NuclearPhynance Archive

@u1234, this is awesome. Much appreciated. To be honest if there’s any more color you’re willing to share on any of the other funds or shops (still curious about quantbot!), I’d love to hear it.



Anyone heard of Athena capital research?



Re: medallion running 3-4 Sharpe on ~3 week holds... it’s tough to believe but with many uncorrelated signals and diversification across a huge universe of instruments, I could imagine it’s doable... what would you estimate is the avg. hold of other places running 2.5+ (PDT, quadrature, etc)?

@u1234



Thanks! for the color.



@sigma



Sharpe of 27 isn't really unusual for HFT. It's quite common for shops like Virtu to report never having a down day. That implies a Sharpe of 16 or more. At some point it doesn't really matter, since more of the standard deviation is driven by day-to-day variation in the volume rather than position risk.



@levkly



It's feasible But difficult. You'd need to diversity across a large set of orthogonal factors/exposures, and ruthlessly hedge out any exposures that you don't have signal on. Some back of the envelop math...



Let's start by saying you have the equivalent of 225 orthogonal factors. Obviously there's a lot more than this number of securities in the world. But with a $4 billion fund levered 5X, to get 1/225 of your portfolio requires $90 million in exposure, so most of your returns are driven by blue chip stocks, liquid futures, and major FX pairs. 225 would be the upper-limit even for a globe-spanning operation.



Let's say the typical factor has a daily volatility of 50 basis points. This is roughly the idio vol for a typical US stock. This number doesn't matter too much, the logic mostly holds if you want to scale it up or down. At a three week horizon, that's 193 basis points of volatility.



3.0 annualized Sharpe means a Sharpe of 0.72 at the three week horizon. Across 225 uncorrelated factors that's an individualized Sharpe of 0.048. Using the above volatility, the average trade needs to net 9.2 basis points of return after costs. That doesn't seem huge, however using the three-week horizon window the net signal needs to achieve an R-squared of about 0.2% on the orthogonal factor returns. That's pretty substantial at a weekly horizon.



Obviously this is a toy model, but I think the numbers don't look substantially different for reasonable extensions or modifications.



That being said, I'm relatively skeptical that the bulk of Medallion's returns come from strategies at these horizons. For one, there's the ITG lawsuit from 2009. Basically they were trying to arb out discrepancies between the Posit dark pool and lit exchanges. That certainly looks very HFT-like to me. For another, before merging into Medallion, Nova was doing 10% or more of the volume on the NASDAQ back in the 90s. It seems unlikely given the shifts favoring electronic trading that they would have significantly substantially retreated. Another factor is the amount of computing power they supposedly have. You don't need that many FLOPs to evaluate interday strategies.



I do believe, that there's probably some definition of "average holding time" for which the statement is true. Especially if they're counting very long-lived portfolio hedges or cash instruments. Longer-term strategies tend to be hedged and levered, whereas short-term strats bias flat and don't hedge nearly as much. At any given time the median position in the portfolio, weighted by notional, could be held for a long-time. But I'm fairly confident that a substantial proportion of Medallion's profits come from intraday signals and strategies.

Voloridge is another fund rarely heard off. I think last I heard they had around 1.5bn and hard closed with a waitlist of 13bn of interested capital. Sharpe of around 1.7-1.8 from what I heard 2 years ago. David Vogel is a ML expert.



@EspressoLover

A friend who worked at one of Japanese banks in NY did the lit vs dark Arb for a while. Sharpe of 5 while he worked there. Never made money after he quit to launch his own strategy. The strateg went dead by 2013-14. It was mostly execution alpha. More of a technology thing than alpha from what I heard. At peak he made only 5mn for the bank, barely anything. The alpha from this Arb was really small even at peak.

My advisor has sent many students to Setauket.

All of them must be using techniques very close to what we picked up in our research because the application is quite obvious, and it's likely the basis for improving a high capacity strategy that should outweigh the intraday piece at their size.

The speculation is fairy dust as usual.



/* Nova was doing 10% or more of the volume on the NASDAQ back in the 90s. */



90s is a long way back.

My team averaged almost 10% ADV in the entire of US cash equities 2006-2009.

Ported a few of the same strategies in 2016 when I had to deal with US equities again.

Barely grazed 1% ADV.



/* You don't need that many FLOPs to evaluate interday strategies. */



Deeply mistaken.

DE Shaw's cluster is larger than Rentec's, at least 5000 nodes.

PDT's is about half of that.

Tower has about 8-9 nodes per employee.

Go figure.

I interviewed some guys who worked there for a year. Told me that their experience was that they were promised big bonuses to plug in their code in to their servers and once the code / strategy was implemented, the guys were let go. Not sure how much of this is true.

"Told me that their experience was that they were promised big bonuses to plug in their code in to their servers and once the code / Strategy was implemented, the guys were let go. Heard an exactly similar story with Domeyard from candidates I interviewed recently."



Do you think that IP has any real value once you let go of the trader that designed it? Just curios, I keep hearing about this type of stuff, but I can't imagine code having much value without the people that found the alpha and built it.

"Do you think that IP has any real value once you let go of the trader that designed it?"



@Strange I think the continuity of IP value is a function of the type of IP you're referring to and the firm's culture. In my experience, it's definitely possible to retain the use of the base predictors, market simulation, model-related data structures, quant dev tools and common optimization routines even for years after the author has left. Of course you're not going to rewrite the same couple of souped-up imbalance signals every time 1 guy leaves right?



Despite some reusability, I don't think it ever economically makes sense to promise bonuses and churn staff immediately after they implement some code. There's usually a lot more to it and you're just hearing the resentful side. For starters, I feel it's a lot costlier to hire and train employees than it is to retain them at fair market value. Emphasis on 'fair market value', because I know many employers encounter the situation they cannot offer a higher payout to a senior PM, researcher or trader because there's equally competent employees willing to do the same work for less. This is a different case where their employees are not involuntarily being terminated but rather they are voluntarily deciding to leave.



It also depends on the culture of the firm. It's easier when you have good coding practices in place and cross-pollination among your team members. I'm not a fan of turnover, but I actually admire a few of our competitors for having built up an incredible training cycle where they've been able to turnover their "head of trading" every one or two years and still keep on killing it. It's not possible to do so unless you have a good retention rate or slow alpha decay on your IP and a strong management team. Good for them.

> I can't imagine code having much value without the people that found the alpha and built it.



It happens more than you may think.



"Signals? We have literally hundreds of different signals, across equities and futures. We've been in the business since the 1990s. We use this really smart allocation engine to switch weight between signals."



"Right - so what sort of Sharpe can you get from hundreds of signals and smart allocation?"



"0.75. We've produced nothing but volatility since 2013."



I must have had this conversation dozens of times.

@prikolno

Not important, but good to hear your side of the story. Best .

@ronin yeah it’s a nice story you seem to hear a lot. The straightforward problem is (apart from the possibility that some of the signals actually have negative sharpe) that the benefits of combining multiple signals rapidly diminish when the signals have even modest positive correlation. And at low frequencies building a nontrivial allocation engine is subtle.