Understanding Jane Street

1st August 2022

Note: Today's post is pretty long. Consider reading it on the site, since your email client may truncate it.

The highest-stakes gambling events in the world are typically very discreet, invite-only affairs. One that might be close to the top in terms of available winnings happens at the end of Jane Street internships: interns get a stack of 100 poker chips and spend half a day getting asked brainteasers and then betting on their confidence in the answers. Some of these questions might be pure math and probability questions, some might be more abstract bets on making a market in some outcome, and apparently one of the questions is a tough probability question where part of the prompt is to bet on how long it will take to get the answer.1

This would actually be a pretty fun way for a math-minded person to spend a few hours, if it weren't so high-stakes: the winners get a job from which people routinely retire rich in their 30s, and the losers... don't. Jane Street is a quantitative prop-trading firm, i.e. they use their internal capital to make markets in liquid assets around the world, balancing completely automated strategies with human intervention. In terms of expected value, this might be the single best field for a smart and technical person to work in. Starting compensation can be up to $425k. (Don't feel too bad for interns who don't get an offer, either; they're apparently getting $16.5k/month.)

The point of the poker chip challenge is not just to test out people's probability chops in a thematic way. It's to measure what's arguably the most important bundle of traits for a trader to have: a sense of when to risk what, and enough self-awareness to calibrate confidence. It's also a way to filter out one particular form of overconfidence: no trade is good enough that it's worth the risk of losing everything. For what shall it profit a man, if he shall make a bet with unbounded positive expected value, and lose his entire bankroll? The history of finance in general and quantitative finance in particular is littered with examples of very smart people who had good track records and then blew up. The brilliant people are table stakes. Building a lasting institution means embedding risk awareness into everything.

Jane Street is not the only big prop trading firm, and not the only firm that thinks this way and tries to hire from the same pool. They're worth focusing on for three reasons: first, the combination of secrecy about some things (and more on this below) plus aggressive recruiting means that they have a bigger surface area for outside research than most firms; second, a few of their decisions, like systematically buying black swan insurance and using the obscure Ocaml language, make them an interesting outlier; and third, it's hard to argue with success: Jane Street earned $6.3bn in the first half of 2020, up more than 10x from the year before ($, FT).

In one sense, every investor is a market maker and the only difference is their timeline. Jane Street is far along the continuum towards strict market-making: being willing to buy and sell assets at a price close to, but not exactly at, the market price.

Two truisms in folk economics are that liquidity is free and that payment, which doesn’t involve credit risk, is also basically free. Both are much more true than they used to be, but not quite true—both of these things are expensive, best provided with the kind of scale that only humans working closely with computers can offer, and (especially as they get cheaper) catastrophic to offer on the wrong terms.

The price of liquidity reflects basically three factors:

So market-makers are, in the average case, getting a steady return in exchange for taking some big tail risks. Jane Street's early specialization was in exchange-traded funds, but they also trade equities, bonds, options, commodities, and crypto.

The exchange-traded fund market is a great case study in how these kinds of firms can make a profit and provide liquidity. ETFs are a nice, simple structure where a tradable asset represents a basket of securities (which could be an index fund like the S&P, or oil futures, or a set of positions returning triple the inverse return of an index, or whatever). "Authorized participants" can exchange a block of the relevant securities for a block of 50,000 ETF shares, or vice-versa. So a market-maker can continuously make bets on the gap between the price of the ETF and the price of its constituents, and if they’re an AP they can periodically close the arbitrage by delivering ETF shares or the underlying.

That's the trivial part, which you could do manually in excel if you had quick reaction time, and also a time machine that would take you back to the late 90s. Today, things are more complicated. And Jane Street has grown up along with those complications: their original system was, in fact, written in Excel (with lots of VBA and C#). Today, their systems are written in Ocaml, a language that would be fairly obscure if it weren't for Jane Street itself.

The tricky part is finding the million tricks that help you spot adverse selection, detect patterns in the market, determine when a 99%-hedged position is a better deal than paying the transaction cost for a 100%-hedged position, and reselling some of this liquidity-provision skill to other investors. One example of this is NAV trading (Jane Street has a paper here), where an investor wants to place a large trade in an ETF and agrees to buy it at some future point at whatever its net asset value is, less some small fee. The fee example they give is five basis points, so for their hypothetical $30 million trade, that's a fee of $15,000. It's a slim margin for a low-risk trade, but it's not a no-risk trade. In this trade, what the client is paying Jane Street to do is a fun act of financial alchemy: transmute the liquidity of an ETF's constituents into liquidity in the ETF itself. Jane Street can do this by just buying the underlying assets of the ETF, by buying the ETF itself, by buying things that correlate to the value of the ETF such that they'd be comfortable being long the ETF and short the other assets as a hedge, or any combination of these. So the real deal is that the counterparty is paying Jane Street 0.05% to find the most efficient way to let them take the risk they want to take, and also paying them for the possibility that selling this risk will be done imperfectly.

Some parts of this process can get automated, and some should require manual intervention. And this gets to what's fundamentally so interesting about the business: Jane Street is in a constant process of finding and extending the efficient frontier of where computers can replace humans in finance. The firm has researchers, programmers, and traders, but the exact definitions of these jobs seem quite porous. A trader's job is partly to take risk, and partly to accumulate data on what kinds of risk can be systematically measured and hedged away.

Full automation is tempting, but it's critical to do it in a judicious way. As Jane Street's Yaron Minsky once put it, "There is no faster way for a trading firm to destroy itself than to deploy a piece of trading software that makes a bad decision over and over in a tight loop." (This was written in November 2011. About a year later, Knight Capital deployed a new trading system that accidentally activated some old code it wasn't supposed to. The company engaged in a whirlwind of trading with itself, turning 212 orders into 4 million accidental orders covering 397 million shares—making a bad decision over and over again in a very tight loop—and, within a few hours, had destroyed itself.)

The best thing for an algorithmic trader is another algorithmic trader who is overconfident but not as good. Let's say one both traders are using one signal, but the smarter one is using a second signal that sometimes suggests the opposite trades. In cases where the signals point in opposite directions, the worse trader is a wonderful source of liquidity for the smart one.

Human discretion is a great safeguard against this: for n-of-1 events, you want humans in the loop to quickly update when a system is going to make bets that no longer make sense. But for n-of-more-than-1 events, human traders are basically training their replacement, finding underlying patterns and then identifying ways to automate exploiting them. This is a great way for a company to approach trading if it has a deliberately flat hierarchy with vague distinctions between roles; a trader is always implicitly doing research on how to effectively implement strategies and how to spot their edge cases, and is ideally going to sketch out an automated fix once the nature of the fix is obvious.

If you think about systematic trading, the fun thing to think about is picking up a continuous stream of income from taking the other side of less sophisticated traders. But what you actually want to spend most of your time on is thinking about ways things can go wrong. It's not a short list.

As with all jobs, you can make a selection effects-based generalization about traders: the experienced ones are obsessed with risk, because the ones who aren't obsessed with risk get the experience of being fired.

Market makers have a special reason to care unreasonably about risk: a market-making strategy with no edge is exactly equivalent to writing put and call options on the underlying security. That's worth unpacking in two senses, one more impressionistic and one more theoretical:

So a market-maker is, by default, in a position where they're selling insurance against extreme market moves. This is a tough place to be. Every time a trade successfully executes, there's some probability that it's because the trader on the other side just wanted liquidity, and some other probability that every other market maker saw that it was a bad trade and you're the one who got stuck making it.

But there are ways to mitigate this. One of them is to devote a lot of effort to finding every possible indicator of adverse selection, continuously model them out, seek out counterparties who are less likely to be well-informed, etc. In fact, this is one advantage of human traders: they're humans, dealing with other humans, and can get a sense of one another's personalities, propensity to bluff, etc. If liquidity provision is a market for lemons, where any big counterparty can represent a colossal risk, then traders' relationships are one way to mitigate that.

You can view the entirety of market making as a grand effort to write lots of theoretical options, collect a juicy premium, and develop an endless toolkit for dealing with adverse selection.

The other mitigation strategy is: just buy some puts. Markets usually don't crash upwards, but they do have a habit of crashing downwards. And for a market-maker, a crash is a uniquely interesting situation: volume is high, spreads rocket up because people are afraid to trade or don't have the liquidity, so an active participant can make a staggering amount of money. (I liked this Reddit AMA: "Yeah, 08-09 was insane. I've heard stories. No one knew what the fuck was going on and everyone was on edge. Then it all turned out fine and everyone got PAID.")

Jane Street has been in the habit of buying out-of-the-money put options and rolling them over. AQR has a nice paper on how this typically works; the chart on page two shows that buying puts loses money almost all the time and has little blips of outperformance during crashes. For most investors, it's a bad deal. But for a market-maker, it means that they're better-capitalized than ever at exactly the time when they have the most opportunities for profitable trades. They've been looking at tail risk for a long time. At year-end 2007, they had the largest exposure of any institutional investor to Bear Stearns puts ($, WSJ), though that may have been part of a different strategy.

A market-maker might be the part of the private sector best-equipped to profit from tail risk insurance. They have an immediate way to put it to good use for themselves, and meanwhile for the rest of the market participants, it's handy to have someone willing to trade—transaction costs won't be the bulk of your losses during a market crash, but they do add insult to injury, and one reason markets crash is that liquidity disappears. So firms that make markets and own out-of-the-money puts are probably saving investors tens of basis points on transaction costs when those investors trade during a panic, but might also be saving them percentage points of capital if the crash is less severe because there's a natural well-capitalized buyer.3

Finance is intimately tied to language: money is the high-order bit, and finance is one of the easiest ways to convert a communications slip-up into a sizable loss. One fun place this shows up is in the industry's insistence on verbal disambiguation, at least in the parts of the industry that are still phone-centric: people and stocks get nicknames because you don't want to sell the wrong "Liberty" on behalf of the wrong "Bill" (an example not chosen at random; The Bond King apologizes for how many people in the book are named "Bill"). It's also why stocks have tickers; a lot can happen in the time it takes your ticker tape to print out UNITED STATES STEEL CO so you might as well call it X instead. Traders, shell scripting wizards, and haiku writers are at the top of the heap for compressing as much information as possible into the minimum number of characters or syllables.

At a firm focused on automation, that means thinking carefully about programming languages.

Ironically, "thinking carefully" is not the origin story of Jane Street's highly idiosyncratic decision to use the Ocaml language. The origin story is that they had a crufty system built on Excel, and hired a part-time researcher to build some analytical systems. That researcher, Yaron Minsky, chose Ocaml because he liked it, and because he didn't expect anyone else to have to maintain it afterwards. But then he decided to stick around to run a research group that used Ocaml, and a few years later convinced the rest of the company to move to Ocaml, too.

Programming language choices lead to endless debates, and in fact one of the reasons to use a popular language is that if everyone slightly hates it, at least you'll avoid factions where one group loves it and one loathes it. One solution to the political problem is to make an esoteric choice early on, and then filter hires in part by whether or not they think it's a bad one. That can be tricky, because a company can end up wedded to a subpar technology, but Jane Street has found a novel solution to that, too; they contribute lots of open-source libraries, and help keep the language alive and up-to-date. The more people there are who can throw together side projects in Ocaml, or who are writing other libraries for their own convenience, the better-off Jane Street is. Other companies can free-ride on this somewhat, but there are two likely cases here: first, if they're not prop traders, it's good for Jane Street to increase the population of Ocaml-fluent programmers and the number of contributors to new libraries. Second, if they are prop traders, Jane Street can probably outbid them for talent, since returns in systematic trading tend to compound.4

Given the endless fractal nature of programming debates, and given that the readers most interested in these debates already have strong opinions, I'll focus on two areas where using Ocaml makes a difference.

First, Ocaml's design requires programmers to be very explicit about what data they're using and what they're trying to do with it. Excel is happy to tell you that the square root of today's date is 2:22pm on July 29th, 1900; Ocaml will insist that you figure out what you meant to take the square root of first. This has the obvious advantage that it spots certain kinds of bugs before code goes live; it simply won't work if you make certain categories of mistakes, even if what you're doing just consists of known mathematical operations on numbers. It won't catch every bug, of course, but this is helpful.

More subtly, this system makes it easier to reason about complex systems, since it's easier to see what code is executing when, and what it's trying to accomplish. Complex systems generate edge cases, and in trading edge cases generate bankruptcies. The more a trading system resembles a theorem instead of a recipe, the easier it will be. And code that explicitly declares types also has a form of documentation that's automatically audited by the compiler—you can predict often that the code will do what it says because if it doesn't do that, it won't do anything.

A second reason is consistent performance. A trading strategy needs many things to go right at the right time. Suppose there are four different components: a data feed, a pricing model, a trading system, and a risk monitor. The entire strategy's latency is ideally determined by the slowest—and I say "ideally" because if your trading system is acting right this second on what the risk system said ten minutes ago, and that's a ten-minute period during which volatility exploded and it overwhelmed your risk system, you can find yourself in a very bad situation. Pure speed matters, all else being equal, but one of the things that is not necessarily equal is the predictability of that speed.

Exchanges run a sort of bug bounty program where the prize is not necessarily a positive number. Their code is imperfect, and those imperfections are more likely to show up in extreme market events, both because of unanticipated price changes and because the high volume of trades can lead some systems to fail. (One admirable event in the chaotic spring of 2020 was CME's decision, on April 15th, to warn traders that oil prices could go negative and suggest that they test how their system would react in that event. Five days later, it happened.) Prop trading firms need to work around the deficiencies of the platforms they interact with, and having confidence in their own systems is a good way to speed up triage.

Using an unusual language is also a way for a firm to stay truly technical, when the temptation is to drift in the direction of a more traditional finance culture. A sufficiently technical firm can find ways to appropriately compensate people who do boring but necessary work like fine-tuning and refactoring old code. Even if they don't get a performance improvement, reducing future maintenance costs is valuable, and when traders, researchers, and developers are all somewhat fungible, you can pay someone appropriately for cleaning up an existing program, eliminating some redundancy, and making the code shorter and more readable. That's right: they can reverse the classic finance formula and pay a big bonus after someone obliterates a bunch of lines.

One of the meta questions to ask about an automated trading firm is: what do they want to ensure is the most reliable piece of code they have? Yaron Minsky has answered this in an interview: "We have lots of ways of turning things off, including a literal physical big red button." There are many ways to avoid blowups, but the future contains plenty of unprecedented events, and the only way to deal with them is to find a way to stop doing anything else until it's clear what to do next. And this is a case where readability, clarity, and reason-about-ability are all critically important: as anyone who has worked for a big government agency, corporation, or other bureaucracy knows, the hardest feature to engineer is an off-switch.

One thing you'll notice about quants is that they do not like to talk in detail about what they work on. Depending on the context, you might hear about an asset class, or a trade frequency, but actual strategies are kept very close to the vest. This is a notable contrast with non-systematic investors, who are all too happy to tell you what they like or don't like (at least once they're done putting on a trade). But this secrecy extends beyond talking about what they're doing—quants are even reluctant to talk about what strategies they used to use, even if those strategies no longer work, or about what they've looked into, even if it didn't produce a profit. This love of secrecy can reach truly pathological levels: when Bloomberg did a story on a (retired) founder of Jane Street, the company said it would comment and then declined to after four requests and the article's subject refused to disclose the names of his cofounders (information that's publicly available).

The secrecy about live strategies makes sense: if you have a source of alpha, you're splitting it with everyone else trading on the same signal, and the most pleasant split is to keep 100% for yourself. But the core of the trading business is not just executing current strategies, but constantly replacing them; most signals decay over time, whether it's because other traders catch on to the opportunity or because the traders who create that opportunity figure out what they're doing wrong.5 Avoiding information leakage even happens at the trade level. Every trade expresses a view, and one constraint on traders is that they want to be careful not to tip their hands. There can be times when it's optimal to take a position more slowly, and perhaps get a smaller one, rather than trading aggressively and prompting your counterparties to figure out what you know. Secrecy at every level of abstraction is a good way for a firm to get the most value out of the ideas it develops and the people it hires.

This creates some fascinating unit economics for prop trading firms. Every employee they hire has a high opportunity cost—when they're doing multiple rounds of interviews with multiple employees in each, they're burning lots of six- and seven-figure time vetting someone. Hiring well would be expensive even if the new employees weren't getting paid so well.

So when the firm looks at new opportunities, whether it's trading a new asset class or adding a new signal to an existing strategy, they have to run some basic calculation: expected alpha from the signal; expected upfront time spent on research and software implementation; the expected ongoing cost in terms of capital, time, and distraction from having human traders help implement the strategy; and the risk that either the research won't pan out or that it will produce a signal that looks good in theory and doesn't work in practice.

In this model, knowledge about which strategies don't work is incredibly valuable! It saves everyone a lot of time, and in a world where the value of existing signals is persistently decaying, that time is enormously valuable. Some strategies work for a while, fade out, and then start working again—perhaps the trade migrated from specialized hedge funds to banks in the 2000s, but banks were too risk-constrained to pursue it after the financial crisis, so the trading avenue opened up again. And some strategies might work by analogy; perhaps a market anomaly shows up in any country that hits middle-income status and develops a domestic base of equity investors, and knowing that it worked in Japan in the 70s can tell you something about how well it will work in Poland today.

This has another even more interesting takeaway. When unit economics are this visible, and when specific strategies are likely to become obsolete, it means the scope for strategic decisions is radically reduced. A strategy is either positive-EV or not; a new data source is either going to produce incremental profits that pay for it, or it won't; adding a new asset class might lead to some mild spillover effects—perhaps getting into currencies will add some alpha in pricing US-based exchange-traded funds tracking foreign companies—but a firm that's hedging out that part of the risk is basically outsourcing sophistication to their counterparties.

This doesn't mean that prop trading is a completely strategy-free business. It just means that the overwhelmingly important strategic choices are around hiring, company culture, and supporting Ocaml. Those are the cases where there are potential nonlinear returns, and complementary opportunities whose upside exceeds the sum of their parts. So one reason Jane Street spends so much energy on finding hiring the right people, running job ads on programming theory blogs or posting puzzles, holding guest lectures and after-hours chess tournaments, etc. is that every company reaches the point where "do something strategic!" is an itch that's worth scratching, but in prop trading there are very few legitimate ways to scratch it, so companies are forced to excel.

Is "liquidity provision" a socially useful thing, or is it a way that millionaire options traders rationalize the fact that they're not building something socially useful but less lucrative? This is a genuine, and difficult, question. If a firm asks employees to constantly think about edge cases, novel risks, and whether or not their behavior makes sense from first principles, it's only a matter of time before they go meta and start worrying that the big risk is that they're wasting their lives.

Maybe this is why so many of them retire in their 30s: the utility of money is roughly log(wealth), but the disutility of guilt seems more like a linear relationship. And if one of your hiring screens focuses on introspection, you may accidentally filter out the sociopathic personality traits that minimize guilt.

But there are two moral cases to be made for why someone who's good at programming and math should take the offer from Jane Street and not, say, go work for a fusion startup or try to cure cancer:

Either way, pushing out those efficient frontiers is a fun and engaging job. And that may be its own justification. We don't demand that chess champions use their skills in something with more real-world applications, or that concert pianists find a more practical outlet for their manual dexterity and attention to detail. Some jobs take a field with a general category of problem and create the cleanest, most abstract, and most challenging form of it. Sometimes it's not until you push a problem to its limits that you figure out what problem you're really solving.

Subscribe now

Further reading: The author of The Laws of Trading worked at Jane Street, and it's a good book on the philosophy that systematic trading inculcates. This interview with Yaron Minsky is a great look at their decision to use Ocaml. If you're interested in more about the mechanics of exchanges and trading, the Hide not Slide Substack is good. As a starting point, here's their writeup on Jane Street. Max Dama on Automated Trading is old, but a very helpful overview of the industry for technical people. If you want to learn Ocaml, Jane Street's Yaron Minsky has coauthored a good book on it.

Tegus is the first port of call for M&A professionals and institutional investors ramping up on an industry or company.

Get access to a database of 35,000+ expert call transcripts, spanning 5+ years, or schedule expert calls through the platform for a fraction of the usual cost.

When thousands of research analysts are pooling their expert calls into an on-demand database, using Tegus is table stakes. It's the leading platform for due diligence and primary research.

See the power of a Tegus subscription, and get up to data parity with your competitors, with a two week free trial through the Diff.

Try Tegus

US public pensions mostly report annual returns as of June 30th, and so far the results don't look great. Pension funds have been a looming problem for a long time, and early in the pandemic it looked like they might finally reach the point where they'd need a bailout and restructuring. For now, they're still targeting higher returns than they can likely achieve, which forces them to move into riskier asset classes to meet their benchmarks. This means that events that reduce asset valuations, i.e. increase expected future returns, hit them even harder because their portfolios are more-than-optimally concentrated in the riskiest assets.

Amazon has announced plans to shut down its Amazon Drive service at the end of next year. The classic Hacker News thread about the announcement of Dropbox includes some skepticism from users who see it as a marked-up version of Amazon S3. There's a lot of value in bridging the gap between cheap storage and a product people will actually pay for, and bridging that gap is nontrivial even for companies that make a point of customer obsession.

Disclosure: long AMZN.

Drone delivery is booming as a way to deliver contraband to prisons. New technologies often have a comparative advantage in the military or crime, because those are two cost-insensitive fields where unit economics are strongly influenced by physical risk. One response is to restrict their use, but that's a good way to ensure that crime still has a comparative advantage (the person with the lowest marginal cost for flying a drone illegally is the one who's flying a drone to get away with something else illegal).

Uber is showing drivers how much they'll be paid if they accept a trip. Ride-sharing is a business that scales to the point where churn is the dominant concern, and as it's matured, it's gotten harder for companies to compete on coverage and arrival time. Making concessions to driver preferences will mean that some users won't get as much service, but it also provides another dimension for the company to determine the optimal pricing tradeoff.

Early in the Ukraine war, I argued that the run-up in energy prices due to concerns about the war made it more likely ($), by increasing the cost of energy sanctions. This is still happening: a recent plan to ban insurance for Russian oil shipments has stalled because the UK is reluctant to sign on ($, FT), partly because higher energy costs would be so hard to bear. This presents the interesting possibility that the length of the war is partly a function of the world's economic health: a severe recession, and an ensuing drop in oil prices, would reduce the relative cost of sanctions and make continued conflict untenable for Russia.

Diff Jobs is our service matching readers to job opportunities in the Diff network. We work with a range of companies, mostly venture-funded, with an emphasis on software and fintech but with breadth beyond that.

If you're interested in pursuing a role, please reach out—if there's a potential match, we start with an introductory call to see if we have a good fit, and then a more in-depth discussion of what you've worked on. (Depending on the role, this can focus on work or side projects.) Diff Jobs is free for job applicants. Some of our current open roles:

Hire With Diff Jobs