I do independent research on EA topics. I write about whatever seems important, tractable, and interesting (to me). Lately, I mainly write about EA investing strategy, but my attention span is too short to pick just one topic.
I have a website: https://mdickens.me/ Most of the content on my website gets cross-posted to the EA Forum.
My favorite things that I've written: https://mdickens.me/favorite-posts/
I used to work as a software developer at Affirm.
I'm conflicted on this: on the one hand I agree that it's worth listening to people who aren't skilled at politeness or aren't putting enough effort into it. On the other hand, I think someone like Sapphire is capable of communicating the same information in a more polite way, and a ban incentivizes people to put more effort into politeness, which will make the community nicer.
I am not clear from your explanation on whether health impacts are talking about the effect on the mother or the effect on the stillborn child. If you are considering the effect on the stillborn child, it seems that you should consider increasing reproduction as approximately as good as decreasing stillbirths.
it seems crazy to imagine a baby dying during labour as anything other than a rich, full potential liife lost, but if we extend that logic too far backwards then we might imagine any moment that we are not reproducing to be costing one “life’s worth” of DALYs. Time discounting is a difficult science
It seems to me that you have to accept one of these or the other. Treating a failed conception as ~0 DALYs but a stillbirth as ~5 DALYs doesn't make any sense to me. Either they should both be ~0, or they should both be ~70. But if the 5 DALYs number is the effect on the mother, then that's fair. (Although I'm guessing the effect on the mother would be much smaller than that?)
(I have no comment on the rest of the investigation as it's somewhat technical and outside of my expertise. It looks like you're looking at the right questions, at least.)
A couple of questions related to this, not directly relevant but I've been wondering about them and you might know something:
Your recent comment got me thinking more about this. Basically, I didn't think the last few years of underperformance was good evidence against factor investing, but I wasn't sure how to explain why. After thinking for a while, I think I have a better handle on it. You're (probably) smarter than me and you're better at statistics than I am, so I was kind of bothered at myself for not being able to convince you earlier, and I think your argument was better than mine. I want to try to come up with something more coherent.
Most of this comment is about estimating future expected returns, but one thing I will say first: you are right that my original estimate of standard deviation, where I took the historical standard deviation was too optimistic. Originally I looked at vol over the last ~25 years and assumed it would stay the same. But I just looked at the performance of value and momentum back to 1926 and I noticed the last few decades have had a standard deviation 2–3 percentage points lower than through most of history. Higher standard deviation decreases optimal leverage. You were correct that I was relying too much on a short sample. (The return for a value/momentum/trend strategy back to 1926 is about the same as the return 1991–2017.)
(I believe in my head I was thinking, "historical volatility is a good predictor of future volatility, so I can just take this sample and use that to predict future volatility." Which is somewhat true but not good enough because volatility isn't that stable over time.)
Regarding future returns:
Between "nothing beats the market" on one side and "it's reasonable to expect a median +X% return over the market for this particular fund or similar funds" on the other side, I can see a few intermediate steps:
This comment focuses on #2, and touches on #4. I'm going to focus on value investing and not momentum because it's much easier to reason about. My argument for momentum investing could be summarized as, "if momentum investing didn't work for reason X, then reason X would also be visible in the value factor, but no such reason shows up in the value factor data." Basically I can't see any reason why the momentum factor would stop working but not the value factor.
So, value has performed poorly recently. What is that evidence of? That depends on why it performed poorly.
There are basically two hypotheses about why value might stop working:
For now let's define 'value' as buying stocks with low P/E ratios and shorting stocks with high P/Es.
It is useful to break down a stock's price as
and a stock's price return as
Call the multiple expansion and the structural return (and roll the dividend yield into the structural return).[5]
(Note: When the P/E multiples for value and growth converge, that's multiple contraction for the value factor, and is good for value. When they diverge, that's multiple expansion, and is good for growth.)
Then we can divide the stock market into a value stock bucket vs. a growth stock bucket and look at the differential structural return and multiple contraction between the two buckets.
(This is basically the same methodology as Research Affiliates' Reports of Value's Death May Be Greatly Exaggerated.[6][7])
If value became too popular, we should see the value spread narrowing as money floods out of growth and into value (and value probably outperforming while that's happening) and then sitting at a narrow spread for a long time. That's empirically not what happened—the recent underperformance is driven by a widening spread (I will give some numbers below). So I think this hypothesis is almost certainly false.
If value stops working because high-P/E companies start systematically outperforming low P/E companies[4], we should see a negative structural return. I find this hypothesis much more plausible.
If value underperforms because of a negative structural return, that suggests something is going on with value companies that's making them less successful, and it provides evidence that value investing won't work in the future.
I looked at some numbers using the Ken French data library. Value investing has had a poor streak since about 2007, so let's look at the data from before and after 2007. From 1952 to 2006, the value factor (= long low P/E, short high P/E) had a return of 5.7% = <0.5% multiple contraction + 6.2% structural return. It's had lots of multiple expansions and contractions over that period, but they all pretty much canceled out, so most of the long-run return was structural.
From 2007 to 2022, the value factor returned –1.2% = –5.3% multiple contraction + 4.1% structural return. This shows that the poor return was driven by multiple expansion—if the multiple hadn't changed, value would have beaten growth by 4 percentage points. My biggest concern here is that the structural return dropped by over 2 percentage points relative to history, which means value companies' fundamentals did worse 2007–2022, although the value factor still had a positive structural return. Hard to say whether the lower structural return will persist, but it's fluctuated a lot in the past and it was lower in the 1930s than it was in the 2010s, so it could be bad luck[1]. A 4.1% structural return still suggests that, if multiple expansion evens out, value should perform pretty well. There's some possible concern that the structural return will continue to go down, which I'll talk about later.
For the last few years: I can't look at the exact period since I wrote the original article, since the data only gives annual P/E updates. So I calculated value factor returns for 2019–2022 and 2020–2022:
2019–2022: –1.1% return = –12.2% multiple contraction + 11.1% structural return 2020–2022: 10.8% return = –20.1% multiple contraction + 30.8% structural return[2]
(Note: My data set treats years as ending in June, and the value factor performed badly since June 2022, which is part of why these numbers are higher the return for VMOT you quoted (the other big reason being that VMOT combines value investing with other strategies, and the way it implements value isn't the same as the way I'm testing it).)
Even though value underperformed over the last 3 years, IMO the recent few years look pretty optimistic for value investing because value had a strong structural return, and the poor performance is fully explained by historically fast multiple expansion (comparable to the 1998–1999 multiple expansion at the peak of the dot-com boom). I think the only way you could reasonably explain this much multiple expansion is by saying (1) AGI is imminent, (2) AGI is going to revolutionize the economy, but in a specific way that benefits stockholders, and (3) it's specifically going to benefit the stockholders in growth companies but not value companies. (Which is definitely in my hypothesis space, but it's a pretty conjunctive hypothesis so I don't think it's all that likely.[3])
(I repeated this analysis using Price/Cash Flow and Price/Book instead of Price/Earnings and they gave pretty similar results.)
What does all this say about the prospective expected return of the value factor? Hard to say because you could make a lot of different assumptions. Like, will the valuation multiple revert to the historical mean? How long will that take? Will it overshoot the historical mean and contract to the point of being narrower than average (like what happened in 2000–2007)? Maybe it will never fully revert? What will happen to the structural return? Will it look like the historical average? The recent average? Will it go lower? Will it be negative?
For the multiple to revert to the historical average, value would have to return about an extra 70% in total. If it takes 20 years, that's 2.7% return per year due to multiple contraction. If it only goes halfway back to the historical average in 20 years, that's 1.5% per year. (Or eg it could fully revert in only 10 years, which would be a 5% annual return.) Then for the structural return, I'd say a reasonable median estimate is somewhere around 3–6%, somewhere between the historical average and a bit below the more recent average. If we say 2% multiple contraction + 3% structural return - 2% overhead costs, that's a 3% return for the value factor. There's a wide range of numbers you could justify:
These both sound at least weakly plausible to me. (9% is more like the "full inside view" number...probably much too high, but there's an argument for it.)
Since we were talking about VMOT, not the academic value factor, VMOT differs in a few ways:
The value component of VMOT should still be fairly well correlated with the value factor as I defined it above. For the momentum/trend factors, there's no way to break down the returns like I did for the value factor, so it's much harder to say anything useful. Maybe the most useful thing to look at is whether momentum appears oversubscribed, and it probably doesn't, although it's hard to say with confidence.
VMOT (or a similar fund) won't perform the same as an academic-style value/momentum/trend factor portfolio:
On net, I'd expect VMOT's implementation of value to perform about as well as the traditional value factor.
I should also mention that, while VMOT is my favorite factor ETF, the difference between VMOT and the 2nd best ETF is probably small. (Not sure what the 2nd best would be, maybe [QRPIX](https://www.etf.com/TRTY][TRTY]]? or [[https://funds.aqr.com/funds/aqr-alternative-risk-premia-fund) looks really nice but it's not an ETF.) Most "factor" ETFs don't actually have much exposure to academic-style factors (I think the biggest US value and momentum ETFs are VTV and MTUM, respectively, which both have pathetic factor loadings), but there are a few good ones out there.
Some people (like Dan Rasmussen) say that, even though value has been underperforming, the value spread is widening, so this is actually evidence that value will perform better in the future. I'm not sure I would go that far, but let's go with that claim for a minute. Is it a fair claim? I find it suspicious because, on its face, you are taking a streak of bad performance as evidence that value works. But presumably you'd take good performance as evidence that value works, right? So any outcome is treated as confirming evidence.
I thought about this a bit, and I don't think it's true that I'd always take good performance as evidence that value works. Consider four possible worlds:
(#2 and #3 could have either positive or negative performance for the value factor, depending on which force wins out.)
I would rank the forward-looking prospects of the value factor for these worlds as #3 > #1 > #4 > #2. So the best world (#3) is one where value may be underperforming due to multiple expansion (as we saw in the last decade and a half, and during the dot-com bubble), and the worst world (#2) is one where value may be outperforming due to multiple contraction (for the record, the biggest times where this happened in the US were a ~15% structural drawdown 1994–1995 and a ~45% (!) structural drawdown 2002–2004, where both times the drawdown was basically canceled out by a multiple contraction. now that I'm looking at it, the big 2002–2004 structural drawdown + contraction probably explains a good chunk of value's 2007+ underperformance).
That said, I don't think that's the whole story. I'm more pessimistic than some people because, by some measures, the value factor has had a worse structural return since 2007 than it did historically. It was still positive, and it's not unusual for the structural return to fluctuate (eg it had a big drawdown starting in 1934 and didn't recover until ~1970), and this depends on how exactly you measure it and what time period you look at, so I'm not too concerned about it, but I wouldn't rule out the possibility that value could have a still-lower structural return in the future. And I'm not super confident in the argument that the valuation spread is historically wide so it must narrow—it might get even wider. (In the 80's, Japan's market got to the highest valuation ever seen in history, and then the valuation doubled again before reaching the peak.) It can't keep getting wider forever, but you don't know how far it will go or how long it will take to revert.
In a simple two-hypothesis universe, what is the odds ratio for "value/momentum/trend works" vs "doesn't work" based on out-of-sample performance?
I made the following assumptions:
(I'm talking about VMOT but similar logic should work for any value/momentum/trend strategy, I just happen to like VMOT the best)
To calculate odds ratios, we need to know the mean and standard deviation of VMOT – Market. Using the above numbers, NH gives mean -2%, stdev 7%. AH gives mean +3%, stdev 15%. (Possible that I made a math error)
The lower volatility for NH means that sufficiently bad underperformance is evidence in favor of AH, which seems counterintuitive. This would follow from the fact that, under NH, VMOT has to have beta close to 1, so it's very unlikely to out- or under-perform by a lot. To avoid this counterintuitive result, we could say value and momentum are real risk factors that drive returns independently of market beta, but that they have zero (gross) expected return. So say NH is that relative return has mean -2% and stdev 15%. In that case the odds ratios over 2.5 years are: (calculated as norm.pdf(x, 3, 15 / sqrt(2.5)) / norm.pdf(x, -2, 15 / sqrt(2.5)))
At -5%, odds ratio = 0.73:1 At -10%, odds ratio = 0.56:1 At -15%, odds ratio = 0.42:1
VMOT's 6.5 percentage point underperformance May '20 to Nov '22 (using the assumptions above) is an odds update of 0.68:1. VMOT's ~8% annualized underperformance since inception vs. the MSCI World index is an odds update of 0.39:1.
FWIW I'm not well-versed in this sort of calculation, you probably know more about it than me. I know that this estimate is wrong in various obvious ways (eg it's using a normal distribution, only two hypotheses, it's not allowing for the possibility of being wrong about the volatility) but it's at least some sort of approximation.
For some more historical context, I used data from the Ken French data library to test the performance of a VMOT-like strategy[9] vs. the US market going back to 1926. The 3 biggest relative drawdowns were
69.6% -> 2010-05 to 2020-06
62.2% -> 1983-08 to 1991-01
31.1% -> 1969-02 to 1973-03
The big 69.6% drawdown gives an odds ratio of 0.074:1 which is pretty substantial. (Although this is cherry-picking the peak and doesn't correspond to the out-of-sample period.)
To explain why I still have a high credence that value will work in the future, I would change this two-hypothesis model into a multiple-hypothesis model that includes (at least)
(There are other hypotheses too, those are the main ones I can think of.) The poor streak is an update against #1 and in favor of #4. To get a better comparison, I'd specifically look at the odds ratio for the structural return component of the value factor. The odds ratio depends on the null hypothesis. If NH is that structural return = 0%, then the recent structural return is still high enough that it's an update against NH. If NH is that structural return = 4.1% (which is what it was 2007–2022 using the way I defined the value factor) and AH is structural return = 6.2% (the historical average), the odds update is 0.64:1. Although the measured structural return varies a lot depending on how exactly you implement the value factor.
Anyway, I don't have a mathematical justification for this[11] but my general sense is that "big" odds ratios in investment performance don't matter much:
I am not sure about the apparent conflict between (a) you can get pretty big odds updates by looking at periods of bad performance and (b) the consensus among smart investors I follow (eg Ken French[10]), and the apparent conclusion from history, is that periods of bad performance don't matter much, and most investors way over-update on losing streaks—institutional investors tend to take money out of funds/strategies right before they outperform and move them into funds/strategies that are about to underperform (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1523736), such as CalPERS scrapping their tail hedging strategy in Jan 2020 when it would have made a killing in March.
(There's also a question of how to deal with in-sample vs. out-of-sample evidence. If you treat them the same (which I don't think you should), then you'd have a very high credence that factor investing works, and it would take something like a century of contrary evidence to change your mind—see https://blog.thinknewfound.com/2018/06/factor-fimbulwinter/.)
There's an anthropic argument for expecting future returns to be much worse than historical returns. You should expect to invest in a strategy when it's becoming the most popular, because that's when most people are investing. And if a strategy becomes more popular, that should bid up prices and squash future returns. But for something like value investing, we can refute this by looking at valuation spreads and seeing that they've widened, not narrowed.
The other way value investing can fail is if the economy fundamentally changes in a way that makes it not work (irrespective of how popular it is). I don't think the anthropic argument is as strong here, so there's more of a need to explain why value stopped working right now rather than at some other time, so this hypothesis demands more evidence than the "it's too popular" hypothesis which you can easily justify with anthropics.
[1] It's not really luck, it's probably driven by economic forces—for example, modern tech companies established strong moats which helped them keep persistent earnings growth more so than most companies did historically. This is debatable, but I'm inclined to say that the high profit margins on companies like Alphabet and Microsoft are temporary and competition will eventually drive margins back down. I have a moderately strong prior that markets are competitive and companies can't maintain moats indefinitely.
[2] The numbers for 2020–2022 are kind of insane...my interpretation of what's going on here is by starting in July '20, this data range is effectively picking the bottom of the COVID crash. Value companies' profits got hammered by the economic shutdown, and then the stimulus rebounded the earnings of companies in general but especially value companies, and meanwhile the market was kind of manic about growth companies in 2021-ish (see eg ARKK having a massive return and then crash, meme stocks, Matt Levine's "bored market hypothesis"). For my particular definitions of growth and value, the growth basket experienced -10% earnings growth July '19 to June '20 while value experienced -39%, and then July '20 to June '22, growth rebounded with +8% annualized and value earned +38% annualized. But growth P/Es increased while value P/Es decreased.
[3] A quick probability estimate: I don't have any special insight on AI timelines, I'll just say 25% chance of transformative AI in the next 10 years, which is roughly the "EA consensus forecast". 30% chance that it specifically benefits stockholders, and 50% chance that it differentially benefits growth stocks over value stocks by enough to explain the current spread. This isn't well-considered, just a gut estimate. Result = ~4% probability that AI can explain the current valuation spread. I expect you would put a higher probability on #2 because a slow-takeoff world is more favorable to stockholders and I lean more toward Eliezer on the slow/fast debate.
And even given a thesis like this one, IMO it would probably make more sense to invest in value stocks + a handful of strategically-chosen AI-relevant growth stocks, rather than to invest in growth stocks indiscriminately. I'm considering doing something like this but I want to investigate more first.
[4] I should clarify that growth companies actually do systematically outperform value companies in terms of earnings growth, but not usually by enough to justify the higher P/Es. So what I really mean is that growth companies would have to outperform by a significantly wider margin than they did historically.
[5] The P/E * earnings formula makes sense for individual stocks but it's not the whole picture for baskets of stocks because stocks can enter and leave the basket, so the basket's return is not equal to the return of any fixed set of stocks. The return for a basket is really the return due to multiple expansion plus the return due to earnings growth plus the return due to stocks migrating in and out of the basket. The Research Affiliates paper explains this in more detail.
[6] FWIW I would take Research Affiliates' research with a grain of salt because I don't think they have great epistemics in general, but I think this particular paper is pretty good. They do totally gloss over their finding that the structural return has gone down recently, which I think is a pretty important result that deserves to be addressed.
[7] Laursen & Richardson have a similar paper Is (Systematic) Value Investing Dead? which I think uses a better methodology—it looks at the predictive power of value metrics at the individual stock level—but I can't replicate it because I don't have individual stock data up to 2022.
[8] Like I stated before, I now believe 13% is too low. I wrote this part before writing the earlier part. A higher volatility would make the odds ratios closer to 1.
[9] The simulated strategy puts half in the top value decile (measured by B/M) and half in the top momentum decile, and shorts the market any time the market's 12-month SMA is below 0. I added a -2% annual cost.
[10] A quote from Ken French: "[P]eople are crazy when they try to draw the inferences that they do from 3 or 5 years or even 10 years of performance on an asset class or any actively managed fund or a non-index fund. [...] They see five-year track record and say, "Aha, I know I have a great manager here," when in fact you'll learn almost nothing about a manager's skill from five years of performance." From the Rational Reminder podcast, episode 100.
[11] I haven't really looked into this, but one mathematical justification that comes to mind, and intuitively seems to match the data: rather than drawing from some fixed underlying distribution, markets experience "regimes" where the length of the regime follows something like an exponential distribution. That could explain things like the 1994–1999 tech bubble, where instead of 5 out of 6 years independently experiencing rapid market growth with a widening valuation spread, these years were all correlated because they were part of a particular regime, which we can say in retrospect lasted 6 years but at the time it had indeterminate length. This "regime" perspective matches the way people talk about investing, although I wouldn't put too much credence in it because people might just be finding coincidental clusters in independent data.
My ideal self spends most of my EA Forum time reading technical posts about various cause areas, both to stay up to date on the ones I know a lot about and to learn more about the ones I'm less familiar with.
My actual self disproportionately reads Community posts because they take a lot less energy to read.
But I reserve almost all my upvotes for more technical posts to help nudge myself and others toward reading those ones more.
For instance, given utilitarianism, the Equality Result probably implies that there should be a massive shift in neartermist resources toward animals, and someone might find this unbelievable.
I would make the same claim more strongly: "modus tollens" / "reductio ad absurdum" (as in, "this assumption gives a conclusion I don't like", rather than "this gives an internally inconsistent conclusion") style ethical reasoning is, broadly speaking, not good. Unless you believe standard 21st century morality is correct about everything, you should expect your ethical assumptions to lead to some unexpected results. Ozy wrote something about this that I really liked:
[I]f your moral reasoning doesn’t produce conclusions that seem absurd on the face of it… why are you bothering? I want to be the sort of person who would have come up with the absurd conclusion that slavery is wrong, or the absurd conclusion that women should have rights, or the absurd conclusion that sodomy shouldn’t be illegal; therefore, right now, I am the sort of person who comes up with the absurd conclusions that eating meat is wrong, malaria net donations are morally mandatory, and global warming is really important.
OP said something similar, which is a less general argument but, I think, harder to dispute:
Essentially, it amounts to casting doubt on a broadly empirical conclusion—that particular animals have certain capacities that allow them to realize some amount of welfare—based on a moral conclusion.
I thought this was the most powerful sentence under the "Balking at the Implications" heading.
This is good news! I'm not familiar enough with the law to know whether this ruling is constitutionally justified, I would have preferred to see the Supreme Court ban animal confinement entirely on 13th Amendment grounds, but society is not currently at a point where that has even a distant chance of happening, so I'm happy with this Prop 12 ruling.