Beating the Spread

Malt Liquidity 104

Anyone who has traded less liquid markets than ES/SPY understands a certain pattern when trying to make a trade:

The spread is wide, let’s say .35 x 10 — .6 x 5. You place an order slightly better than the bid, but nothing doing. .4 x 1 — .6 x 5. All of a sudden, the level of your bid is populated with 20 contracts, and you realize you’ll have to up your bid to try and get a fill on this spread that clearly nobody wants to handle. This is worth experimenting with to truly grasp how algos “follow” real orders — as I’ve written before,

And in fact, one of the earliest option strategies I came up with fell along these lines. I’d go look for relatively wide spreads that were seeing some volume, and try and walk the book in on either side, while avoiding the midpoint. If I got a fill, that’s a bit of edge in and of itself, and I’d try and work the fill alongside price movement across the book to get out. It was LFT — shitty, unhedged low frequency trading.

Visible, automated quotes are just the most advanced version of a concept we’re all aware of, called “dynamic” or “surge” pricing. Really think about what Uber’s supply-demand sensitivity actually is — it’s just market making cab rides rather than options trades. A certain route has a given bid (whenever the user puts in their desired location for a price quote) and an ask (what price is offered to the driver), with Uber collecting the spread in between. Imagine this process in a vacuum, where there is only one rider and one driver: internally, Uber has to have a relative pricing mechanism for that route, but it will be quite wide because of uncertainty. As more riders populate the area with “real” bids for rides, rather than relying on the pricing mechanism directly, the relative pricing mechanism (e.g. the “market making algo”) can congregate bids and asks to create a “real demand sensitive” price. This is why rapid quoting of a single ride ups the price so quickly — it’s not gouging, but when you quote a ride from KGB bar on 4th to Martiny’s twice, you’ve created 2 real bids for a route that previously had zero. The spread has become more expensive due to your own bid (say, $25 x 0 to the rider, $20 x 0 to the driver —> $29 x 2 to the rider, $20 x 0 to the driver) — what’s the probability of having two rides being demanded from the exact same stop, outside of an airport? You’re competing with your own bid in a remarkably thin spread. Needless to say, the proper Uber strategy is to avoid refreshing the app instance when ordering a ride.

We can quickly observe that pretty much every relative pricing mechanism follows this model, whether it’s Amazon pricing or Facebook ad auctions (in fact, I’ve heard that their ad model was constructed by essentially implementing a trading MM strat.) However, there’s a particular theory that this type of pricing, when implemented poorly, is a cause of inflation. The theory essentially works like this:

Take our Uber example from earlier. In the right locality, theoretically, you can compare against yellow cabs and Lyft. As humans, we can sanity-check these prices pretty quickly — a $40 Uber vs a $25 cab ride is clearly indicative of some sort of algorithmic oversensitivity. However, what if the prices across providers are monitored algorithmically? The algo doesn’t know how to populate the spread by any mechanism other than looking at prices. So if Lyft sees a certain route being priced by Uber at $40, while it is offering $28, it might rapidly adjust upwards, creating a feedback loop of sorts. Or, as described in Amazon pricing structure, here’s a funny scenario:

In 2011 an Amazon seller called “profnath” listed a second-hand, out-of-print academic textbook (The Making of a Fly, by the geneticist Peter Lawrence) for $9.98. And there it would have stayed, had another seller not listed the same book using automatic pricing, which caused the original listing (also set with automatic pricing) to rise in turn. Each seller just wanted to make a slightly higher margin than the market rate, so the two listings kept bidding each other upwards until the book cost well over 23 million dollars.

The flaw here is one of hidden book depth. Something that’s overlooked about market makers is that they’re quite literally plugged in to the market — designated MMs have deeper access to incoming order flow and order book population than basically any other party. This is why they’re obligated to provide quotes on both sides, but this is also why when prices actually move, the algo prices are never the ones getting hit on the way up or down. Only the resting orders get hit. But in the Amazon example, it’s unknowable whether the other seller listing the book has gotten a real offer or is simply adjusting relative to the market. Previous sales are tracked on the Amazon site itself, but in the case of a rare book, how does the offeror come up with the ask? Perhaps they have sold a copy in another venue (eBay, brick and mortar stores, etc.) and are pricing it as such… or they’re just winging it. In a situation where “real orders” are indistinguishable, these kind of feedback loops proliferate — all liquidity available is not created equally.

Whether this is inflationary, though, is another question entirely. Certainly, this impacts prices locally — one need only go to an intersection with multiple gas stations to see what type of price coordination overly-visible price quotes results in — but it’s hard to take the article’s argument seriously that this could create some sort of measurable, large-scale inflationary effect:

Is this inflationary? The idea that prices across the economy are made “supra-competitive” by tacit collusion certainly sounds inflationary. The Competition and Markets Authority published a working paper in 2018 warning that pricing algorithms could “reduce competition and harm consumers”, and it is notable that the areas in which so-called “greedflation” is most identifiable – such as fuel – are those in which automatic pricing is most widely used.

Presumably the intuition here is that if gas station prices are incorrectly (or abusively) relatively priced locally, the national fuel price calculation that works into consumer inflation metrics would also be inflated. At the very least, this separation in quoting mechanisms should intuitively explain why oil prices at the pump fall much slower than spot. But this ignores that the process of pricing oil is mostly an effect of cartel theory and a prisoners dilemma — OPEC sets the production caps, but

The inability to cooperate between cartel members is nothing new if you have ever followed OPEC. Countries might agree to curb production to increase the price per barrel, but surely enough, a country might sneakily increase their production because the increased profit is too tantalizing. And of course, at the next EIA report, this behavior shows itself and the price of oil tumbles.

In practical terms, though, how do we avoid having to cross the spread in every interaction with an algorithmic model? Well, think about what I talked about earlier with “below-screen liquidity”:

Really think about what a midpoint fill is - when you are being quoted the bid and ask, why would someone fill you at the midpoint? Overwhelmingly, you’re not transacting with someone else trying to sell an option at the midpoint - you’re interacting with a market maker filling the order. As such, there is a reason why the MM might want to come in on the spread, which is in your favor - you do not have to cross fully to the bid or the ask to transact.

Automated market making, whether it’s Uber rides or options, preys on our desire for convenience. Booking a cab requires much more forethought than opening an app, but in a sense, we have to stop providing “real orders” because every order you place conveys information. It’s why I’m against the concept of stop losses — resting orders intuitively are negative EV because that’s the order that gets hit when liquidity is pulled. A human can probably adjust faster when bailing out on a trade than pulling a quote in a book where liquidity has dropped off. The only workaround, however, is to put in the work. It’s why I can find endless articles on the median rent for an apartment in NY, but know precisely one person that pays that amount or more (who happens to be a price-insensitive consumer, the true American dream.) Call the landlords directly, offer the teenager next door a flat fee to take you to the airport, and make the effort to check the in-store price yourself. You might be surprised at what you find.