What this covers
Sonali Basak, Chief Investment Strategist at iCapital, sits down with Cliff Asness, Co-Founder AQR Capital Management, to discuss why today's "expensive" and "bubble" are not the same thing, and why the difference matters more than most investors think. With pockets of froth across today's market, Asness explains why he sets a deliberately high bar before calling a bubble, and why profiting from a bubble you correctly identify is far harder than it looks. The conversation moves from there into the mechanics of staying disciplined when markets get extreme. Asness makes the case that quantitative investing is an extension of valuation discipline rather than a replacement for it, that value has evolved beyond simply buying cheap toward paying a fair price for quality, and that crowding has quietly become one of the biggest risks to any strategy as more capital chases the same trades. His core conviction is that the edge is not calling turning points, but in navigating dispersion, respecting the limits of prediction, and avoiding false precision.
Listen to The Bridge by iCapital: Apple Podcasts: https://podcasts.apple.com/us/podcast/the-bridge-by-icapital/id1896316886 Spotify: https://open.spotify.com/show/670eItrvT1LeLXRsBt7Zsb Amazon Music: https://music.amazon.com/podcasts/538ca236-fb6a-4046-a5c1-a595304a633a/the-bridge-by-icapital
Source description (no synthesized summary yet).
Asness argues that while current market valuations show dispersion at the 75th percentile historically—wider than normal but not at bubble extremes—diversified quant strategies can still generate attractive risk-adjusted returns by exploiting pricing inefficiencies across stocks, and that the evolution of quantitative investing toward more sophisticated factor models and machine learning represents an appropriate shift from simple factor-based approaches toward more holistic, behavior-resistant strategies.
- Value spread disparity is at 75th percentile, not the 95th+ that characterized the 2000 and 2020 peaks, making current conditions attractive but not bubble-level extreme
- Market concentration in mega-cap tech creates opportunity rather than constraint for diversified long-short portfolios that maintain industry and country neutrality
- Evolution of quant from simple momentum/value factors toward profitability, risk measures, and machine learning better captures the holistic value investing principles of Graham, Buffett, and Munger
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
A typical 60/40 portfolio of stocks and bonds is not truly diversified because stocks are about three times more volatile than bonds, so you are mostly diluting equity returns rather than achieving true diversification of returns.
“a 60/40 portfolio stocks are about three times more volatile than bonds. Bonds in their normal form...If you buy, say, the 10 year treasury, you're mostly diluting equity returns, not diversifying the direction of bonds.”
The value spread—the ratio of expensive to cheap stocks on valuation multiples—historically varied between 3x and 6x over 50 years, and peaked at 12-13x during the dot-com bubble, which enabled Asness to conclude numerically that prices were not justified by reasonable growth assumptions.
“for about 50 years, if you did that calculation, it varied between about three and six times...At the peak of the tech bubble, we saw that thing hit 12 or 13...at the peak of the tech bubble...you sit there and go...are there any numbers that could justify 13 times the spread? And the answer we came up with was, no.”
Underfitting (using overly simple models when the true underlying relationship is more complex) is a problem equally serious as overfitting, and modern machine learning better balances both risks by penalizing overfitting endogenously while allowing appropriate model complexity.
“underfitting is a problem on a par with overfitting...There's actual true complication out there and you're not picking it up 'cause you are using overly simple models...One thing that modern ML is go is better at than traditional statistics...it is better...balancing overfitting and underfitting, penalizing, um, endogenously to creating your model penalizing overfitting while allowing it to do some and some reasonable amount.”
Every investment strategy, particularly those known to more than two people, has a bad left tail (the risk of large losses), and the larger and more popular a strategy becomes, the worse its left tail becomes because the strategy's performance depends partly on the actions of similar investors.
“There's not a strategy in the world, particularly one that more than two people know about that doesn't have a bad left tail. And the bigger and the more popular, the worse the left tail because you don't control your own act...if someone very similar to you...panics and sells, you're going to suffer.”
Good risk control requires stress-testing portfolios against scenarios worse than the historical worst case; if your portfolio survives such stress tests, then even if conditions get worse than any historical precedent, losses should be manageable.
“we've looked at the worst cases in the past, you've said, what if it's substantially worse than that? That should hurt, right? If you say it's the worst case for my strategy and it's worse than that should hurt. But if it's survivable, you know, if you own equities, you should know that they can go down 20% in a day.”
Betting against bubbles is famously difficult; if you are 6-12 months early, you will make money over the full round-trip but endure an unpleasant 6-12 month period where the bubble inflates further before bursting.
“Betting against a bubble is famously difficult. If anyone gets it to the day, I'll be shocked. If you're six to 12 months early and you hold on, you're gonna make a lot of money round trip. But it's not a pleasant six to 12, right?”
Asness cannot fully interpret what individual dimensions in NLP vector representations mean, and even AQR's most expert ML researchers cannot easily explain what specific components of the vector contribute to the strategy, illustrating the limit of understanding that ML models require.
“if you ask me this, that vector of numbers that represents each earnings call, what does number, what does the seventh number mean? I'm gonna be hard pressed to figure that out for you. And if I turn to a far younger, far smarter A QR employee who's fully up on these things, they're gonna be hard pressed...you are giving up a little bit.”
Market corrections and retracements are healthy for long-term strategy sustainability because they reset investor entry points and prevent strategies that seem too good to be true from becoming so widely adopted that they're arbitraged into non-profitability.
“a lot of strategies that have worked for many years that I think will still work, would not work if they didn't have these periodic tough times. It doesn't create any kind of flushes out the market, Right? You need a new entry point for investors to get in...If it's at all publicly known...it doesn't last very long at all.”
A strategy with a 0.3 Sharpe ratio (nearly 3x better than the stock market's long-term Sharpe) that is publicly known won't last long before being arbitraged away, but a strategy with a 0.3 Sharpe will experience terrible short-term periods that scare away casual investors trying to use it.
“If something is a 0.3 sharp ratio, that's a little less than the stock market's been long term...it has some horrible short-term periods and just when everyone thinks I want this in my portfolio forever...it'll go through one of those periods and flush it out.”
Even when a market-neutral long/short portfolio is carefully balanced by industry and country, it is not riskless because individual stock bets can go wrong; the portfolio wins on average but experiences periods where 'the wrong stocks win.'
“it doesn't make it riskless. There are times where you choose what you think are all the right stocks, con country and industry neutral and, and the wrong stocks win.”
Market timing based on pure valuation is a 'suicidal quest'—Asness has attempted it and concluded it does not work reliably despite having identified expensive valuations historically.
“timing the market based on pure valuation is, I, I've, I've tried to do this. It's a, a suicidal quest.”
Market concentration does constrain traditional long-only managers because they cannot easily express negative views—not owning an overweight stock like Nvidia is a large implicit bet, while even the 100th largest S&P 500 company is still large enough that underweighting it moves the dial very little.
“Negative views are hard to express in a long only, all you could do is not own something. Um, if you have a negative view on Nvidia, you can express that. 'cause it's such a big weight in the index that, for instance, if you don't own it, that's a fairly big bet.”
Information edges about speed are the most arbitrageable of all edge types; even a small speed advantage will be competed away quickly, whereas behavioral bias-resistant signals about valuation and risk tend to last longer because they resist the human tendency to believe certain stories.
“information edges that are about speed, I think are the most arbitrageable out there. Things about valuation and risk where you're up against huge behavioral biases that people want to believe in...are harder. I'm a little faster than everyone else. It's not gonna last that long.”
Alternative data strategies typically show very high risk-adjusted returns when first deployed, but experience steady decline as the information becomes more widely known and less tradeable, eventually reaching zero or near-zero alpha.
“The lifecycle of alt data tends to be fairly high risk adjusted return when you're early to it...then a steady decrease. It may not go to zero. You may still want it in your model at a smaller weight or it may really go to to zero. The world has figured this out.”
Many quant hedge funds suffered losses in the 2007 quant crisis, but more levered multi-strat funds that added quant without deeply understanding it were forced out of business, whereas dedicated quant firms like AQR that understand their strategy survive such events.
“at some point people said, you know, we, we need some alliteration here...a lot of your peers were out of business by the end of that. Um, some of the more levered ones, I I think most common was, uh, multi-strat funds that really didn't understand quant. That added quant in a six, seven year bull market for quant.”
When AQR started in the early 1990s, the two major academic findings in quantitative investing were that low-multiple stocks outperform and momentum stocks outperform; the third finding (small-cap outperformance) was not believed by AQR, and low multiples paired with momentum formed the core early strategy.
“When we started a QR, you were looking for really two things...you were looking for low, multiple good momentum stocks. Um, those were the two major findings in academia. The third was that small stocks outperformed. We never believed in that one particularly.”
AQR and modern quant managers pride themselves on validating strategies against both live trading and back-tested results, and ensuring models make intuitive sense even when they don't fully understand every mathematical component, to guard against overfitting.
“we've always prided ourselves...being better than a back test, but both counting and does it make sense to us? Do we understand every drop of what it's doing and do we get the intuition?...the reason you do this part is if you only do what's worked in the past, you're just mining the data and you tend to grossly overfit.”
The quant factor model approach can explain a significant portion of famous active investors' returns by analyzing their factor exposures even though they pick individual stocks, showing that good stock picking typically aligns with exposure to systematic factors like profitability, risk, and value.
“if you use the quant factors simply to explain, if you have a style to what you do, even if you're picking individual stocks and owning a concentrated portfolio, if you're looking for reasonable multiples and profitable companies with low volatilities and low betas, you end up pretty correlated to the quant versions”
Benjamin Graham's value investing approach and the Graham-Dodd tradition have been misrepresented as seeking 'cigar butt' ultra-cheap stocks, when in fact Graham was always closer to seeking good companies at good (not dirt-cheap) prices, as Buffett and Munger emphasize.
“people will quote him and, and, and, and the, and the Graham and Dodd book...they'll make it sound like he's the deep value cigar butt destroyed company, but selling for pennies. And it was never that, it was always closer to what Buffet and Munger talk about.”
The August 2007 'quant crisis' or 'quant quake' lasted six days, during which the value spread moved from the 50th percentile to the 95th percentile, and while painful for AQR, it was survivable with no risk of ruin.
“The mother of all of them was the August of '07. Um, it was called the quant crisis for about 10 years...this was all of about six days. Um, for us it was very survivable...the value spread again, went from the 50th percentile at the beginning to the 95th over six days.”
Alternative data is ultimately a form of fundamental momentum—a way to measure changing business conditions faster than traditional metrics—and relates to the long-held quant insight that earnings revisions and surprises tend to continue.
“alt data by the way, is ultimately a form of fundamental momentum forever...simple measures of fundamental momentum that quants have looked at since maybe the eighties, at least the nineties, are earnings revisions and surprises where they've tended to continue...All data is just a way to try to be as quick as you can on, on those things.”
Warren Buffett experienced horrific three-year periods both in absolute and relative returns, and this demonstrates that even exceptionally skilled investors face extended stretches of underperformance, highlighting that markets are hard even for the best.
“Warren Buffett, when we studied him, he had horrific three year periods, both relative and absolute tend to be different periods. Um, the market, uh, tells you it's hard. It's, it's not impossible to beat.”
In August 1998, the stock market dropped about 20% over a month (the 'crash that nobody remembers'), which AQR survived profitably in their first month of operation by running a market-neutral strategy that was long and short similar amounts of stocks.
“In August of, of 98, Russia defaulted on its debt...the biggest thing was in August of, uh, of that year, stock market dropped about 20%...the crash that nobody remembers...the first month of our existence, the market drops like 20% and we're up.”
In bubble peaks, firms tend to sell significant amounts of their own shares because they recognize their valuations are high; near-term monitoring of insider selling will be an early warning signal for potential bubble conditions.
“you do see near the peaks of bubbles, firms selling a lot of their own shares...Keep an eye on, um, could happen. Doesn't mean it will happen...the true and tried like golden day investor used to look at the Wall Street Journal and look at the insiders stock sales too.”
Asness has identified two genuine bubbles in his career: the dot-com bubble peaking in March 2000 and a bubble peaking in October 2020, despite being a disciple of the Efficient Market Hypothesis through his dissertation committee chair Gene Fama.
“I have screamed 'Bubble!' twice in my career...I was a disciple of Gene Fama, the Efficient Market hypothesis. He, he co-chaired my dissertation committee...the.com or tech bubble, the end of the nineties, uh, peaked in March of 2000...the other peaked in say October of 2020.”
To create uncorrelated returns in a market-neutral strategy, the simplest approach is to be long and short approximately equal amounts of stocks, balanced by industry and country, which reduces exposure to market, country, and industry risk while allowing stock-picking skill to drive returns.
“if you want to be uncorrelated, start out by being long and short, a very similar amount of stocks...if you're long, a thousand stocks and short a thousand stocks...mainly balanced by industry and country...you can neutralize the market part. And in fact, a bunch of other parts, the country part, the industry part.”
AQR added capital to their positions near the bottom of the August 2007 quant crisis after hearing that Goldman Sachs had injected cash into their suffering long-short equity fund, identifying Goldman as the last weak hand to delever.
“we actually added a little bit to what we did at almost the exact bottom of that. Um, because we heard Goldman Sachs had had injected cash into their very, very suffering long short equity fund. And we thought they were the last kind of, uh, weak hand at the time to, to delever.”
AQR's versions of value have held up well since COVID, particularly because they are globally diversified (about half US) and maintain industry and country neutrality rather than taking a directional short in technology, which avoids the concentrated tech-short bet of traditional value indices.
“our versions of value since COVID have been pretty good...we're fully global, which amounts to being about half the US...being industry neutral, we're not sitting there. Short tech, we're long and short within industries.”
AQR has deliberately not disclosed some of its alternative data sources despite being asked by journalists and interviewers, because publicizing them would accelerate their arbitrage and reduce their value, a form of 'helping our own destruction.'
“I'd love to give you an example, but our, our heads of, of stock selection have asked me not to because it is proprietary...unlike some of the other things we do that we think is fairly hard to arb away, here, we'd be helping our own destruction by by doing that.”
Credit card receivables data was one of the first and most successful examples of alternative data, where collating public web data on credit card usage provided an early signal of retail company performance, particularly for identifying changes in spending patterns.
“The classic one about the only one I am willing to talk about, um, is credit card receivables. It was one of the venerable first ones...people built databases that were essentially, this is legal public data...The lifecycle of alt data tends to be fairly high risk adjusted return when you're early to it...for retailers knowing who in the last three months...has seen an increase in credit card usage...if you have that data a little faster than other people, that's an advantage.”
The next 18 months after August 1998 were the blow-off top of the dot-com bubble, and AQR experienced a two standard-deviation event in their first year and a half, which is not desirable to experience early in a fund's history.
“the next 18 months were the blow off top of the.com bubble...it wasn't as geeks would say, a 10 standard deviation event. It was a two standard deviation event. But you don't wanna have a two standard deviation event in your first year and a half in, in business.”
Jim Simons' Renaissance Medallion Fund performs dramatically better than AQR's strategies, but Medallion only accepts the partners' own money and does not take outside capital, so comparisons of 'which is better' are misleading—AQR's value is in accepting client capital.
“my favorite question...is the Medallion Fund better than you guys? And I go, oh hell yes. But I think we're very good in making your portfolio better and we will actually take your money and they won't.”
Alternative data consists of proprietary databases that did not exist before, often created by specialized firms to be sold to investors like AQR, and represents a new category of information sources beyond traditional financial statements and market data.
“Alternative data is exactly what it sounds like. It's, it's databases that didn't exist before that somebody often proprietary new firms sometimes designed just to create this database and to sell to people like AQR.”
One major use of machine learning at AQR is natural language processing (NLP), which analyzes earnings call transcripts to identify positive and negative signals faster and more nuanced than traditional word-counting methods.
“one of the big uses we use for ML is, uh, called natural language processing...where you take textual data...and you say, is this good or bad news? Fundamental momentum to us...the way quants did this for a trillion years was get the text to the call...add up good words and phrases and bad words and phrases.”
According to an AQR analysis, Warren Buffett's returns can be explained approximately 40% by profitability (low-risk, profitable companies), 40% by low-risk characteristics, and 20% by valuation multiples, showing that multiples are a tiebreaker rather than the primary driver.
“we explained a lot of his returns. This is very rough, but call it 40% looking for more profitable, uh, companies, 40% looking for lower risk companies and about 20% valuation.”
The quant industry has historically mislabeled the 'value factor' (low multiples) when it should be called something broader, because true value investing is holistic and a high-multiple company can be valuable if it has sufficient growth potential, making low multiples a useful signal but not the definition of value.
“I don't think the quant should have called low multiples. The value factor, it's not. Value is more holistic...You could have a high multiple company that is value if you think it's gonna grow enough...low multiple still works. because on average people just go too far.”
Quant strategies are valuable partly because they experience their worst periods when the market is doing well or in bubbles, creating an 'accidental hero' dynamic where quants make money after losing it during the bubble peak, providing portfolio diversification when most needed.
“when it lost money, everything else was going straight up in a bubble. And when it made more than all of it back was when people needed it the most. And that's not our goal. And those funds, our goal is to be market neutral. We were accidental heroes on that.”
Asness defines a bubble as prices that are 'beyond the pale' and require that he try to construct assumptions that aren't ridiculous in order to justify them; this is a subjective framework but it prevents him from seeing bubbles everywhere.
“a bubble to me has to be beyond the pale prices...can I come up with assumptions that aren't ridiculous? With ridiculous, again, being a somewhat subjective word that could justify these prices.”
Modern quant models incorporate DCF concepts implicitly through looking at current price, growth expectations, and discount rates, making them a grand bet that DCF works across a diversified portfolio rather than for any individual company.
“It's a grand bet that a DCF works. It's not a bet that a DCF will get any one company. Right?...modern quant models look more, not entirely, but look more like a very diversified version.”
There are two kinds of crowding: strategy-level crowding where capital compresses expected returns over the long term, and short-term crowding where many similar investors panic-selling creates temporary dislocations; both exist but operate on different timescales.
“I separate crowding. Uh, and this is too binary. There's all points in between. But to just think about it, it makes it simpler. Uh, two kinds of crowding. One...a strategy with too much capital in it is crunching down...more long term...crowding is more typically talked about in the short term, the risk of a very bad short term event.”
Retail investors and institutional quant investors tend to be on opposite sides of bets: retail tends to chase expensive, low-quality stocks while quants short them, so net capital flows across the market are not necessarily adverse to quant strategies.
“the retail world, I think we're usually on the other side of them. I think they're usually on the faddish stuff and that's become very popular. So yes, some parts of the market can grow, but it's the net that matters.”
In 2019, AQR wrote a piece recommending a 'venial value sin'—a small position increase into value investing at historically extreme valuations—because even if conditions get worse before improving, owning some value at extremes is warranted.
“I wrote a piece where we've said market and, and timing factors like value is an investing sin...We say it in a joking manner, but we're serious. We say we recommend you sin a little, um, at true never before extremes...We wrote a piece saying it's time for a venial value sin. A we got a little Catholic there, you know, the more minor sin, not a cardinal sin.”
Giving companies credit for genius cost of leverage when they simply execute what they thought of is dismissive; Warren Buffett did this for 40 years successfully, which deserves credit, not dismissal.
“everyone says that kind of dismissively like, ah, that's all he did. And I'm like, yeah, he did it and nobody else thought of it. And he did it for like 40 years. Um, you get credit Yeah. For that.”
AQR has historically thrown out more alternative data sources than it has kept, and core factor approaches are less dynamic in their updates, suggesting that alternative data is inherently more experimental and less reliable long-term than traditional factor models.
“We've probably thrown out more from alt data than we have from the core factor stuff over our whole history. Um, and measuring it when you alter a core factor and measured a little better, have you thrown it out? I don't know. But there's certainly a lot more dynamic change a few years later. Alt data has to be refreshed and replaced with new sources.”
Comparing tech companies to textiles on valuation multiples is inappropriate because industries have different structural growth prospects that justify different multiples, and AQR avoids making concentrated industry bets despite the dispersion in valuations.
“we think comparing tech to, I'll make up an extreme to textiles on valuation multiples is a little crazy. There are industries that will grow for a long time different than each other and should sell at different multiples.”
Some valuation factors may work primarily in the extreme deciles (cheapest and most expensive stocks) but be flat in the middle, and some factors may work on the short side but provide less information on the long side, complexity that traditional linear models cannot capture.
“Some factors might work in the tails, call it the top decile and the bottom decile, but be kind of flat, not give you a lot of information. Some might work on the short side, but not provide a lot of information once you get into the reasonable, uh, area.”
If there is extreme market concentration that reflects a bubble-like attitude where value is thrown out the window and only momentum works, that can affect which strategies work and doesn't work; however, concentration itself has not been problematic for AQR's performance to date.
“if it's indicative of a bubble...that can affect what works and doesn't work...it might be coincident with things...If you tell me the market got super concentrated and, and people threw value out the window and the only thing that worked was momentum, I think that's part of the story...it has not bugged us to date.”
Signs that typically precede bubbles include massive equity issuance, and we have not seen that recently—only recently beginning with SpaceX IPO after a years-long dearth of IPO activity, which is not yet at bubble-level issuance.
“one of the major ones on the, on the list is issuance in, in the other major bubbles we've seen, we've seen pretty massive issuance just starting to ramp that up. Obviously the SpaceX IPO was a pretty big fricking deal. Prior to that, we had seen a dearth of IPOs for a number of years.”
Asness has intentionally slowed AQR's transition to machine learning to maintain the firm's historical commitment to understanding models intuitively, representing a philosophical choice to trade some short-term returns for long-term security and decision-quality.
“I often say I slowed us down a little on, on moving to ML Really...I think it is the old man at the firm's job when you had something that's worked for 20, 25 years and you want to have a philosophical change, even if it's a modest one...any change to a philosophy you've stuck with and talked about should be done slowly and carefully...I do think I probably cost us a little money by going slow.”
NLP represents earnings calls as vectors of numbers (where the meaning of each number is often opaque), and the strategy created from these vectors correlates reasonably well with other fundamental momentum measures, providing reassurance that NLP is measuring what it is supposed to measure.
“NLP actually does is represent each earnings call as what's called a vector of a whole bunch of numbers, and then do empirics on those numbers...The intuition at the end, you create a strategy that is actually fairly decently correlated with other ways to measure fundamental momentum. We just think it's better...The correlation gives us great comfort 'cause it means we are measuring what we're trying to measure.”
The value spread during the COVID crash and recovery reached a new hundredth percentile, exceeding the dot-com bubble peak by approximately 25%, though by the time of this interview it has narrowed to the 75th percentile historically.
“that disparity, which hit a new hundredth percentile during COVID surpassed the dot com...The widest ever...the hundred 25th percentile...it was about 25% more than the prior one...We're about 75th percentile now on our measures.”
At the 75th percentile on the value spread, Asness does not believe conditions warrant calling a bubble; instead, this represents an attractive time to own value and other diversifying assets, but not an extreme requiring screaming about a bubble.
“I don't think I should scream bubble at the 75th percentile for me. Uh, 75th percentile makes it maybe a better time than normal to own some of this in your portfolio...it's not screamingly crazy.”
The most important lesson from AQR's first year is 'Never high five in this business' because overconfidence after early successes invites subsequent losses; one should only high five once when fully retired from investing.
“Never high five in this business. You get to high five once when you retire and you're like all in safe assets and you're not doing this anymore. You can high five once.”
A junior AQR employee challenged Asness' concern about losing intuition in ML models by pointing out that if something is fully intuitive and obvious, then ML providing better results means it's doing something not obvious—if the improvement weren't from hidden complexity, there would be nothing special for ML to do.
“a fairly junior person said something like, if it was all super intuitive and obvious, what do you think the machine learning is doing that's special...If there's not a part that's at least hard. And uh, one of the big uses we use for ML is, uh, called natural language processing.”
Asness anticipates that AQR's successor may eventually exhaust the returns of a strategy that has worked for decades, inheriting a situation where the accumulated capital flows have arbitraged away the edge that the current generation benefited from.
“you know, the good news is you've won the AQR Game of Thrones. You're, you're in charge. The bad news is I used up all the return forever. Could happen.”