Investing - Theory, News & General • The Day The Factors Died

As a former Boglehead who is deeply involved in using factors to manage my personal assets now, I actually think many people here are missing the point. That's not surprising. If a lot of people understood what the factors were all about, it probably wouldn't be as effective indeed.

There is a very simple principle that factors do work. And it is not someone else who illustrates this, but A.Y. Chen himself. He proved in another paper that accounting factors found purely by data mining do work out of sample.

So why did he find the opposite in this paper?

Quite simply, he not only overestimates transaction costs, but more importantly, he picks a time period that includes two major "factor bear markets": the collapse of the Momentum factor in 2009 and the collapse of the Value factor in 2018-2020. In addition, he didn't include the crazy bull market in factors from 2021 to the present. If you account for the latter, you'll find that value + momentum still works in the US.

In addition, Value + Momentum in Europe and Asia is still profitable even for the 2005-2020 period. Further, while factors have underperformed in the US, the US stock market has better liquidity, which partially makes up for it.

That said, there are still some opportunities if you are willing to design your own system to trade. But if you're going to do this by choosing long-only funds, perhaps this will be more difficult.

Right. So it works, when it works, where it works. If you choose the right factors, at the right times, in the right markets, using the right funds. If it works for you you look/feel like a genius. If it doesn't work for you, factor people will tell you that you picked the wrong funds, the wrong times, or didn't wait long enough. That you just didn't do it right. Small was great once upon a time, but maybe not so much anymore. Value maybe still there if you look international, but decayed. Momentum sounds super, but hard to identify, let alone capture. It is, by definition an active strategy with high turnover and trading costs. Maybe that's why its there.

How do you know which ones will derive a premium in the next era? You're always looking at the past, either recent or remote.

If you can long-short maybe you can capture more of the premium than most long-only investors, but then you're deep into active management territory and really have to figure out how to manage costs. How much does one pay to short-sell when SOFR is over 5% (I'm not being sarcastic, I have no idea, would love to know).

Anyway, congratulations on all your success, sounds like you're doing very well with your strategy. But I have a day job.

Regarding Chen's work:

Maybe he overestimated trading costs. I'm not qualified to answer that. Maybe you can minimize your own costs, maybe you need to pick the right fund to minimize costs for you. Also, maybe when interest rates are low, trading costs are low. When interest rates are higher, maybe not so much. Maybe maybe maybe. Again, I have a day job and a pretty boring 401k and brokerage account. I'm fairly confident that I'm a small boat in a big ocean and the lowest transaction costs probably go to the bigger ships.

And saying "he didn't include the crazy bull market in factors from 2021 to the present." Is not really accurate. His paper discussing this stuff was published in 2020. It takes months, sometimes years to write a paper. There is always a lag in data before publication. You can't compare formal research findings to running a quick backtest in portfolio visualizer or whatever.

While shorting gets you a higher rebate during periods of higher interest rates, I don't use shorting to get a factor premium because the borrowing cost (for borrowing shares) of shorting is too high and shorting is also too risky to manage (if you're not just doing it to avoid taxes).

Yes that's my point. You mentioned that investors using long-only funds might have trouble. I assumed that meant you short sell. Apparently not. You have to borrow to short sell. You have to pay interest to borrow. It costs money. It costs more money when rates are high.

Momentum has been doing well in Europe and India, but not in the U.S., while East Asia is known for reversals. I'm not quite sure what you mean by momentum being difficult to capture though.

These two sentences are internally inconsistent. Momentum has been doing well in one place and time, but not another. So how do I capture a premium from it if I can't determine where and when it works? That's exactly what I mean be it being difficult to capture.

Regarding Chen's works, many of his papers point out that factors are unlikely to be the product of data mining alone as you said (I also doubted factors as a Bogleheads before I read enough papers):

The Limits of P-Hacking: A Thought Experiment

https://papers.ssrn.com/sol3/papers.cfm ... id=3358905

Most claimed statistical findings in cross-sectional return predictability are likely true

https://arxiv.org/pdf/2206.15365

And the paper I mentioned before:

High-Throughput Asset Pricing

https://arxiv.org/pdf/2311.10685

We use empirical Bayes (EB) to mine for out-of-sample returns among 73,108 long-short strategies constructed from accounting ratios, past returns, and ticker symbols. EB predicts returns are concentrated in accounting and past return strategies, small stocks, and pre-2004 samples. The cross-section of out-of-sample return lines up closely with EB predictions. Data-mined portfolios have mean returns comparable with published portfolios, but the data-mined returns are arguably free of data mining bias

Yes, Chen's work and many others have shown that the distribution of past returns has shown factor premia. And many of those factors or "anomalies" persist out of sample. I do not doubt that. Few who are academically inclined would. But that's not his or my point. His point is that the premia has decayed. My point is that even if the premia is there, the average retail investor is unlikely to persistently capture it.

His points about data-mining are interesting. The practice is not favored among academics in my field (not economics) because it lacks explanatory power. I know all about "p-hacking". I'm pretty good at it when I want to be. Chen's arguments about it being more useful than previously acknowledged are interesting. But there is some survivor bias there. By that i mean that you can "data mine" a predictive model and have it work really well. Then you can look back, pat yourself on the back and claim victory. But you can also do it and have it fail. The world only remembers the or remarks on the models that in retrospect are successful.

And hat's all fine, but it wasn't the point you made about Chen. You said that his work did not take into account the bull market in factors since 2021. His paper and on that subject predate that bull market. The 2024 paper you cite does not address that topic. If you think he should update his studies, send him an email and see if he responds.

There's another problem with that Chen article, he uses returns instead of alpha. Since most of the factors are negatively exposed to market factors, this causes them to underestimate the alpha that can be gained.

No. "factors" are an academic construct. The single factor long-short portfolios that FF generated, for instance, were explicitly designed to isolate the effects of market on factor estimates. They also explictly exclude trading costs and interest rates. The whole point of the studies was to isolate factor effects from other effects. That's one of the reasons these studies were designed the way they were, as opposed to portfolios that real investors could actually hold.

Return is another story. In the real world investors are subject to market risk, interest rate effects, etc. I think that was Chen's point. Once you introduce those elements, premia decay. You can argue with his methodology. But that's the point he is making.

At last, an individual investor will actually have lower transaction costs because you are trading less. The reason institutional investors need all sorts of weapons to lower transaction costs is because they are whales.

Edit: Supplement the reply

I don't know. I have to trust you on that. My impression is that regular joe with a fidelity account probably isn't winning the bid ask spread by writing market orders when competing with large investors and high frequency computer algorithms, but maybe you're right.

Short sellers actually gain interest, not the other way around. The main problem is that attractive short selling opportunities have higher stock borrowing costs and short squeeze risks. I just wouldn't try it. Maybe you should know the difference between stock borrowing costs and rebates. The latter increases when the interest rate increases.

East Asia's momentum factor underperforms mainly because their value factor is always too strong and they are negatively correlated. When you strip the value factor from price data, the remaining returns will have momentum. Of course, it's true that you can't be sure that a certain type of factor in a place will perform well over time, which is why it's necessary to diversify and stay the course as you do for market factor.

You're just repeating your argument that you don't think data mining can be useful, but hopefully you'll read the paper before responding to what he mentions. In fact, Chen finds that factors with more and "better" explanations perform worse out-of-sample. The machine learning community has also found that the pursuit of explanatory power impairs out-of-sample predictive power. Perhaps what we need is to embrace the intelligence of machines rather than hoping we can explain their results. Bad explanations would hurt your performance. They are even worse than nothing.

As for your later statement that the factors are constructed, you don't seem to understand what I'm talking about, because I didn't say that the factors aren't constructed. The point here is that longs on factors tend to have lower beta and shorts on factors tend to have higher beta, so it is unfair to calculate the role of factors solely in terms of the difference between long and short returns because the consideration of market premium considerations would actually increase excess returns. In addition, the interest rate story actually favors factor investors, which allows them to explain 2009-2020, but I think the story lacks foundation.

In fact, the premium "decayed" less than I thought it would, a difference that was not at all unexpected considering that 2005-2020 included two major factor bear markets, while 1984-2000 was with a period of particularly good factor performance because many retail investors come in and gamble their money.

Also, in order to get good trade execution, you should not (only) use market orders. However, the fact is that if you use robinhood, maybe the market maker will think you are a WSB YOLO APE and give you good execution. I never tried it. After all, you don't actually have to win over high frequency market makers, you can't. You just need to get less bad trading costs.

Edit: Maybe you should know HFT companies and hedge funds are different. Just search VIRT. It is a public-listed HFT company. You can't beat HFT, but you can beat hedge funds in the aspect of trading costs

Edit2: Of course, I understand perfectly well that A.Y. Chen knows about the factor bull market after 2021, which is well known in the industry, just like the momentum crash of '09 and the value super-bear market of '18-20, so I don't see the need to email him. It's just a basic fact, and as you say, maybe he didn't have time to take that into account when he was writing that paper.

However, in any case, it's clear that you don't have any valid counterarguments to my original argument in this regard. Obviously, even if the paper was written at an earlier time and therefore may have limitations, that doesn't mean we can't talk about it with the addition of later circumstances to point out its limitations.

Statistics: Posted by folkher0 — Sat Aug 10, 2024 3:05 pm — Replies 208 — Views 16419