NO CRYING IN THE CASINO

Quitting trend following at the worst time: AQR's Managing Director Peter Hecht "In 08 it actually did well. So that's actually why everyone got in after 08." "If trend following is based on underreaction to news, if there isn't much news, there is very little to underreact to." "A lot of people got out of trend following after holding it for 10 years and it was sort of mediocre, and they got out in 2021, and in 2022 the industry was up, let's call it 30% or 40%." "Sometimes holding it with equity beta is a way to force discipline." "If you're just holding managed futures standalone and it's just having two, 3% returns, you're sort of like, ah, is that a good use of my capital?"

Slugging vs. hit rate: Balyasny's founder Dmitry Balyasny "If you contrast it with high frequency trading, which is a hit rate type of business, you're going to have 99% of your trades going to be profitable or a tiny loss. You've got to do tiny, tiny, tiny, and repeat, repeat, repeat." "We have 170 investing teams at the moment. When you hire an investing team, chances are, depending on their track record, maybe 75% of them will wind up working out." "If you can control the risk, you might lose 10 million, 20 million, 30 million on somebody who doesn't work out, but the ones that do work out, they'll hopefully be with you for 10, 20 plus years, and you might make hundreds and hundreds of millions off of them." "In order to enable the slugging, you have to have very well-defined risk boxes within which people will operate, to enable them to bet on the things that they're really good at betting on and try to exclude as much of the other stuff as possible." "For every strategy, we'll have stops, vol targets, vol limits, stress limits, liquidity limits, et cetera."

Answer every question before it's asked: Citadel Securities' Technology COO Jeff Maurone "For 50 years, the process of sequencing amino acids into folded proteins was a bottleneck on biology research." "Researchers would spend quarters, sometimes years, folding just a single protein until Demis Hassabis and the team at DeepMind fundamentally changed this." "Using models, they reduced that process from years to call it six or seven weeks." "Even though they were using a model to accelerate the process, it still worked first in, first out." "Instead of working that way, why not just fold every single protein we've ever seen?" "More than 200 million proteins folded, became an incredible resource that changed biology forever." "Using what is essentially an early form of agents, rather than answer questions when they arise, but to preemptively answer every possible question before it is asked, is one of the real emerging ways that we are seeing AI and agents being used to achieve incredible results."

AI and the holding-horizon problem: Balyasny's founder Dmitry Balyasny "I don't think you can replace any for the foreseeable future, but I think it's a really good adjunct." "Markets are quite different. And the longer your holding horizons, the more different they are." "For that stuff, you can use AI to really automate your trading because the rules are quite clear, like what causes a stock to go up in the next 30 seconds? Well, it's more buyers and sellers." "But over a longer period of time, is this employment number that comes in hot good or bad for equities? Well, that depends on the environment that you're in." "When a company misses their earnings number, but the revenues are better, is that good or bad? Well, it depends."

The 55% hit rate: AQR's Managing Director Peter Hecht "In stock selection, the way it ends up working on average is, let's say, if we're right about 55 percent of the time, we hit our target return." "So that means 45 percent of the time we're wrong." "Let's say one year we were 52 percent right and we hit our target return." "Let's say we're having a bad year and you must be like, well, instead of 52%, you must be at 37%. No, we're at 49%." "The difference between hitting your target return or having a good year versus a bad year, it's small, right? It's a small edge." "It's not like we're right 80% of the time. It's more like low 50s, mid 50s for our stock."

Small edge, high reps: Balyasny's founder Dmitry Balyasny "A simple way for folks to think about it is like a casino. The edge that a casino has is relatively small." "When you have millions of people coming through the casino, their earnings are very consistent, because they have a little bit of an edge, and any one player might win, which keeps people coming back." "You need high enough reps to put the statistics in your favor." "How do you optimize your trading efficiency to lower the slippage, your cost of moving the stock when you're buying or selling, execution costs?" "You still might only have a 55% chance of being right, but you're not gonna lose that much if you're wrong and you'll make a lot if you're right, because your view is sufficiently variant." "If it does happen, it's a big surprise to the market and the security gets repriced."

The out-of-sample gauntlet: AQR's Managing Director Peter Hecht "Start off with an economic thesis on why this signal should predict returns." "Let's say you first test your hypothesis in U.S. large caps and it works. Then we're going to say, all right, but your theory also says it should work in U.S. small caps. Let's test it there." "It should work in European markets. It should work in Japan. Let's go grab European markets and test it there and in Japan and Australia. It should work in emerging markets." "Some of our signals, if they predict a single name stock, maybe they should also predict country level indices like the S&P 500 versus the FTSE versus the Nikkei." "So we're always pushing the data with all of this out of sample verification, even though we're starting with sort of an economic foundation, to try to minimize the chance that we're just finding patterns in randomness."

The three traits that scale a PM: Balyasny's founder Dmitry Balyasny "The things I look for are how does the person respond to adversity. Because if they had an amazing track record before they got to you, you can be certain they'll have a drawdown once they're here." "You need people who have the perseverance and mental fortitude to use the difficult periods to re-underwrite what they're doing, come up with some fresh ideas, improve their process, improve their team, as opposed to pack it in." "A real learning, adaptive type of mindset, where they're constantly looking for better ways to do things and aren't overly wedded to a particular trade style approach." "The folks that are really scaled over time are ones that have a business building DNA." "They enjoy hiring people, they enjoy managing people, they enjoy running their own business within a business, and those kind of folks do a better job of scaling over time."

The airline problem for GPUs: Citadel's CEO Ken Griffin "GPUs, like airlines, have a very low marginal cost to run. Like the incremental seat on an airline that's not full, if you can sell that ticket for $20, you're making money in that seat." "That's one of the reasons that the airline industry is such a nightmare: the moment you have overcapacity, pricing goes into freefall." "We saw this actually with the telecom build out back in 2000." "We built more fiber than we had demand for, and the price of telecommunications just went straight down." "If we reach a point where the provisioning of GPUs exceeds demand for GPUs, marginal pricing will actually go into free fall. And that really upends all the investment assumptions that people are working with today." "Right here, right now, there's people that are earning extraordinary pricing on GPUs that are five or six years old. That probably doesn't persist for another five or six years."

Prompts vs. embeddings: AQR's Managing Director Peter Hecht "You literally ask ChatGPT or Claude, is Apple cheap or expensive? Give me your analysis and give me the details why you think it's cheap or expensive. That's prompt, okay?" "Then there's a way of using a large language model where you feed in documents, financial documents, broker reports, earnings reports, and you go under the hood of the large language model and you take what we call a word embedding, which is a numerical representation of the text." "When we use large language models, we're doing the latter. We're not doing the prompt-based approach." "All of the special sauce is not in the large language model. The fact that you can get a numerical representation of the data is sort of a commodity." "The hardest part is can you find whether that numerical representation of the data can predict future returns." "That is very difficult to do, requires a ton of skill and it's highly subjective. And that's why I don't think there will be convergence."