Medallion is not unique, there are other firms with comparable win record (Virtu, a Czech one, an Israeli one and a couple of British ones at the very least) but only 5-10% of the size; of all these, only Virtu is public and verifiable, the others aren’t but you can find people who will confirm it off the record.

People were begging Simons to take money. He wouldn’t let them into medallion (why should he share?) but he did start a higher-risk, lower-reward business and let’s people into that.

As far as I can tell, the commonality among those always-winning firms is high frequency low latency algorithmic trading. These days it takes millions of dollars per month just to pay for the infrastructure you need to be able to be competitive - and then you also have to have some nontrivial edge, without which there isn’t all that much profit in having low latency.

What’s Virtu’s or RenTexh’s/Medallion edge? I don’t know. In the past, they seemed to like people with speech/hmm background. But that was before the DNN / differential computing revolution. I have no idea where there edge is now (and actually whether hmm was their edge in the past - but it did seem to be quite common background among their recruits)

That said, they may or may not be tax frauds as well - I have no idea. But I don’t see any reason to suspect they are doing retroactive allocation of successful trades.

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.

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)

And that is why I like the idea of a "trade arbitrarily slowly with limited price change" market versus the current approach of "trade fast with an arbitrary price change" market.

See https://news.ycombinator.com/item?id=24760841 for an explanation of how the trade arbitrarily slowly market could work. Under normal conditions, it would look a lot like the current market does. Except that you're paying less to the HFT folks.

What I didn't describe there is that you could even have a chain of slower and slower markets. With a maximum rate of price change varying from 1% per day to 1% per minute. With the idea that ordinary folks would trade on the 1% per minute market while large institutional orders would be likely to go in the 1% per day market. (And when the price of two markets cross, open orders on the one can match as open orders on the other.)

My understanding is that

a) either there is a huge price lag to the fast market and say you offer some good cheaper that the fast market. Then the HFT would come and buy your stuff and sell it more expensive on the fast market. Until

b) your market becomes illiquid.

In both cases there is little incentive to use your market. It would only make sense for huge trades (similar to take over offers, etc).

If the slow market is outside of the bid-ask spread, then HFT will be happy to move the price to the bid-ask spread. So you're liquid in the direction that moves the price to where it needs to be and not liquid in the other direction.

If the slow market is inside of the bid-ask spread, then HFT is likely to be willing to buy/sell on the fast market and complete the other half of the trade on the slow market. That is, they don't snap up the slow order right away, but they will snap it up to complete trades. Getting a guaranteed trade is better than holding the stock and not trading. This gives liquidity in both directions.

The incentive to use this market for smaller orders is that you are likely to get a price somewhere between the bid-ask spread. I don't have recent data, but a decade ago the bid-ask spread for small stocks was around 2%, lowering to 0.6% for the top 20% of stocks. (As you go to the behemoths, the spread drops farther.)

I'm sure it is smaller today, but if you are a day trader, that spread is a hidden tax that is going to kill you over time.

Even if nobody used the slow market, HFT would guarantee that you get no worse than the spread on a market order. But when traders use the slow market directly, they bypass the HFT middleman and save themselves money.

"liquid in one direction" is a nice way of saying that no trades are happening, which is to say illiquid.

> If the slow market is inside of the bid-ask spread, then HFT is likely to be willing to buy/sell on the fast market and complete the other half of the trade on the slow market.

This seems to hinge on an unrealistic model of HFT as perfect-arbitrage machines that need to immediately close positions. In fact, it would be quite surprising if they were willing to do this, based on how they currently behave. If they were willing to close the loop like that, they would equivalently be willing to do it on existing exchanges, which they could do by posting an order inside the spread (which would, of course, shrink the spread). The fact that that is described as "inside the spread" is a pretty clear indicator that they are not doing this.

> The incentive to use this market for smaller orders is that you are likely to get a price somewhere between the bid-ask spread.

The reason that you don't currently get filled "between the bid-ask spread" is that the whole point of the spread is that it is the region inside which no one is currently willing to trade. If they were, the spread would be smaller. By what magic are they willing to trade inside the spread on your exchange, but not on traditional ones?

Similar principle... IIRC it delays execution to try and prevent HFT. A big part of Flash Boys by Michael Lewis was chronicling the history of what led to this exchange being created.

They take away a lot of the tools that HFT uses to increase their edge, simplify the ordering structure, and try to make as many trades as possible to be between traders instead of the HFT market makers. But they have not fundamentally redone the structure of orders such that prices are guaranteed to move slowly at the cost of indefinite delays in execution.

I don't believe Medallion would be classified as a high frequency trading operation.

The $5 billion in unpaid taxes is a small fraction of their total returns. So again, not a great explanation.

Everything I've read (and you obviously have to take it with a grain of salt) doesn't agree with this assessment. The core fund trades commodities and stocks/options in a pair format with short holding times. To make their public fund, they needed to adopt more scalable strategies which meant longer holding periods.

So it's not a matter of choosing/assigning ex post - they are fundamentally different approaches. They said up front that the public fund wouldn't replicate the internal fund. Whether people listened to them or not is another question.

I get that we should be skeptical, but by the same token I don't think you can says fraud is happening without any evidence. Yes, they had a tax case - but that was related to their derivatives contracts and the tax handling of them. Clearly they were wrong on that - and they've stopped using them - and still been up huge after that.

I'm just not clear why we should assume fraud just because they are successful. To me, it looks like tiny profits magnified by enormous leverage - but with holding periods and market neutral positioning to reduce risk.

Mostly likely based on what grounds? Rentech was very profitable for decades before they ever started their public funds.

> Nobody knows how Rentech makes money

There's nothing extraordinarily special about Rentech's returns, they just employ short-term stat arb type strategies that require relatively little capital to execute, so if you express their returns as a percentage of invested capital you get an eye-popping number. But it's not comparable to the returns that a traditional buy-and-hold fund makes (in particular because those returns don't compound). There are plenty of other quant shops and prop trading firms that would make huge (>Rentech) annual returns if they attempted to phrase their earnings in those terms, but they typically don't, because if you don't need a lot of capital then you don't need to raise money from the clients and outside investors (can just trade the partners' money) and you don't need to brag about your returns in public.

The larger "public" funds are for strategies that can't be used at Medallion but still provide value for others.

> Noah Smith himself worked at SUNY Stony Brook, which is heavily funded by Simons.

Simons is very clear that his money is to go to the Math and some science departments. Noah is out of their Finance department.

Even if you take the senate report at face value...it only explains 1/5th % of their yearly returns.