4 ms·

Virtu is a market-maker. Comparing their win record to RenTech makes no sense. They have a high win-rate but so did brokers in the 70s...Virtu is doing the same

Virtu is a market-maker. Comparing their win record to RenTech makes no sense. They have a high win-rate but so did brokers in the 70s...Virtu is doing the same thing as them (they are also APs for ETFs...again, the innovation there has really been able to make markets at very low cost). There are hundreds of other quant firms with public records. RenTech has better numbers because they stayed smaller. It is difficult to generalise but firms either grow assets to a point where the market moves against them when they trade/returns drop or they go into strategies with lower returns at scale (btw, both things are common outside of quant too). They aren't doing HFT. Some quant strategies are tangential to HFT, for example front-running news was a big strategy in the early 2010s...it is somewhat latency-based but is still distinct from HFT, which tends to refer more to making markets. The book says they tried hmm/speech stuff and it didn't work. It is likely they are doing more complex things now but Nick Patterson said they were using linear regression for most of the 90s. Generally speaking, this is a common misconception: people believe that because the results are good, the model must be more complex. This reflects how university courses are organised but the real world isn't like that (one big advantage that RenTech had was data, they had data that no-one else had for a very long time, another big factor is execution...these kind of practical edges are far more important than people think). Also, they use a ton of leverage...their returns actually compare pretty well to what fundamental managers can achieve outside of a public fund. Having investors is a significant limitation because they will often force you to behave in a way that reduces returns (i.e. redeeming at the worst time, asking for risk reductions at the worst time). The structure is very kind to gross returns. Retroactive reallocation of successful trades is very old. The SEC cracked down on this in the 80s, it is very easy to prove, and it is very unlikely that someone doing this would hire a bunch of scientists and then give them a bunch of equity in the fund...it doesn't make any sense.

hardwaregeek 6y ago
Yeah in the Zuckerman book they mention an employee who worked on getting and cleaning data for decades, far before data science techniques were common in finance. I could see RenTech having good quality data going back decades being a serious advantage.
beagle3 6y ago
Indeed, but it’s important to differentiate between model execution complexity, and model optimization complexity. A linear model, if the inputs are e.g. squares and variable products, is a quadratic equivalent. A logistic regression yields a linear model; you could tell people it’s linear regression and they’ll likely believe you, but won’t be able to replicate. There’s a huge issue with itrelevant inputs and how to identify them - Emanuel Candes has done a lot of work on that, as did Rob Tibshirani. Saying “linear models” is saying little more than “using math”, even if that’s true, and even saying “linear regression” doesn’t give much information about what is actually being done. The bottom line is that the decision boundaries are usually simple and often have linear form - but the variables in that linear form are not raw data, but rather nonlinear transformations of it (e.g. order imbalance)

hardwaregeek 6y ago

Yeah in the Zuckerman book they mention an employee who worked on getting and cleaning data for decades, far before data science techniques were common in finance. I could see RenTech having good quality data going back decades being a serious advantage.

beagle3 6y ago

Indeed, but it’s important to differentiate between model execution complexity, and model optimization complexity. A linear model, if the inputs are e.g. squares and variable products, is a quadratic equivalent. A logistic regression yields a linear model; you could tell people it’s linear regression and they’ll likely believe you, but won’t be able to replicate. There’s a huge issue with itrelevant inputs and how to identify them - Emanuel Candes has done a lot of work on that, as did Rob Tibshirani. Saying “linear models” is saying little more than “using math”, even if that’s true, and even saying “linear regression” doesn’t give much information about what is actually being done. The bottom line is that the decision boundaries are usually simple and often have linear form - but the variables in that linear form are not raw data, but rather nonlinear transformations of it (e.g. order imbalance)