cuckqueen

@_Dave__White_ second one was saying, ur setup is nice analytically since there's a clean solution to the symmetric random walk given boundaries. i was remarking w this bias setup it's a bit more hairy to intuit end result, but still tractable

@_Dave__White_ she can swap out optionally when she's losing for a cheaper value one, or if the idio risk of a scarf one spikes, etc. she has the ability to swap so i was saying, the bias could be related to how big the spread is in the set that alice defines (i.e. ocelots w wizard hats)

I've been working on a highly speculative new NFT primitive: Martingale shares, or "Mortys." Mortys represent fractional ownership of classes of NFTs. They do not require buyouts or oracles, but instead rely on a random Martingale settlement process. paradigm.xyz/2021/09/martin…

@_Dave__White_ The current characterization lends itself nicely given the hitting time of the process with the boundaries has a clean solution, but even with biasing the process the closed form isn't too bad.

@_Dave__White_ Nice! Given Alice has embedded optionality to swap out her Ocelot, seems like this can be characterized as a supermartingale for her, say where the drift of the process is proportional to the variance of the Ocelot set as a penalty term for max pain in Ocelot swap?

@lightspringfox folks also run a sim mirror to prod and feed in the divergence of the two back into a prod learner (not just stopping once exceeds threshold), since that itself would signify value (i.e. potential covariate shift, etc.)

@_Dave__White_ DM'd

@cuckqueeen i understand like 15% of this, what's the quickest way to get up to speed?

what are finance nerds who aren't in crypto excited about these days?

@tarunchitra Was going to originally mention this when @_Dave__White_ wrote his TWAMM post a few weeks ago but wasn't too relevant at the time albeit tangentially related

@tarunchitra e.g. if you're masking an exec algo (i.e. say twap) by switching between sampling from two gaussians and injecting it into trade {size, price, time}, an adversarial agent can game this by running hard EM, but right before they converge, you'd ideally toggle hyperparams of noise

⚠️ 📰💧🚨 Tired but true: Private Uniswap can’t exist Wired: *Differentially Private* Uniswap does exist This month’s paper (9th in a row!): can’t get the mechanism we want, but there exists one that we need eprint.iacr.org/2021/1101 (Storytime! 📓) x.com/0xbunnygirl/st…

@tarunchitra To be clear, this isn't predicated on MEV, right? i.e. you can have many adversarial agents that are training their own online learners with a (t_i, p_i) sequence and construct t_i+1_hat that they could monetize as soon as they see t_i without waiting for t_i+1 to hit mempool no?

Why online learning? It’s because of MEV, stupid! If an adversary sees a subset of trades + prices, say (T1, p1), .., (T5, p5), and that’s enough for them to predict T6 with high probability, that’s bad (front runs T6!) Notions of privacy for CFMMs *need* to represent this

@0xdoug Also why there are small dedicated teams within larger outfits working solely on approximating L3 books from L2 feeds for both sim'ing maker strats but also extracting features. ++ literature from the heavy diffusion / mean field theory / (obvs) queueing theory space as well

Knowing queue position is huge. When I was TradFi HFT market making (in a thick book market), queue position easily constituted 50% of the signal. There are entire firms living off quoting in the deep book, and selectively canceling at top of book until they get good queue.

@tarunchitra wen tower

End of an era — everyone needs to sip on retail flow to survive in 2021

@tomhschmidt That doesn't track for me— why does the ability to cycle in and out of liquidity on v3 pools imply matching with uninformed flow? Is the predicate that all DeFi AMM (i.e. v3) flow is uninformed, or that the ones who continue to trade on pools with this behavior are uninformed?

@cuckqueeen It's true -- it would break the way people traditionally use AMMs and would potentially change the flow -- but this is a bet that you'll still get a large amount of uninformed flow you can scoop up. Right that it would be great if this changes.

Free idea based on @mevintern / @hosseeb: a public Uniswap V3 vault that uses Flashbots to deposit and withdraw liquidity before and after big trades. No toxic flow, no IL. Just fees. Works even better at scale -- can socialize gas costs of the strategy.

@0xfoobar @Inuaraashi @tarunchitra Well, in the metric embedding of the NFT you'd incorporate in the human constructed features, but the exact community detection (if I may) of that joint space can be amenable to unsupervised techniques.

@0xfoobar @Inuaraashi tRanSFer LeArNiNg /s

@0xfoobar @Inuaraashi Need an unsupervised layer to partition the NFT space into learnable sets to feed into downstream SLERMs, and perhaps just ensemble them together by training a metalearner? Unsupervised layer is non-trivial, need to embed NFTs into a joint metric space amirite @tarunchitra

@0xfoobar @Inuaraashi Agree feature eng, per usual, is key. Did quite a bit of work on this a while ago and recurring problem was constructing a reasonable grouping of observations for training a supervised learner since individual NFT ecosystems were too sparse for stat sig results

@tomhschmidt @danrobinson @mevintern @hosseeb But, if you ask me, I'd rather use this as a way to source higher passive fill rates in a multi-legged strat rather than any short dated arb which I think constrains the possibilities of where this fits into an algo

@tomhschmidt @danrobinson @mevintern @hosseeb Without saying too much, if you know their team, a good portion of it is lead-lag (i.e. momentum across cefi -> defi) where they already have the first leg of linear delta that they can instaneously source a high prorata fill with cycling liquidity.