Posts
Comments
For those who are:
- Mathematically literate, but
- Not familiar with this particular analogy (of proofs <-> agents)
Do you know of a good reference for how to interpret discussions like this?
For example: " tries to prove that , and tries to prove " -- If A and B are propositions, what does it mean for a proposition to try and prove another proposition?
(There might be more language that needs interpreting, but I got stuck there.)
As a follow-up: I did this for a while, but I've become convinced there are a couple effects that make this not as good as it sounds:
- Futures have taxes paid in the year gains are made, which significantly reduces returns in simulations I've run. In an ETF or mutual fund, you can instead let those gains ride.
- Futures have an implicit financing cost, and portfolio performance is very sensitive to this cost if you're using a lot of leverage (e.g. for intermediate term bonds).
- Leveraged ETFs fluctuate a lot, and need to be rebalanced with the rest of your portfolio. This causes taxes like above. If you don't rebalance, your leverage ratio changes which causes the portfolio to behave poorly as well.
With all of these effects accounted for, the gains from leveraging look very modest and depend a lot on what time period you look at. Given the risks, I've decided against it for myself.
I've got no attachment to the phrase, I meant it in the sense of (From Wikipedia):
In simple terms, a negative externality is anything that causes an indirect cost to individuals.
I think e.g. paying for labor increases the demand for labor, thus increasing the price everyone else pays. That's an indirect cost to them. I didn't intend to make any claims about rights.
I disagree that believing there's a negative externality of production leads to the position you're arguing against -- for instance, I might think the negative externality is very small compared to the positive gains from trade. But I appreciate you pointing out exactly where you disagree with my framing.
The part that is missing from your question is, in a parallel reality where you didn't buy the PS5, how did you spend the extra money
Correct, and intentionally so. This is why I compare to setting the money on fire instead. Maybe I have the wrong framing, but it seems valuable to me to look at the marginal value of a single spending act in isolation. I might do many things with the money (like invest, spend, or donate it in various ways), and it would be interesting to know each of their values independently so I could compare them on even footing.
Maybe the parallel reality we should look at is the one where you worked less, got less money, and therefore didn't buy the PS5
Maybe! I think if you're trying to influence social policy, you should definitely think about that. But as an individual, I'm curious about the separate positive/negative effects of more work vs more spending.
some traditional alternatives are "you extort money from people at a gunpoint, then spend the money to buy PS5", which do not generate so much positive value for the others.
Thanks for bringing this up, it hadn't occurred to me.
I basically agree with the rest of your answer but, I'm still on the hunt for quantification!
I strong downvoted this, because it doesn't answer the question and instead seems to me to be making a political point.
Insanity Wolf would tell you you're absolutely right.
About what? I don't think I believe any of the things you're arguing against. I'm just wanting to get a quantitative sense, whether it supports or opposes the political point you're trying to make.
Thanks for this post, these kinds of details seem very useful for anyone wanting to attempt this path!
A worry I have: there are people who long for the imagined lifestyle and self-description of being an independent AI alignment/agency researcher. I would categorize some of my past selves this way.
For many such people, trying to follow this path too enthusiastically would be bad for them -- but they might not have the memetic immunities that protect them from those bad decisions. For instance, their social safety net might be insufficient for the level of financial risk, or the career tradeoffs might be very large. This post is enthusiastic, but I think many people need to be urged caution when making major life changes -- especially around such high stakes causes, where emotions run high.
So for my past selves, I'd disclaim:
- It's ok (and good) to prioritize your own financial and social safety net. You can revisit your ability to contribute from a better position in the future. The risks of things not going as well for you are very real.
- When starting down such a path, you should have a clear fallback plan that does not involve immense suffering. For instance, make effort for X time period before attempting to find an alternate job if you have not achieved Y income. Do this only if you have confidence you will not take a too-large psychological hit from the failure.
For buy-and-hold strategies: leverage (borrowing to invest) can be used to increase returns. The Sharpe ratio of a portfolio is a measure of it's risk/reward trade-off; there's a theorem that given a set of assets with fixed known distributions and ability to borrow at the risk-free rate, the Kelly optimal allocation is a leveraged version of the portfolio with highest Sharpe ratio. You can calculate that optimal leverage in various ways.
In practice, we don't know future Sharpe ratios (only the past), we don't have assets with fixed known distributions (to the extent there is a distribution, it likely changes over time), and there's other idiosyncratic risks to leverage.
Using past data, however, is highly suggestive that at least some leverage is a great idea. For instance, a 40/60 stock/bond portfolio has higher Sharpe ratio than a 100% stock portfolio in almost all historical models since bonds and stocks have low correlation. It's not unreasonable to assume that low correlation will hold going forward. If I leverage up a stock/bond portfolio to the same volatility as the corresponding 100% stock portfolio, I see a 2-4% increase in yearly returns depending on time period. This is easily achievable by buying leveraged ETFs.
Naive Kelly models will usually recommend more risk than the stock market, so if your risk tolerance is even higher the (potentially naive and misleading) math is on your side. Be careful if you might need the money soon: these models assume you're never withdrawing!
My opinion: ~1.5-2.5x leverage on stock/bond portfolios looks reasonable for long-term investors, although it really depends on the assets -- short term treasuries for example are so stable that even 10x leveraging isn't crazy.
In the real world, you can leverage by:
- Buying leveraged ETFs. These work just like buying stocks or regular non-leveraged ETFs. Vanguard doesn't allow these, so I use Fidelity. I might recommend SSO and TMF, along with international + extended market diversification. Since the leverage on these is set daily, they eliminate some of the tail risk of leveraging. Downsides: volatility decay in sideways markets and ETFs can simply fail to track their (leveraged) index.
- Using margin accounts. This is where your broker lends you money, using your other investments as collateral. Rates are often really bad/not worth it -- be careful. Also they'll sell your other investments if you become overly leveraged due to a downturn (a "margin call") -- which means forced selling during a downturn, which can cause you to miss out on any recovery. I don't like this.
- Directly taking on debt / counter-factually not paying off debt. If you take out a mortgage instead of paying cash for a house, the money you saved can be considered to be "on loan" at the mortgage rate. Investing it is then a form of leveraging. This is bad if the rate is too high, but mortgage rates are really low right now.
- Futures. These allow higher leverage than ETFs and adoption of various strategies if you're willing to monitor them daily -- for instance, you can leverage/deleverage on a daily basis with no (additional) tax penalty. Great for getting really high leverage, e.g. for stable investments like treasuries. Downsides: your positions will be closed if you run out of margin -- which means you can be forced out of the market during a bad enough downturn, potentially missing the recovery. For high-leverage strategies, this means you probably need to monitor daily and de-lever as appropriate. Gains from futures are immediately taxed at 60% short-term 40% long-term capital gains, introducing tax drag that you wouldn't have with ETFs.
- Options. I don't know much about these. Here's a book talking about passive indexing approaches using options, which I haven't read: https://www.amazon.com/Enhanced-Indexing-Strategies-Utilizing-Performance/dp/0470259256.
- Other stuff (box spread financing?)
You can kind of backtest various strategies for yourself using e.g.:
- https://www.portfoliovisualizer.com/backtest-portfolio#analysisResults. Entering a negative percent for CASHX until your portfolio adds up to 100% is an approximation for cost of borrowing at the risk-free rate. They also have ticker data for leveraged ETFs, although most have short histories.
- https://www.bogleheads.org/wiki/Simba%27s_backtesting_spreadsheet. Entering a negative percent for 3-month T-bill rate under "Compare Portfolios" until your portfolio adds up to 100% approximates the cost of borrowing at the risk-free rate.
Keep in mind that the past doesn't predict the future, and naive back-testing might not track historical reality perfectly or may be missing some idiosyncratic risks of leveraging certain ways.
For your maximal laziness, here's an ~1.9x leveraged ETF allocation I made up just now that's around 40/60 world stock market/long-term bonds: 36% TMF, 16% SSO, 8% VXF, 40% VXUS. Full disclosure: Lots of people would think this is a terrible allocation since "everyone knows" the bond market is due to crash due to current low interest rates.
For increasing salary vs investing: do these really funge against each other? Why not both? Generically I'd think income should dominate before you have a lot invested, and then investment strategy becomes important once your yearly average investment returns become similar in scale to your yearly income.
This is mentioned in the "don't screw up" section. By market cap, crypto was 1.7% as big as the world stock market, and Bitcoin 1% as big, so those seem like good starting points -- adjust from there for your desired level of risk vs reward.
IMO, self picked stocks are dumb. You give up diversification benefit for no reason, unless you think you know better than the market (which you don't).
Great point; I agree. Also a great example of missing an obvious risk; I hadn't noticed that before linking.
The calculator here allows simulating withdrawal rates by asset allocation, although it only has data back to 1970 so is a bit limited. I get the same safe withdrawal rate (4.3%) for 30 year retirees using either 100% US or 50/50 US/ex-US over that time frame. 100% Japan had a 1.5% safe withdrawal rate.
Learning about the existence of state guaranty associations has decreased my sense of how big I think the counter-party risk is; thanks for sharing this.
Re: running out of money, I've added a section on the risks of retiring too early to address this concern in more detail. I now agree that annuities might be a good idea to address this if you are old enough, and I was probably overly worried about counter-party risk.
Re: the 4% rule, it is indeed more of a guideline than a guarantee. More details are available here: https://thepoorswiss.com/updated-trinity-study/. The link shows a 100% stock allocation with a 3.5% withdrawal rate has a historical 98% success rate over 50 year periods starting from 1871 -- if you are never going to generate income again and are never going to increase real expenses, you may be able to buy quite a bit of safety by going to 30x instead of 25x and holding a 100% stock portfolio.
I would worry about the counter-party risk with annuities; if a single company goes out of business, you might be bust. Even if you distribute across many companies, I'd think it's more likely that the whole sector goes bust than that your portfolio devalues to 0 in some other way.
For that reason I'd lean toward not putting too much of my assets in annuities -- but maybe it works out so that the counter-party risk is smaller than the risk of running out of money otherwise.
For posterity's sake: I became convinced this is practically doable (using either treasury futures, leveraged ETFs like NTSX, or maybe options which I don't understand as well) and probably a good idea/not very dangerous if done correctly. I think that fact is slightly info-hazardous for a couple reasons:
- You shouldn't trust most people to correctly advise you on financial products, to not be delusional, or to have your best interests at heart. So it's hard to figure out exactly what to do. Index funds overcome this problem through the sheer size of their giant pile of empirical evidence and expert consensus; basically everyone agrees that they work as advertised, and no one reports getting accidentally burned using index funds -- except when the whole market crashes, where they behave as expected.
- If you learn that it's probably a good idea when done correctly, you might feel obligated to go do it, and then you might do it incorrectly and foreseeably lose a bunch of money.
- Because the pile of empirical evidence is less giant, it might not turn out to be such a good idea in retrospect, so it's fundamentally riskier (even taking into account the risks people calculate). I'm sure someone would argue the pile is giant, but even if true that's probably only the case if you're sufficiently expert to judge more obscure evidence piles which most of us are not.
So I'd STILL recommend you not do this unless you're extremely curious in this area, have no hang-ups, feel competent and trust your own judgment around things like intimidating financial products, have no track record of unwise gambling behavior, and have a stable enough life that if you fuck up you won't be in a bad situation.
Here's some resources. If you're not interested enough to read and enjoy stuff like this, probably avoid doing this:
- https://www.aqr.com/Insights/Research/Journal-Article/Why-Not--Equities
- https://www.amazon.com/Enhanced-Indexing-Strategies-Utilizing-Performance/dp/0470259256
- https://www.bogleheads.org/forum/viewtopic.php?t=143037
But I'll probably do it myself and might write a blog post about it.
After reading around for half an hour, I think there's a decent chance that some form of leveraged investing via e.g. ETFs might be a good idea. The basic idea makes sense to me.
This is currently completely out-weighed by my "being too clever in markets is a great way to lose all of your money" prior. But I'll probably look into it more and see how convincing I find the numbers and historical evidence. If I'm pretty convinced I could see myself allocating 10-20% of my investments in a leveraged strategy at some point in the future.
A cursory look at box spreads makes me think it's the kind of thing with so many caveats that I'd never feel certain I'd eliminated enough tail risk from it.
Re 1: This is a good point; I did the math on this at some point for myself and ended up still landing on traditional by a large margin (even though I wanted it to turn out pure Roth for simplicity). But it'll be dependent on your expected tax rates, opportunities for low-tax conversion, and your retirement timeline (more tax-advantaged money dominates on longer timelines).
Also yeah, I skipped backdoor and mega-backdoor to keep things simple. The goal was to give people a linkable 90/10. The outcome I was aiming for is that people read, get the important parts of the memetic package, and then some of them will dive in more to find things like that for themselves -- for instance, they're mentioned in the flowchart I linked.
Re: tax advantages of homes, yeah . Homes are probably a better long-term expected value than I make them sound for this and other reasons. Still, transaction costs suck, and I think they stop being a good idea if you buy/sell them too often -- I've heard you need to hold ~5 years on average to break even on transaction costs vs mortgage advantage, but haven't done the math myself. I am biased toward the flexibility of being able to change my physical location in order to take better jobs, decrease my commutes, decide to decrease my expenses, etc. If you do buy a home, you can ameliorate these concerns by just buying a small home so that you can more easily decide to eat the proportional transaction costs if necessary.
Re: financing things at low interest rates to instead invest in the market, I agree this is correct to maximize expected value but 1. it's not all that big of an impact for things smaller than a house or new car and 2. I think most people would do better by avoiding the over-spending tendency that credit brings, and the negative psychological effects of future spending obligations. If you're buying a house or new car regardless and can get a low interest rate, I agree you should take out low interest debt instead of buying it with cash.
The intent isn't to neglect these advantages; rather, I (probably wrongly) assume that everyone is familiar with these advantages -- this is my intent in noting the psychological difficulty of dropping from a 500k home to a $1k/month apartment.
The intent is instead to bring to attention the nature of the financial trade-off. I use a pretty simple model, but you can dive into the counterfactual yourself: what's the price you're paying for owning your home instead of owning an index fund and renting? How low could you go on rent, and what would be the impact on your financial life? Is that tradeoff worth it to you for the advantages
The answer can certainly be "yes" -- but I think people are biased toward assuming yes when they haven't actually examined the issue.
I roughly agree with this. My biggest concerns around buying housing are:
- The transactional friction of buying/selling homes causes opportunity costs.
- People tend to buy too much due to low-interest credit.
But you correctly point out upsides that I don't dive into.
Skimming that link, I think it shows backtesting; have you actually beaten the index yourself with real money? For what time period / amount of assets?
I mostly avoided leverage in this post because I don't use it and kind of don't trust it. But if I had to give a better defense of avoiding it, it would be because 1. it's really easy to lose a bunch of money if you use it wrong and 2. I'm not sure there's a reliable way to borrow money at low enough rates to get good results. Most of what I've read about leverage pretends the interest rate is 0, which it's not -- looks like Robinhood offers 2.5%? What's the most reliable interest rate people can get, and does this rate kill results?
Definitely agree not to discount bonds though; without leverage, small amounts give you large risk reductions for small reductions in growth. I personally am 100% stocks because I'm preferring to maximize (unleveraged) growth, and have decided I can tolerate the increase volatility over 10-20 year time spans.
Re: 1., I'm personally unwilling to move outside the U.S. but agree it could make sense if you can make it work for you while maintaining a high salary.
Re: 2 and 3, I completely agree. I think in particular about longevity medicine as a potential future expense. You can certainly build up support for higher-than-current expense levels to address these risks. You might also retire to less profitable or more risky activities that you find more enjoyable (but that supply >0 income), or simply stay in your current profession -- but with the advantage of having higher option value.
I agree the company's current price should be lower than $10 million. But if it starts at price P and I expect it to go up at the risk free rate r, then at a time T later the company's price should be in expectation. At some point, that'll be substantially more than the $10 million I expect it to pay out.
- I'd be interested for both 100% growth and 100% dividend stocks -- I'm not sure why they'd behave differently w.r.t. to this.
- By underlying value, I'm not sure. Something like the real dollar value of all of its capital -- or the real dollar price someone would pay to own it in its entirety?
- By price, I mean what you can buy the stock for on the market.
Let me try clarifying: The volatility argument seems formal rather than empirical, so I'm wondering what we formally need to assume to make it go through.
I'd summarize the argument as "since stock prices are volatile, they're expected to go up over time (more than the risk-free rate)". But then why are stock prices volatile? I assumed that they're volatile due to "underlying value" being volatile/hard to predict.
So my hypothetical is a company whose "underlying value" is hard to predict, but where the expectation of its "underlying value" is constant over time. To make it easier, assume the company is a magic vault that currently contains $10 million real dollars, and the money will undergo a stochastic process with the given property, and everyone knows this. Maybe it will disperse it's full value as a dividend at some random future point.
It seems obvious to me I shouldn't expect this company's price to go up faster than the risk free rate, yet the volatility argument seems to apply to it. So I'm trying to identify what I'm missing.
Question: say I have a company whose underlying value is volatile, but whose expected underlying value after any time span is the same as today's value. Both of the above arguments seem to suggest I should expect the company's price to increase over time, but wouldn't this unanchor the company's price from its underlying value?
This makes sense!
Do you know anything about the state of evidence re: to what extent this is happening and/or driving stock returns? I'm not sure how you'd pick this apart from other causes of currency devaluation.
Rephrase attempt: If you were to buy and hold a company's stock, and you don't expect to know better than the market, then your anticipated gains over time are independent of what you think the company's underlying value will do (since the market has already priced that in). But you should still anticipate gains over time due to volatility's effect on pricing.
Ok, so the argument would go:
- Stocks' expected values (in terms of time-discounted dividends or similar) have volatility, from things like business decisions, technology development, and capital re-investment.
- A stocks expected value serves as an anchor point for its current price (kind of?)
- A stock's price will change around that anchor until it has expected returns that justify the volatility risk.
- Thus a stock price will increase when either its expected value goes up, or its volatility goes down (without changing expected value).
- In an efficient market, knowledge that changes your expected value of the stock is probably already priced in, but you can still capture the gains due to volatility.
Does that seem in line with what you're saying?
You may have to hold my hand on this one: I can agree the value of the stock (in the time-discounted future dividends sense) will go up after 10 years due to time-discounting -- the technology would enable value production that "comes into scope" as it gets closer in time.
But is there any reason other than time-discounting that the PRICE won't go up immediately? For instance, if I expect the time-discounted dividend value of the stock is $50 today and will be $5,000 ten years from now, and the rest of the market prices it at $50 today, then I could earn insane expected returns by investing at $50 today. Thus, I don't think the market would price it at $50 today.
Is this the same as positing "the market is continually surprised by the pace of technology"?
E.g., say I value company X's stock at $100. Then I learn a new fact that there's a 50% independent chance the company will discover a technology that doubles its value by 1 year from today. If I ignore all other factors, my estimate of the company's value 1 year from today should then be $150. If the company discovers the technology, I'll value it at $200, and if not, I'll value it at $100.
For the market to trend upward as it does, it seems like either:
- Everyone's consistently getting those bets wrong, and underestimating how often/how much technology will pay off, OR
- There's something else to the story, like time-discounting or a different way of thinking about risk.
For growth stocks, why is the expected future growth not already priced in? If I know the company will be re-investing into future growth later, why not invest now?
There may be uncertainty, but if stocks on average trend upwards, doesn't it mean that the market continuously underestimates the amount that companies will re-invest?
Stock prices go up even in the absence of technological advancement because stocks are tied to the bond market via arbitrage.
If I'm understanding correctly, you're suggesting they should go up at least at the bond market nominal rate -- but they tend to go up much faster than that? I didn't fully read through the argument, so I might be misunderstanding.
For anyone like me: it's easy to read this advice as "if you're not curious, you're therefore bad/doing something bad", which might suggest attempting to brute force an emotional state of curiosity. I think that's probably emotionally harmful.
It can be the case that:
- Curiosity is very useful for being a good listener
- You are not curious about (this person) in (this situation)
From there, you could:
- Hide your current lack of curiosity and go through the motions as best you can. I think this is the best option quite often!
- Tell them your honest feelings. Maybe you're not very interested at the moment but are worried about them feeling not cared about, and you can tell them that. Might be a bad idea if they're vulnerable or not trustworthy.
- Investigate why you are not curious (either internally or with the other person), which might spark curiosity or suggest how else you should move the conversation.
- (any other option)
I think it's healthier to grow curiosity as a natural extension of your desires instead of shoving it in as "ah, now I have to take the curiosity action to perform this task". I don't think the author was suggesting the latter, I just noticed my inclination to read it that way.
No worries, was worth clarifying. I edited the post to link this comment thread.
Yes, I understand this point. I was saying that we'd expect it to get 0% if its algorithm is "guess yes for anything in the training set and no for anything outside of it".
It continues to be surprising (to me) even though we expect that it's trying to follow that algorithm but can't do so exactly. Presumably the generator is able to emulate the features that it's using for inexactly matching the training set. In this case, if those features were "looks like something from the training/test distribution", we'd expect it to guess closer to 100% on the test set. If those features were highly specific to the training set, we'd expect it to get closer to 0% on the test set (since the model should reject anything without those features). Instead it gets ~50% which means whatever it's looking for is completely uncorrelated to what the test data looks like and present in half of the examples -- that seems surprising to me.
I'd currently interpret this as "the discriminator network acts nonsensically outside the training set + generator distribution, so it gets close to chance just because that's what nonsensical networks do."
Thanks for sharing thoughts and links: discriminator ranking, SimCLR, CR, and BCR are all interesting and I hadn't run into them yet. My naive thought was that you'd have to use differentiable augmenters to fit in generator augmentation.
You can ask him on Twitter.
I'm averse to using Twitter, but I will consider being motivated enough to sign-up and ask. Thanks for pointing this out.
"compression" is not a helpful concept here because every single generative model trained in any way is "compressing"
I am definitely using this concept too vaguely, although I was gesturing at compression in the discriminator instead of the generator. Thinking of the discriminator as a lossy compressor in this way would be... positing a mapping f: discriminator weights -> distributions, which for trained weights does not fully recapture the training distribution? We could see G as attempting to match this imperfect distribution (since it doesn't directly receive the training examples), and D as modifying weights to simultaneously 1. try to capture the training distribution as f(D), and 2. try to have f(D) avoid the output of G. Hence why I was thinking D might be "obfuscating" -- in this picture, I think f(D) is pressured to be a more complicated manifold while sticking close to the training distribution, making it more difficult for G to fit it.
Is such an f implicit in the discriminator outputs? I think that it is just by normalizing across the whole space, although that's computationally infeasible. I'd be interested in work that attempts to recover the training distribution from D alone.
I think it's decently likely I'm confused here.