix of funding instruments in one go?

Getting to this world state is our real passion. Imagine a fund manager running a cross-geography equities and a credit book. Any trade they want to do will involve rates and currency risk on top of the actual delta. We want to make it easy to, say, sell some European debt issuances in euros to fund a US equities position in dollars while re-hedging curve risk, all as part of one atomic and frictionless transaction with a pre-trade known cost basis.

You can read our ATS-N here: https://www.sec.gov/Archives/edgar/data/1692652/000169265220...

[0]: https://otctransparency.finra.org/otctransparency/AtsData

The harder problem that large traders face is executing blocks and portfolio trades. How do you figure out what your total transaction cost (market impact / cost of liquidity) will be if you're buying 100x the displayed volume? Being able to express where you are flexible (e.g. individual security prices) and aren't flexible (aggregate price, atomicity) helps lock in the uncertainty pre-trade.

So we're actually mostly going after the large scale stuff, more than the small scale.

We raised a series A in 2019 led by Green Visor (who has been excellent btw, and with us from the start)

> it seems like you've been building for ~5 years ... i'm guessing it cost quite a bit to develop

Yep, you're spot on that it's a complex product. The biggest cost has been making it feel for the user like it's not. What that boils down to is an enormous amount of iterative feedback and development w/ the industry. Between that and the regulatory process, a lot of the "cost" has been more duration than cash burn. We've kept things lean from the start in anticipation of that.

> unless you built out all of the components yourself

We've developed the tech in house, with some hands on help from our friends at Imandra mentioned in the OP. On the research piece: that's been happening in the background for many years, and we're definitely building on the shoulders of giants in the worlds of mechanism design, algorithmic game theory, and deep learning. We're lucky to have some great academic advisors involved (like Kevin Leyton-Brown since the early days) as well.

On substitutability: if you want to sell $2m of some sector basket you wouldn't put in a limit order for $2m in every security (overfill risk). An expressive order can enforce a $2m global constraint across the basket but show full size in each security. When lots of people are doing this, it solves the "ships passing the night" problem where people are looking for offsetting exposure at a high level but can't express anything but single stock orders.

On risk management: some constraints certainly are restrictive, e.g. conjunctive constraints like 'a' AND 'b' AND 'c'. It would be very unlikely that we find the exact opposite of that constraint, so the auction is multilateral: we can stitch together the contra with individual single orders for 'a', 'b', 'c'. Key to the liquidity aspect of this is our objective function: it rewards more aggressive pricing and larger quantities. So the principle here is that by gaining atomicity (and reducing uncertainty) people can be more aggressive on price and qty. This is especially important for liquidity provision: how much larger size could market makers quote if they could automatically hedge new positions they enter into?

All that said, bootstrapping liquidity is the hardest part of any venue launch. We're obsessively focused on making sure we have the right blend of participants trading on different horizons for a healthy pool.

https://www.cs.cmu.edu/~sandholm/vickrey.IJEC.pdf

but in the context of second price auctions.

lpage might be alluding to something having to do with their proxy bidder implementation but the above paper actually discusses how proxy bidders themselves lead to untruthful bidding (so maybe lpage is suggesting their implementation is better?).

Both the multiunit dynamics and the specifics of our uniform clearing price mechanic minimize ex-post regret. Double auctions suffer from the winners curse/adverse selection, as limit orders are always "traded through." Multiunit uniform clearing price mechanisms like OneChronos can lessen or eliminate that by incentivizing buyers and sellers to truthfully report aggregate supply and demand curves, and Expressive Bidding enables the reporting of supply and demand curves (among other things). NB: we are not an IC direct mechanism. We are balanced budget and individually rational.

I love the part about eventually determining your value-add by comparing to a counterfactual vanilla market -- sounds a bit like Shapley value? If not exactly Shapley value?

It's a hot take on both Shapley values and VCG (while avoiding the issues with both), and it's about to become an active area of research for us!

Our value add is mainly a team that understands these fields and the extreme nuance of capital markets. That's allowed us to generate novel IP and a purpose-built solution.

[1] https://www.nobelprize.org/prizes/economic-sciences/2020/sum...

[2] https://arxiv.org/abs/1706.03304

Information theory tells us that no universal bidding language (there's a representation of any package of interest) is uniformly more compact than the power set representation. Nonetheless, a good bidding language makes "common" bids compact and easy to communicate. We thought about this problem deeply and realized that functionally pure computer programs mapping proposals (packages of goods) to valuations (how much the bidder will pay or would want to receive) are about as natural as it gets. There's a direct analog in asking a human or a pricing algo for a price in a bilateral trade setting. However, our optimizer doesn't know what to do with an arbitrary computer algorithm, and exhaustively querying one to get the power set of prices out is computationally infeasible. However, using formal methods, we can (in the right setting) convert a computer program into an equivalent representation in a logic fragment called mixed integer real arithmetic. And that (via SMT solving) is something that an optimizer can work with.

You can see what Proxy Bidders (the pure functions that create expressive bids) look like here [1].

[1]: https://www.onechronos.com/docs/expressive/bidding-guide/#in...

Folks also plan on using Expressive Bidding to enter other business lines that high startup costs or low margin (ex ongoing technology costs) previously kept them out of.

Thanks! And yep, similar but all the way down to the venue/match level, e.g. as opposed to a broker taking on some legging risk to shield the end investor.

> How do you think market makers are going to react to this? Expanding on Kelly's take - the big thing it does for market makers is allow them to manage momentary risk. When a market maker gets filled on an exchange, they are immediately looking to hedge/offload what they took on which involves a sequence of transactions.

Here, the hedge is baked in. So for example, they may enter an order that looks like "buy and/or sell any mix of these 200 securities, if and only if the net change in risk (e.g. change in exposure across several factors) is within some tolerable distance from 0". So that would look like a traditional bid-ask spread across a series of symbols, but with a global exposure constraint. The key outcome being they can quote larger sizes across symbols safely.

NB: the MM doesn't need to know anything about the composition of the complex order on the other side. On top of that, they may be filling one leg, and a natural or other LP filling another etc..

1. There are two places where deep learning and prior-based approaches can come into play for combinatorial auctions. One is pretty analogous to AlphaZero, but substitute placing a piece on a Go board with accepting a bid, hoping that upon reaching a terminal state the set of bids accepted is feasible and close to optimal. The second is perhaps more in line with what you mentioned—using ML for hyperparameter selection in an algorithm portfolio. When we go live and have production data, our meta optimizer will measure how different approaches are doing and allocate computational resources accordingly in an online fashion. We always use a vanilla unit double auction as a baseline to measure relative performance within an auction cycle, and if the baseline is better, we use it instead.

2. There's a fun and serendipitous story here. I wrote an extremely early prototype as a tiny lisp and evaluator to go with it. We needed a very restricted and functionally pure language that we could control the execution context of, symbolically execute, and do basic formal methods on. The approach worked for a POC, but it was a far cry from real-world adoptable. We proceeded to prototype a DSL with an HM inspired type system and a more pythonic syntax, arriving at a poor man's ML. Better, but a DSL, and something limited/bespoke that would ultimately be annoying for developers. Then we met the guys at Imandra [1], who convinced us that we could have our cake and eat it too using vanilla OCaml/ReasonML and an ultra-high level theorem prover to keep code in an acceptable logic fragment. As an aside, rust is our systems PL and where we do most of the heavy lifting. Evaluating Expressive Bids isn't computationally expensive relative to the optimization problem.

[1]: https://www.imandra.ai/