> Crypto is flooded with liquidity so market making is like selling ice to an eskimo.

This is not true at all, in general quoting in these markets is absurdly capital-intensive compared to tradfi.

> These days a naive market making strategy in crypto just incinerates capital very reliably as the tiny bid ask spread is a small fraction of the adverse selection risk (whole spread moving past). They did probably make money on this in the early days and got smoked when the sophisticated tradfi players joined.

Are the sophisticated tradfi players here yet? Seems like no?

Of course sophisticated tradfi players are in crypto and have been for years making hand over fist in money.

Source: I work at a firm that does this.

Great! I'm just continually surprised, because my firm is 4 people with a combined 1.5 years of very outdated tradfi experience and these markets are very good to us.

For example, it's surprising that big tradfi players were not able to prevent take-only basis arb bots written in Python running on a un-tuned VPS from printing five figures on individual new listings in mid 2021, or permitted us to click trade tokenized stock quarterly futures minutes from expiry for nearly guaranteed profit also in mid 2021.

How is information on order books non public? Usually crypto order books are public APIs.

Some examples of hidden information that doesn't generally get disseminated in exchange orderbook APIs:

* Attribution: who's making the order?

* Short labelling... are they selling or shorting?

* Non-display or iceberg orders (not common in crypto?)

* Immediate-or-cancel orders... the executions hit the feed, but not the original order details. Also whiffs (order but no fill) don't get disseminated in any way.

* Certain order types that may rest on the exchanges order book but either don't have a specific price or display doesn't make sense... market orders, midpoint orders, pegged orders, auction order books (less common in crypto)

EDIT: On the attribution side -- they could also know the leverage any customer is taking and use that adversarially (which was the straw that broke their camel).

Thanks, yeah attribution is an interesting one. You usually have the cryptographic address of the maker but not any link to FTX account that might be responsible, and as you say the ability to correlate an order with other positions

> Certain order types that may rest on the exchanges order book but either don't have a specific price or display doesn't make sense.

I suppose orders than are designed to only be consumed by a matching engine don't need to be made public unless they are matched.

The order book from the API is at the very minimum delayed (by network and protocol latency if nothing else) and aggregated/sampled (full feed is too big), so an insider can have an advantage of more complete and timely data.

That's for honestly run APIs, then an exchange can play some games with that feed if they want to...

It depends on the time, it's public after the fact, but not-public during the time gap where someone like Alameda could do front-running by managing to get in other transactions before the loser's orders get executed, or before any others get the chance to take that order.

Alameda was in general deeply incompetent. Front running your customers is not all that doable if you insist on using python.

You don’t need speed to front run someone you need special access, which Alameda had.

That said they are being accused of something much less sophisticated. They were allowed to take money out when they made money but didn’t have to pay money in when they lost.

Even if you have special access, if you need 40ms (500ms p999) to compute a square root, you're not gonna be able to make worthwhile decisions on the basis of this access

Those times are wildly inaccurate on modern hardware with modern python, fwiw.

These are our observed latencies for their risk checks after sending millions of orders per day for months at a time (which doesn't sound like much, and isn't much, but is the entire order rate that they gave us while we were doing 0.5% of the maker volume on their exchange).

So it’s not, as you previously stated, the time to compute a square root?

That seemed to be the most computationally demanding portion of their public documentation of the formulas involved

You’re stating confidently that python is orders of magnitude slower than it actually is because of some pseudo code, which you haven’t seen implemented, and you called into across a network link?

…do you see the problem?