[arzt]

I wonder if anyone has insight into how they have been able to do this consistently in the modern era of quantitative trading (this article had scant detail)? His returns are such an outlier and strategies such a closely guarded secret that they leave people on Wall Street in awe.

[jandrewrogers]

I have anecdotal insight. Many years ago I met with Jim Simons a few times. He had taken an interest in some of my theoretical computer science research (by referral, it was never published). I don't know any details of their strategies but one could infer it from their specific theoretical interests in those conversations.

My impression was that they were doing sophisticated sparse signal reconstruction and then applying some pattern induction algorithms against those signals. The former was, to the best I could discern, absolute state-of-the-art; the latter was merely competent (I've never talked to anyone that was exceptional at this bit). It has been a while but my impression was that a real strength was that this process was highly automated and general, so it could be thrown against almost arbitrary data sources. It is not difficult to imagine how one could build a sustainable and significant edge with this capability. I could be wrong but I don't think I am that far off. Very good math brains, even compared to many of their peers, based on my limited exposure.

I find their performance believable, given the above.

[deepnotderp]

You're not wrong, you're (almost) 100% correct, although they've certainly evolved their techniques since.

EDIT: I'm not saying that's everything, but automated signal extraction is a major part of their secret.

[trhway]

It is also impressive how many people there have intelligence agencies connections. And Medallion , their employees fund, underpaying 7B of taxes by misrepresenting short term as long term gains is very impressive too.

[throwawaymath]

That's right. A lot of people are aware that RenTech scoops up talent from math and theoretical CS departments. But it's less well known that many of Simons' old colleagues from the NSA also contribute math and CS talent by referring them to Simons.

Of the people I know who work at (or used to work at) RenTech, one actually joined after working at the NSA. His PhD thesis was a joint collaboration between Harvard's physics department and the NSA.

[rjdagost]

To my knowledge, all he has ever said on the subject is: "I think people would be quite surprised if they knew how simple our methods are". You probably won't ever hear more information than that, until their strategies stop working.

[Ragib_Zaman]

Flip side to this - I remember studying a series of papers in statistics/optimisation/machine learning with highly non-trivial content published between 1998 and 2008 by the Della Pietra brothers. I had assumed they were mathematicians/statisticians at some major research university. At the bottom of one of these papers it had some personal blurbs which stated they had both been at RenTech since 1995 working on "statistical methods to model the stock market", which surprised me greatly given the amount of research they output. I assume the content of those papers were applied to their strategies, in which case there would also be a good amount of people who would be surprised if they knew how advanced their methods are. I say this especially in comparison to the "quant" strategies I know many other trading firms use/have used, which are actually often quite simple. Indeed, at some places they seem to think any systematic/automated strategy is "quant"...

[VladimirIvanov]

Do you remember what the titles of papers? I'd like to read them.

[throwawaymath]

You can find a bunch of papers published by people at RenTec. Search MathSciNet for "Renaissance Technologies" as the corporate affiliation for the author.

Likewise, search Google Scholar for "@rentec.com", or "Renaissance Technologies."

[Ragib_Zaman]

As throwawaymath mentioned, it's not too hard to find papers written by them, but I'll highlight some papers that I found particularly interesting.

Many optimization problems that arise in machine learning can be viewed as minimizing a particular Bregman distance subject to affine constraints (I'll call these "Bregman distance optimization problems").

In [1], the authors develop some quite general and widely applicable convex analysis type results for Bregman distances and use those results to give the technique of "Auxiliary Functions" which can be used to derive and prove the convergence of algorithms for particular Bregman distance optimization problems.

In [2], the authors show that finding the optimal parameters of two quite different approaches of binary classification, AdaBoost and Logistic Regression, can both be simultaneously viewed as the same Bregman distance optimization problem (with slightly different initial parameters). They then present several algorithms for solving this unified problem (and thus, for optimizing both AdaBoost and Logistic Regression), and prove the convergence of their algorithms with the method of Auxiliary functions developed in [1]. This paper was the first general proof of convergence for AdaBoost which was proposed by Yoav and Schapire (one of the authors of [2]) and earned them the Gödel prize in 2003.

If you don't mind me plugging in some expository work of mine, I wrote an essay ([3]) giving an introduction to Bregman distances and is pitched at a lower level than [1] and [2]. The end of the first chapter discusses in particular the general Bregman distance optimization problem, setting it up in the same framework as [1] and [2], and the final chapter presents the algorithm and proof of convergence originally given in [2] but hopefully with a slightly simplified presentation due to focusing only on the Logistic Regression case of the algorithm.

In case you want to play around with it, I have implemented the algorithm from [2], as well as a related algorithm from [4] which incorporates L1 (Ivanov) regularization into that algorithm, both being available at [5]. The middle chapters contain brief discussions on the relation of Bregman distances with Exponential Families, and on Generalized Linear Models, which are relevant to the overall purpose of the essay but not so closely related to the content of [1] and [2] and may safely be skipped.

If you generally enjoy the types of problems discussed in [1] and [2], you might also enjoy some of these papers: [6] [7] [8] [9].

--------------------------------------

[1] - "Duality and Auxiliary Functions for Bregman Distances". Stephen Della Pietra, Vincent Della Pietra, and John Lafferty, 2001.

[2] - "Logistic Regression, AdaBoost and Bregman Distances". Robert Schapire, Michael Collins, Yoram Singer, 2001.

[3] - "Optimisation of Bregman Divergences". Ragib Zaman, 2018. https://github.com/RagibZaman/mathematical-optimisation/blob...

[4] - "Bregman distance to L1 regularized logistic regression". T. Huang and M. Gupta, International Conference on Pattern Recognition, 2008.

[5] - Notebook containing implementations of Bregman divergence based algorithms for Logistic Regression from [2] and [4]. https://github.com/RagibZaman/mathematical-optimisation/tree...

[6] - "Legendre functions and the method of random Bregman projections." H. Bauschke and J. Borwein, Journal of Convex Analysis, 1997.

[7] - "Inducing features of random fields". Stephen Della Pietra, Vincent Della Pietra, and John Lafferty, 1997.

[8] - "Statistical learning algorithms based on Bregman distances". Stephen Della Pietra, Vincent Della Pietra, and John Lafferty, 1997.

[9] - " A maximum entropy approach to natural language processing". Adam L. Berger, Stephen Della Pietra and Vincent Della Pietra. Computational Linguistics, 1996.

[doctorpangloss]

> anyone has insight

In a world where all their statistical arbitrage has ceased to be viable, financial professionals believe that the outsize returns of the Medallion fund in recent history are siphoned from their institutional funds via shell games.

Then, their famous ability to make money even in down markets is attributed to how they smooth out an extremely profitable short term trade, that may have occurred years ago, over the course of many years. This sort of surfaced in their tax avoidance lawsuit too.

You too can make a "Medallion Fund." First, make a venture investment in something that turns out to be Facebook. Keep that equity secret, even when Facebook goes public. All that time, say that your fund has gained 15% year over year, even during a recession. Do this for years until you have taken your 2% management fee to your liking. You've turned your 200x return that happened all at once into something that looks like the world's greatest hedge fund. Lawfully of course.

[throwawaymath]

No, that wouldn't work. The options basket strategy you refer to did have nontrivial tax advantages, but

1) Those tax advantages can only improve returns which are already fundamentally strong, and

2) There is no "smoothing" effect achieved; the options baskets do not defer returns for years at a time.

I get that the cynical take is, as ever, the attractive one on Hacker News. But speaking frankly, what you're saying doesn't actually make sense. Among other problems with your explanation, there's a straightforward wrinkle. While it's not available to the general public, other institutions like Bloomberg and WSJ have had (and still have) access to audited attestations of Medallion's track record over a timespan of 25 years.

[spac]

“The fund has been closed to outside investors since 1993 and is available only to current and past employees and their families. “ https://en.m.wikipedia.org/wiki/Renaissance_Technologies

[simonpure]

These are some quotes from various interviews -

"But we look at anomalies that may be small in size and brief in time. We make our forecast. Then, shortly thereafter, we reevaluate the situation and revise our forecast and our portfolio. We do this all day long. We're always in and out and out and in. So we're dependent on activity to make money."

Renaissance essentially attempts to predict the future movement of financial instruments, within a specific time frame, using statistical models. The firm searches for something that might be producing anomalies in price movements that can be exploited. At Renaissance they're called "signals." The firm builds trading models that fit the data.

When the trading starts, the models run the show. Renaissance has 20 traders who execute at the lowest cost and without moving markets, crucial requirements for quant investors trading on narrow margins. But the models decide what to buy and sell. Only in cases of extreme volatility, or if the signals appear to be weakening, does the firm sometimes manually cut back. Says Simons, "We don't override the models."

...

"We search through historical data looking for anomalous patterns that we would not expect to occur at random. Our scheme is to analyze data and markets to test for statistical significance and consistency over time," says Simons. "Once we find one, we test it for statistical significance and consistency over time. After we determine its validity, we ask, 'Does this correspond to some aspect of behavior that seems reasonable?'"

...

Many of the anomalies we initially exploited are intact, though they have weakened some. What you need to do is pile them up. You need to build a system that is layered and layered. And with each new idea, you have to determine, Is this really new, or is this somehow embedded in what we've done already? So you use statistical tests to determine that, yes, a new discovery is really a new discovery. Okay, now how does it fit in? What's the right weighting to put in? And finally you make an improvement. Then you layer in another one. And another one.

...

Everyone in the company read the book about LTCM. It makes you wary in a general sense. Our approach is very different. We don't start with models. We start with data. We don't have any preconceived notions. We look for things that can be replicated thousands of times. A trouble with convergence trading is that you don't have a time scale. You say that eventually things will come together. Well, when is eventually?

...

https://www.institutionalinvestor.com/article/b151340bp779jn...

...

"Have an open atmosphere. The best way to conduct research on a larger scale is to make sure everyone knows what everyone else is doing... The sooner the better - start talking to other people about what you're doing. Because that's what will stimulate things the fastest. No compartmentalization. We don't have any little groups that say. this is our system and we run it we get paid because of it. We meet once a week - all the researchers meet once a week, any new idea gets brought up, discussed, vetted, and hopefully put into production. And people get paid based on the profits of the entire firm. You don't get paid just on your work. You get paid based on the profits pf the firm. So everyone gets paid based on the firm's success."

In sum, the secret is:

"Great people. Great infrastructure. Open environment. Get everyone compensated roughly based on the overall performance... That made a lot of money."

[codingslave]

They built a system where any data set can be pushed in, joined with the rest of the data, and then automatically made inferences off of for trading.

[throwawaymath]

This isn't really correct.

Renaissance has always put massive personnel and technology investment into its data processing and analysis pipeline. But there is no "automatic inference" generation. It's not so much brute forcing alpha as it is streamlining the process of hypothesis testing for research scientists so that strategies can be very rapidly generated and examined.

Automatic inferences would be susceptible to two major risks. First, you'd run into spurious correlations at the dimensionality of data we're talking about. Those spurious signals would have to be pruned, significantly reducing any advantage.

Second, you'd decouple the strategy generation from financial domain expertise. The strategies are not developed in a vacuum - contrary to popular belief, quant trading firms do apply financial acumen.

[codingslave]

I work at the intersection of quant finance and fundamental analysis, it can absolutely be automated. The question of what can be determined from raw data like credit card transactions and mobile phone locations is a whole topic in itself, but thinking that you need manual intervention to trade on those signals is completely misguided and its a waste of time for me to argue with you

[throwawaymath]

I didn't say you need manual intervention. I said you cannot do automatic inference generation. What you're referring to does not provide automatic inference generation, i.e. you cannot brute force hypotheses. That's why you still employ researchers.

More to your specific example, I've also worked with the alternative data you're talking about and it doesn't offer automated inference generation. You implicitly have a hypothesis (or several) in mind when you're working with things like credit card transaction data from Yodlee or Second Measure.

Automation is a continuum. What you're talking about is automating time series analysis. I never said you can't do that.

[codingslave]

"you cannot brute force hypotheses", this isn't really true. Credit card data has notorious gaps and bias, but that doesn't mean that an algorithm cannot determine and make decisions about certain situations within that data. For example, if I receive a daily feed file of walmart transactions and the data is increasing in some kind of confidence measure that walmart will beat earnings, I stand to make a good sum by jumping into the market before competitors. It's common that all competitors are aware of the situation, aware of the possible alpha, and competing on speed/accuracy for it. So the superior ability of my model to take a calculated risk from incomplete data (as well as combine other data sources) is one way for me to make money. I may build a model of the common structure of the transactions, ensuring that any signal coming from the data is a real signal, and not one coming from one of the many data quality issues. In the case that my data quality classifier is pushing out high confidence, the result is saying earnings will beat, and other data sources are saying the same, then my model buys. Completing this kind of analysis by hand is too cumbersome (for my usual length of holding period), the money is in who gets there first. There are many ways, some more conservative.

[hnhg]

In what timescale though? There are huge differences between the timescales of "realtime" (say HFT), a second later, a minute later, an hour later, a week later and so on. Do they operate at all of them?

I have no specialist knowledge, btw, I'd sincerely like to know!

[neplus]

If you're interested in this, you'll likely enjoy Gregory's interview on Masters in Business (a Bloomberg podcast) from last Wednesday. Also, Gregory's book will be out in a few days.

If memory serves, in the aforementioned podcast Gregory mentions that the RenTech generally holds most things for a few days (sometimes a few hours). However, they don't engage in HFT or HFT-like trading. This was surprising to me as I assumed it was all reasonably short holdings (relatively speaking), although I knew they weren't a pure HFT firm.

I also seem to recall Gregory mentioning there's some kind of running joke internally that their trading systems aren't nearly as good as they should be (or like what you would find at HFT firms). Given the intellectual and monetary heft within RenTech perhaps that's a bit of false modesty on their part.

I'll be interested in reading Gregory's book as he does seem to have put together a lot of novel information on RenTech. However, he does seem to suggest that very little of the day-to-day workings of the firm will be explored, which would obviously be immensely interesting.

EDIT: RenTech has several funds, it should be noted. Some of which still take outside capital. What I've said above may have only been applicable to the Medallion fund.

[mbostleman]

In the podcast he said two days was average. And on frequency, he called them medium.

[Blammar]

From what I recall, their approach is mostly what you might call "special situations." That is, their analysis looks for significantly incorrectly priced items, and purchases/shorts them.

The Medallion Fund is kept fairly small so it can capture these items without changing their prices substantially. That is, the fund owners have to take their 40% return each year out of the fund.

[trhway]

>That is, the fund owners have to take their 40% return each year out of the fund.

The Medallion is for the employees money. That reminds about salary payment schema in Russian banks in 199x (don't know for today) - employees got to open very special, employees only, accounts paying extremely high, many times beyond the market, interest. The bank account interest got beneficial taxation for the employees, and the bank didn't have to pay various taxes, like social security, etc., which an employer would normally pay on salary. Of course how much an employee could put into such an account had a limit specific for a given employee, and thus the employee did have to regularly take the money out of the account.