Should Earners-to-Give Work at Startups Instead of Big Companies?

By MichaelDickens @ 2021-11-12T22:55 (+85)

Summary

Confidence: Likely

Effective altruist earners-to-give might be able to donate more money if, instead of working at big companies for high salaries, they work at startups and get paid in equity. Startups are riskier than big companies, but EAs care less about risk than most people.

Working at a startup is easier than starting one. It doesn't pay as well, but based on my research, it looks like EA startup employees can earn more than big company employees in expectation.

Does the optimal EA investment portfolio include a significant allocation to startups? To answer that question, I estimated the expected return and risk of startups by adding up the following considerations:

  1. Find a baseline of startup performance by looking at historical data on VC firm returns.
  2. VC performance is somewhat persistent. EAs can beat the average by working at startups that the top VC firms invest in.
  3. Startup employees get worse equity terms than VCs, but they also don't have to pay management fees, and they get meta-options. Overall, employees come out looking better than VCs.
  4. Current market conditions suggest that future performance will be worse than past performance.
  5. Startups are much riskier than publicly-traded stocks, and the startup market is moderately correlated with stocks (r=0.7).

All things considered, my best guess is that more earners-to-give should consider working at startups.

Framing the problem

Suppose you're an effective altruist and you want to donate as much money as possible. Perhaps you've heard the arguments that EAs should start startups. Starting a startup is a lot of work and requires special skills, so you'd rather not. Maybe you'd like to invest in venture capital, but the really good VC firms won't accept your money. However, you wouldn't mind working at a startup. You could also work at a big company that pays a high salary. Which should you choose?

This is how I would think about the problem:

Say a startup offers you an equity package that's worth $X per year at the current valuation. At the same time, a big company offers you a salary that's $X higher than your salary would be at the startup. Both compensation packages have the same face value.

If you work at the startup, you get $X per year of equity. Some number of years later, the startup might go public or get acquired, at which point you can sell your equity for $Y. Over that time, your equity earned a return of Y/X.

If you work at the big company, you could invest your extra $X salary in the stock market. Will that investment earn a higher or lower return than the startup equity? Whichever you expect to earn a higher return is the one you should pick. (Well, that's not really true. Read the next section to find out why not.)

The right question

Which do we expect to earn a higher return, startups or the public stock market?

We don't have good data on the historical performance of startup employees. We do have data on VC returns, so we can use that.

"Have VCs historically outperformed the public market?" is the wrong question, because some VCs consistently outperform the average.

"Have top VCs historically outperformed the public market?" is the wrong question, because future expected performance probably isn't the same as past performance.

"Can we expect top VCs to outperform the public market?" is the wrong question, because startup employees don't earn the same returns as VCs. Employee equity has a worse liquidation preference than VC equity, but investors in VC firms have to pay fund fees, which employees don't. And employees have the meta-option to keep vesting when their company does well, or to quit when it does poorly. We can use these considerations to estimate the value of employee equity compared to VC equity.

"Can we expect startup employees to outperform the public market?" is the wrong question, because we need to consider leverage. Investors in public markets can use leverage to increase their risk and expected return, but startup employees can't.

"Can we expect startup employees to outperform the leveraged public market?" is the wrong question, because an effective altruist's goal isn't to maximize their own portfolio return, it's to maximize the expected utility of the overall EA portfolio. If no other EAs work at the startup where you choose to work, then you're adding better diversification than if you invest in the public market.

"Do startup employees contribute more expected utility to the EA portfolio than if they invested in the leveraged public market?" is the wrong question, because if they did work at a big company and invest their salary, they might be able to invest in something better than the broad market. For example, I have previously discussed investing in concentrated value/momentum/trend portfolios, and I made a rough attempt to calculate the expected utility of doing so. For us to prefer to become startup employees, startups would have to look better than the best possible public investment (whether that's value/momentum/trend or something else).

"Do startup employees contribute more expected utility to the EA portfolio than if they invested in the optimal set of public investments?" is more or less the right question.

Modeling the solution

Do startup employees contribute more expected utility to the EA portfolio than if they invested in the optimal set of public investments?

The answer to this depends on two things:

  1. How do we model the answer?
  2. What values should we use for the model inputs?

#2 is hard. #1 is sort of hard, but luckily, it's already a solved problem. I described an applicable model in Asset Allocation and Leverage for Altruists with Constraints. In short, we set up a mean-variance optimization problem where we assume 99% of the capital is controlled by other people, and we can decide how to allocate the remaining 1%. Suppose we can allocate between three investment choices:

  1. A typical investment portfolio, such as a stock market index fund
  2. The optimal (ex ante) public investment portfolio
  3. Startups

The simplest method is to assume we should put all our money into just one choice. What is the overall expected utility if we put our money into choice 1, choice 2, or choice 3?

If we tell our model the expected returns, standard deviations, correlations between these three portfolios, and a utility function, then the model will spit out the expected utility of each choice.

Let's say the average EA investment portfolio equals the global equity market, which is at least sort of correct. What's the expected return and standard deviation of global equities?

We have no idea how equities will perform in the short run. But in the long run, the market's return is somewhat predictable. And over long time horizons, volatility stays pretty consistent, so we can simply assume the future standard deviation of global equities equals the historical standard deviation.

Similarly, we can approximate the future standard deviation of startups, and their correlation with global equities, by taking the historical standard deviation and correlation and assuming they will stay the same.

The most difficult input variable is the expected return of startups.

What factors determine the expected return of startups?

How to estimate the expected return of startups for employees:

  1. Start with some index of VC returns, such as the Cambridge Associates Venture Capital Index.
  2. These indexes usually provide returns net of fees. VC investors have to pay fees, but startup employees don't. Add 2-and-20 fees back in to get the gross return.
  3. Unlike with public market investors, VC firms that beat the market in the past tend to continue to beat the market. Employees can choose to work at startups with funding from top VCs. Add a premium to the expected return to account for this.
  4. Maybe EAs can pick startups better than top VC firms. Possibly add a premium.
  5. Employees get worse equity terms than VCs, so subtract some discount to account for tihs.
  6. Startup employees get meta-options, which VCs don't get. Add an appropriate premium.
  7. Use the current market environment to forecast how future returns will look compared to past returns. (This is the sketchiest step, so we might skip this and just assume future returns equal historical returns. But I think we don't want to skip this because we don't even really know what historical returns were—more on that later.)
  8. Giving now might be better than giving later. If so, that means we shouldn't compare startups to public investments because public investments aren't the best thing to do with money. Instead, we should compare startup equity to money donated now.

In the next four subsections, let's break down each of these steps.

Returns for VC firms

I used Cambridge Associates' Venture Capital Index 2020 report to find the aggregate historical return of VC firms from 1995 to 2018[1]. According to this data set, VCs had a geometric mean return of 13.1% with a standard deviation of 18.9%. That's a good starting point.

As the saying goes, a person with one clock always knows what time it is. Someone with two clocks is never quite sure. Cambridge Associates' 2017 report has data from 1988 to 2016, which gives a geometric mean return of 14.9% and a standard deviation of 17.2%.[2]

According to these data sets, VC experienced several regimes:

The "historical performance of VC" substantially changes depending on which time period you look at. And we don't know what sort of regime will come next.

Another problem: Other VC return databases give entirely different numbers. For example, I could have used the VentureXpert database, which some (e.g., Koch (2014)[3]) claim is more accurate. (I used Cambridge Associates purely out of convenience.) Cambridge Associates tends to give higher VC returns than other databases (by a couple percentage points).[4]

I will use the 2020 Cambridge Associates report as a starting point. Just be aware that the startup market will likely behave very differently in the future.

Now let's convert net returns to gross returns. If we assume VCs usually charge 2-and-20 fees, this step is pretty easy. Using the Cambridge Associates 1995–2018 data, we find a gross historical return of 18.1% with a 23.5% standard deviation.

Some people (especially VCs) like to talk about how the "top quartile" of VC firms persistently beat the market. This is true, but potentially misleading. VC firms who beat the market in year N are more likely than chance to beat the market in year N+1. But top-quartile firms are by no means guaranteed to stay in the top quartile.

How strongly do top-VC returns persist? Data from Harris et al. (2020)[5] (1984–2014) is presented in the table below[6].

The table uses these terms:

IRR PME
Average VC 14.8% 1.22
Top-quartile VC, backward-looking 45.3% 2.60
Top-quartile VC, forward-looking 26.3% 1.70

Results for VC firms 2001–2014:

IRR PME
Average VC 10.4% (+)
Top-quartile VC, backward-looking 30.0% 2.11
Top-quartile VC, forward-looking 14.7% 1.20

(+) This figure is not provided by Harris et al.

In the full sample (1984–2014), top-quartile VCs retained about half of their outperformance out of sample, although they lost most of their relative outperformance (compared to the S&P 500). In the post-2001 sample, they lost most of their outperformance both in relative and absolute terms, but still showed nonzero persistence.

Harris et al. also did a regression analysis, and found that across all VC firms, one third of PME outperformance persisted.

(Note: Harris et al. used data from Burgiss, yet another source for VC returns.)

Returns for employees

Almost all startups give preferred shares to VCs and common shares to employees. Normally, preferred shares get a 1x liquidation preference. That means if the company exits, VCs are guaranteed to get back at least the money they put in before employees get anything. This makes employee equity worth less than it appears.

Example:

This is basically standard practice. Some startups also give special advantages to VCs. There are lots of ways they can do this, such as:

  1. a liquidation preference that's higher than 1x (e.g., a 2x preference guarantees that VCs get to double their money before employees get anything)
  2. a ratchet, which gives VCs protection against dilution at the expense of employees

These sorts of conditions are really bad for startup employees. You might just want to avoid any startup that offers terms like these. (If you work at a startup with no sketchy terms, and they raise a new round of funding that introduces sketchy terms, that alone might be enough reason to start looking for a new job.) For more on what to watch out for, read Ben Kuhn's How bad are fundraising terms?

Even without any bad terms, employee stock options introduce some other problems:

You can avoid these problems by exercising your options as soon as they vest, or even early exercising if you can. But even if you do exercise, you might end up paying higher taxes when the startup exits because you'll get pushed into a higher tax bracket. You can (at least partially[8]) mitigate this by donating the stock instead of selling it.

(A lot of people can't afford to exercise their employee stock options. Perhaps an EA org could make grants or loans to help EAs exercise their options. That would be difficult to set up and they'd have to carefully vet grant recipients, but maybe it could work.)

According to my rough estimate, a 1x liquidation preference reduces the expected value of common shares by about 10%[9]. That equates to around 1–2% per year, depending on how long the company takes to exit. Let's assume a 3% annual discount due to liquidation preference plus tax disadvantage.

Many startups pay below-market compensation by claiming that their equity is underpriced. Don't buy it. The whole point of working at a startup is that your equity will earn (in expectation) above-market returns. If your employer adjusts for this by giving you less equity, that ruins the (monetary) advantage of working at a startup.

This essay is not about employee equity terms, but it's an important topic for anyone considering working at a startup. For an in-depth guide, see The Open Guide to Equity Compensation by Joshua Levy et al. For something shorter, I recommend Ben Kuhn's checklist for stock option offers.

Startup options are better than they look because employees get "meta-options": your compensation package gives you the option to vest stock options at the current price for the next four years. If the company does well, you can "exercise" your meta-options by continuing to work there. If it doesn't do well, you can quit. Neither startup founders nor startup investors can do this.

I piggybacked on Ben Kuhn's meta-option model (my code here) and found that meta-options are worth an extra 16 percentage points of return (!!). I just did a quick calculation and didn't perform a sensitivity analysis or anything, so this number could be way off, but let's go with it for now. If correct, this number is so large that working at a startup looks more profitable than starting a startup, unless you believe you'd make for an unusually good entrepreneur.

It's worth mentioning that you could work at a big company that offers equity compensation, which also behaves like a meta-option—albeit a much less valuable one, because big company stock is not as volatile. Using similar methodology, I found that meta-options at a big company are worth 5 percentage points. That means startup meta-options provide an extra 11 percentage points of value (16% – 5%).

As far as I know, Ben Kuhn invented the concept of meta-options, and no one has ever rigorously analyzed them. My own modification of his program could contain bugs or logical flaws. The value of meta-options could be large enough to dominate every other factor, or they could be worth nothing. This subject strongly warrants a deeper investigation.

–––––

If we can get around the practical concerns, EAs can easily match the returns of top VC firms by getting jobs at their portfolio companies. Can EAs do even better? Can we outperform VCs at picking winning startups?

Let me say up front that I don't believe EAs in general can outperform top-quartile VC firms. But when I say I don't believe it, what I really mean is I assign it less than a 50% probability. So it might still be worth trying.

(To be more precise, I would give an 80% probability that at least a handful of EAs could pick winning startups better than top VCs, but I don't know how to identify those people in advance.[10] I'd estimate a 30% chance that, if a group of EAs decide to go work at startups and make a conscious effort to pick winning startups, then they will do better than top-quartile VCs.)

EAs as a group are really smart. But professional investors are also really smart, and the overwhelming majority of them still fail to beat the market. Maybe EAs are smarter? Maybe EAs are more rational or clear-thinking in some way that most professional investors aren't? I don't know.

It's possible that EAs could do a better job than VCs of identifying the best startups. On the other hand, EAs might also do a better job of identifying the best public investments. On the other other hand, startups are hard to invest in, so it might be easier to find underappreciated opportunities. If EAs have an edge in public markets, they probably have an even bigger edge in startups.

I don't have strong evidence on this either way, so I'm leaning on my prior that almost nobody can beat the market. We do have at least some evidence, but I'm not sure how to interpret it.

The evidence we do have:

  1. Over the past few years, EA investors have beaten the market. This is mostly driven by a single company (FTX), so I don't know how much we can infer from this.
  2. One person reviewed a few months' worth of EAs' proposed investing ideas and found that they had beaten the market over those few months. (I don't want to go into specifics because this review was not shared publicly, but that's the gist of it.)

In the rest of this essay, I will assume EAs can't beat top-quartile VCs—not because I am confident that this is true, but because I don't know how to evaluate the evidence. It could be a good idea to look into this in more depth.

Forecasting future returns

As discussed above, startup returns tend to vary a lot over time, so past performance does a poor job of predicting future performance. But we can't choose between two investments (in this case, public investments vs. startups) unless we believe something about how they will perform. So what should we believe?

The outlook for VCs looks worse than usual. The question is, how much worse? Should we expect future returns to be 1 percentage point lower per year? Or 20 percentage points?

Well, how much worse to US equities and bonds look? We can reliably predict bonds' long-term returns using the yield. 5-year bonds currently yield around 1%, compared to a 1984–2014 average nominal return of 8% (Damodaran, 2021).

Stock returns are harder to predict. In the short term, they're almost impossible to predict, but we can estimate their return over 10+ year periods with reasonable accuracy. Under a more EMH-y model that assumes no change in market valuation (e.g., AQR, 2021), forward-looking US equity return expectations look around 5 percentage points worse than they did from 1984 to 2014 (4% vs. 9% after inflation). According to a model that assumes valuations will revert to their historical average (such as Research Affiliates, 2021), returns look 10 percentage points worse (-1% vs. 9%). The truth is probably somewhere in the middle.

In theory, all asset classes should have the same risk-adjusted return. Startups are riskier than stocks or bonds. So if return expectations for stocks/bonds go down by some amount, expectations for startups should go down by more than that. But we don't know if this holds in practice, and we don't even know exactly how risky startups are. If bonds look 7% worse, and stocks look between 5% and 10% worse, then maybe we could assume VC will perform 7% to 10% worse in expectation.

As for top-quartile VCs: according to Harris et al., over the full historical sample, they outperformed average VCs by a full 11 percentage points. In the post-2000 era, they only outperformed by 4 percentage points, and had a public-market equivalent performance of 1.2 (which means they only weakly outperformed the S&P 500). It seems fair to assume that the future for VCs will look more like 2001–2014 than like 1984–2000, as the pre-2000 VC market was probably less efficient. We could simply assume top VCs will perform 4 percentage points better than average VCs going forward. But it's also possible that the gap between average and top VCs will continue to narrow.

Giving now vs. later

Working at a startup is comparable to working for a salary and investing it. But if giving now beats giving later, then you wouldn't want to invest your salary. Instead, you'd want to donate it right away. This makes working at a startup look worse because you can't donate your equity until it becomes liquid.

This possibility makes the comparison more difficult, so I will mostly ignore it. It's not as simple as applying a fixed discount rate to the value of your startup equity. Just be aware that my methodology for comparing startups vs. big companies only works if giving later is at least as good as giving now, at least for the next few years.

Putting together the expected return

If we combine all the numbers I came up with in the previous section, we get:

  1. 13% after-fees historical return to VC firms, or 10% after inflation.
  2. 15% real historical return before fees.
  3. Add 4% for the persistence of top-quartile VCs, giving 19%.
  4. Add 0% for EAs' extra outperformance. Still 19%.
  5. Subtract 3% for employee equity terms, giving 16%.
  6. Add 11% for meta-options, giving 27%.
  7. Subtract 9% for the relatively poor market outlook, giving 18%.
  8. Ignore giving now vs. later. Still 18%.

Thus, I predict a 18% real return for startup employees who try to maximize their earnings (by working at startups with funding from top VCs, getting good equity terms, and exercising their meta-options when necessary).

How big are the error bars on each of these numbers? In order:

  1. Historical return depends a lot on what time period you look at. Wide error bars.
  2. Calculating before-fees return just requires knowing fees, which are usually 2-and-20. Narrow error bars.
  3. It wouldn't be too surprising it top-quartile VCs had as much as 11% extra return or as little as 0%. Wide error bars.
  4. Even if we have good reason to expect EAs to do better at picking startups than top-quartile VCs, it seems unlikely that they could perform much better. Narrow error bars.
  5. Liquidation preference matters relatively little. Narrow error bars.
  6. The concept of meta-options is complicated and has received hardly any attention. My best-guess estimate for their value is very large, but I could be way off. Extremely wide error bars.
  7. Future performance is really hard to predict, even over long time horizons. Wide error bars.

By my estimate, startup employees' expected returns could optimistically be as high as 52% (!); or they could be as low as -2%.[12] (Remember, these are expected returns. Realized returns could fluctuate by much more than this. Startups in aggregate could easily realize a 100% return next year, and I wouldn't find that surprising, but I would be crazy to expect it to happen.)

Risk and correlation of startups

Expected return alone isn't what we care about. We really want to know risk-adjusted return.

And we don't just care about the risk-adjusted return of startups in isolation. We want to know how they fit into a broader EA investment portfolio.

There are two equivalent ways of looking at this:

  1. Find the risk of startups, and their correlation to the aggregate EA investment portfolio. Then we can calculate whether EAs on the margin should work at startups instead of big companies.
  2. Find the alpha of startups relative to the EA portfolio. If alpha > 0, that means at least some EAs should work at startups.

I will focus on the first because I find it more intuitive. I also calculated the second and got similar results (not presented in this essay).

I'm looking at the risk and correlation of the startup industry, rather than the average risk/correlation of a single startup. We can think of it as a collective decision by many EAs to work at a diversified group of startups, rather than the decision of a single person.

As with expected return, we have no way to know the future risk of startups, or their correlation to the EA portfolio. But with risk and correlation, we get to make some simplifying assumptions.

We can't learn much about the future return of an asset class by looking at its past return. Markets are reasonably efficient, so if an asset class performs well, more money floods in and performance reverts to the mean. But the efficient market hypothesis doesn't say risk mean-reverts. Studies show that, at least for public equities, historical volatility is a pretty good predictor of future volatility (e.g., Dreyer & Hubrich, 2017[13]).

Three different data sets all give similar(ish) numbers for startups' standard deviations:

Data Set (Gross) Standard Deviation
Cambridge Associates, 1988–2016 21.4%
Cambridge Associates, 1995–2018 23.5%
Harris et al., 1984–2014 28.5%

Note that, according to Cambridge Associates, top-quartile VCs have higher standard deviations than average VCs (25% or 28%, depending on which time horizon you use). So if we only work at top startups, we should bump these numbers up by a few percentage points. Also, startups don't have public prices the same way stocks do, and VCs have some leeway to value their portfolios however you want. I expect that they tilt their portfolios toward low volatility to make themselves look better, so the "true" volatility is probably higher.

Woodward (2009)[14] argues that, because startup valuations tend to lag the market, a naive regression doesn't show the true relationship between startups and public equities. The paper finds that startups have a stock market beta of a little over 2, which corresponds to a standard deviation of about 35%.

For correlation, as with standard deviation, we can assume the future looks the same as the past. Historical correlation between startups and public equities was around r=0.7. (My own analysis found a correlation of 0.6, and Woodward (2009)[14:1] found a correlation of around 0.7–0.8 using more limited data but better methodology. Woodward's analysis suggests that the naive approach underestimates the true correlation. So let's use 0.7.)

Leverage

Many EA investors probably want to use leverage. But startup employees can't leverage their equity: they get however much they get based on their employment contract, and there's no way to borrow money to get more equity.[15] Instead of comparing startup equity to a public investment portfolio, we should compare startup equity to an optimally leveraged public investment portfolio (taking into account that leverage typically costs more than theoretical models assume).

Startups vs. public equities

Now that we've discussed the main considerations, we can return to the original question: is it better for earners-to-give to work at high-paying companies and invest their salaries in the market, or to work at startups and "invest" in startup equity?

Some additional assumptions:

  1. Our goal is to maximize the geometric return of the overall EA investment portfolio. (This is consistent with logarithmic utility of money.)
  2. We can only invest in two things: public equities or startups.
  3. We control 1% of the EA portfolio. We can't affect the other 99%.
  4. EAs currently invest all their money in public equities, and none in startups. (The latter is obviously false, but it's also sort of true: on the margin, earners-to-give can consider working for startups that don't already have any EAs working for them. That set of startups has 0% EA investment.)
  5. If we buy public investments, we can use up to 2:1 leverage.[16]
  6. Public equities earn an expected real return of 3% with a standard deviation of 16%.[17]

Recall from above that we're giving startup equity an 18% expected real return, a 35% standard deviation, and a 0.7 correlation to public equities.

Our two choices:

  1. Work at a big company. Invest our salary in public equities with 2:1 leverage.
  2. Work at a startup.

Given all the stated assumptions, working at a startup is more than four times better than working at a big company (37 expected utility vs. 200 expected utility, according to a scaled logarithmic utility function[18]).

Suppose we hold everything else constant but reduce the expected real return of startups. The return needs to be as low as 1% before the startup looks like the worse choice. (Notice that that's lower than the 3% expected return of the public stock market, even before accounting for leverage. Startups with a 2% expected return are still (barely) preferable to public equities with a 3% return because we're assuming startups have a lower correlation to the EA portfolio.) So even under a much more pessimistic projection for startup returns, they still look preferable to big companies.

Startups vs. an optimized public investment portfolio

Buying an index of public equities might not be the best way to invest one's big-company salary. I personally prefer to invest in concentrated value, momentum, and trend strategies. Some EAs believe cryptocurrency or AI stocks will beat the market. The specifics don't matter too much. What matters is that if you believe some other investment has substantially better expected performance than a broad index fund, then you should use that other investment as your benchmark instead. And startups need to look better than that benchmark.

My best guess is that a concentrated value/momentum/trend portfolio will earn an expected real return of 6% with a standard deviation of 11%. (Of course, as with my estimates for startup returns, these numbers are not remotely robust.) If we also use 2:1 leverage, then value/momentum trend still looks somewhat worse than startups, although not by a as big of a margin (126 expected utility points vs. 200). If startups returned 10% instead of 18%, then value/momentum/trend would be the better choice.

Alternative: Predictionless approach

Alternatively, take the same basic model as above, but don't try to predict the future. Instead, assume asset classes will perform exactly as well in the future as they performed in the past. As I discussed above, this approach has issues—performance fluctuates a lot over time, so past performance doesn't tell us what will happen in the future. But there's also something appealing about this method. Trying to predict future performance leaves lots of room for you to bias the outcome toward what you (perhaps subconsciously) want. It's harder to introduce bias if you just use past performance.[19]

For the predictionless approach, I estimated the expected return to employee equity as:

  1. 13% after-fees historical return to VC firms, or 10% after inflation
  2. 15% real historical return before fees
  3. Subtract 2% for employee liquidation preference, giving 13%
  4. Add 11% for meta-options, giving 24%
  5. Add 4% for the persistence of top-quartile VCs, giving 28%

For public equities and for my value/momentum/trend portfolio, instead of making projections, I used the (estimated) historical return after inflation from 1995 to 2018:[20]

Return Std Dev
Public Equities 9% 19%
Val/Mom/Trend 15% 12%

The three choices have the following expected utilities:

  1. Public equities: 142 utility
  2. Val/Mom/Trend: 288 utility
  3. Startups: 279 utility

Startups look preferable to public equities, but slightly worse than value/momentum/trend.

Alternative: Models are bad. What if we don't use a model?

I love using quantitative models like the one in this essay. I think more people should use them. But most models are bad, including mine. They depend on lots of assumptions, and you can change the model output by making small changes to the assumptions.

How could we reason through this decision without using an explicit model? Let's review some arguments, both pro- and anti-startup.

Argument from risk preferences: Most startup employees don't want to donate all their equity. That makes them much more risk-averse than EAs who work at startups. If they're acting rationally, we should expect them to demand higher equity to compensate for the risk. Therefore, startup equity should look particularly compelling to EAs.

Argument from inefficiency: The market for startups is illiquid and has high barriers to entry. We might reasonably expect it to be less efficient than public markets, which means we have a better chance of identifying startups that will outperform.

Argument from investability: The most reputable VC firms usually don't accept new investors. Even if they can beat the market, you can't invest with them, so it doesn't matter. But there's nothing stopping you from getting jobs at top VCs' portfolio companies.

Argument from overpopularity of startups:

Argument from underappreciated risk: In my experience, almost nobody understands how risky individual startups are. Even medium-sized companies are about 3x as risky as the S&P 500. I don't have sufficiently granular data on startups, but startup-sized public companies are about 5–6x as volatile as the S&P 500, and my guess is startups are even worse. When I see people discussing the value of startup equity, they almost never properly account for this.

Argument from diversification: If you get a job at a startup where no other EAs work, you're adding an entirely new investment to the EA portfolio. That could be a good thing even if that particular investment has a worse expected value than the market. On the other hand, there are other ways of diversifying that might be better.

These qualitative arguments don't obviously lean one way or the other. My intuition from my time working at startups and knowing lots of startup employees is that most people overvalue startups and underestimate risk, which means they probably push down the market rate for equity compensation. But even if most startup employees don't behave consistently with their personal risk appetite, they still might behave more risk-aversely than EAs ought to.

Practical details

If more EAs want to work at startups, there are some ways that people or organizations could support this effort, such as:

Some of these ideas are logistically difficult, maybe even impossible. I'm not sure the best way to provide support for earners-to-give who choose to work at startups, but it's something to consider. I believe it would be valuable if an organization existed that helped EAs with these practical details.

Conclusion

Assuming my model is approximately correct, what type of person might want to work at a startup?

Who might not want to work at a startup?

My analysis suggests that working at a startup has good expected value under ideal conditions. If you get a job offer from a startup, remember to pay attention to the specifics of the offer:

  1. Is your total compensation competitive with what you'd get at a big company? (Taking startup equity at face value)
  2. Does your equity contract include any sketchy terms?
  3. etc. (Too many specifics to list all of them)

Areas for further research

Many subjects warrant a deeper investigation:

Acknowledgements

Thanks to Linchuan Zhang for commissioning this research project and providing support. Thanks to Charles Dillon for feedback.

Changelog

2024-04-25: Upgraded confidence from "somewhat likely" to "likely" based on the fact that readers did not spot any major flaws in my reasoning, and based on calculating the value of meta-options using Black-Scholes (see here for explanation) which roughly agreed with Ben Kuhn's script that I'd used previously.

Appendix A: Startups for founders and investors

This essay has looked at startups from the perspective of employees. How do startups look for other types of people?

Founders: Similar to employees in many ways. The upside is you get a lot more equity. (Hall & Woodward (2012)[21] found that VC-backed startup founders on average made much more money than salaried employees.) The downside is you have to actually start a startup, which is much harder and may require an entirely different skillset.

Other people have written about whether EAs should start startups:

VC limited partners: If you give your money to a VC firm to invest, this is probably worse than being a startup employee (although it does have the advantage that you don't need to get a new job). You have to pay VC fees and you don't get meta-options. You do get a better liquidation preference, but that's usually not worth as much. For a more detailed discussion on investing in VC, see Mulcahy et al. (2012)[22].

Angel investors: If you become an angel investor, you don't have to pay VC fund fees, but you do have to evaluate startups on your own.

Appendix B: Some important tangents

The points below are all important, but they distract from the thesis of this essay, so I'm not commenting on them in detail.

  1. Within the context of my model, some big companies' compensation packages behave more like startups'. Companies such as Facebook and Google offer equity to employees. Unlike with startups, you can sell the equity as soon as you get it. But you usually have to wait a year, and a big company's equity grant still behaves like a meta-option.
  2. There are many non-monetary pros and cons to working at a startup. For instance, see Big companies vs. startups and Early career EA's should consider joining fast-growing startups in emerging technologies.
  3. An unimportant point that I nonetheless want to address: Startup employees' equity will get diluted by future fundraising rounds. This doesn't matter because VCs will get diluted by the same amount, so it doesn't make employee equity look worse relative to VC equity (unless the employment contract contains sketchy terms around who gets diluted, in which case maybe you shouldn't work there). Normally, VCs have the option to invest more money in future rounds to negate their dilution, but this also doesn't matter because it doesn't change the return on their initial equity purchase.

Notes


  1. The report includes VC returns up to 2020, but it only includes detailed data up to 2018. So when I did my analysis, I used the 1995–2018 data. ↩︎

  2. Somewhat concerningly, these two data sets show different numbers even for years where they overlap. For example, the 1988–2016 data set quotes a 60.09% return for the year 1996, whereas the 1995–2018 data set claims a 63.46% return for the same year. This discrepancy is at least partially because the 1995–2018 series includes more VC firms, but I haven't read the Cambridge Associates reports in enough detail to say if that's the only reason. ↩︎

  3. Koch (2014). The risk and return of venture capital. ↩︎

  4. Woodward (2009). Measuring risk for venture capital and private equity portfolios. ↩︎

  5. Harris, Jenkinson, Kaplan & Stucke (2020). Has Persistence Persisted in Private Equity? Evidence from Buyout and Venture Capital Funds ↩︎ ↩︎

  6. Figures are copied or inferred from Harris et al. (2020), Table 1, Table 2, and Table 4. ↩︎

  7. Kaplan & Schoar (2005). Private Equity and Performance: Returns, Persistence, and Capital Flows. ↩︎

  8. At least in the United States, if you donate stock to charity, you can only deduct up to 30% of your income. If you get lucky and make a bunch of money when your startup exits, your startup equity could account for something like 90% of your income. You can only deduct 30%, so you're stuck paying taxes on the other 60%. ↩︎

  9. Very roughly, a startup has a 70% chance to be worth 0x, 20% chance of 0–1x, and 10% chance of >1x. Liquidation preference only matters in the 0–1x case, where common shares are worth about half as much as their face value.

    If you don't filter out sketchy terms, the appropriate discount is more like 36%. ↩︎

  10. Actually, I know a few people who I believe could do a good job of identifying top startups if they took the time to conduct lots of interviews and due diligence. But they're not going to do that because they're busy doing other important EA-related activities. ↩︎

  11. Chingono & Rasmussen (2015). Leveraged Small Value Equities. ↩︎

  12. For the optimistic estimate, I assumed: top-quartile VCs' average out-of-sample return of 26% (or 23% real) fully persists; EAs perform 5% better than top VCs; and meta-options are worth 20% (which follows from optimistic assumptions about how meta-options behave).

    For the pessimistic estimate, I assumed: public equity valuations fully mean revert, and startups perform even worse due to higher risk; top quartile VCs' returns do not persist at all; and meta-options are worthless (probably because there's some flaw with them that I haven't thought of). ↩︎

  13. Dreyer & Hubrich (2017). Tail Risk Mitigation with Managed Volatility Strategies. ↩︎

  14. Woodward (2009). Measuring Risk for Venture Capital and Private Equity Portfolios. ↩︎ ↩︎ ↩︎

  15. Technically, if you get employee stock options, your equity is leveraged. But the amount of leverage approaches zero as the stock price increases. And most stock options are only a little bit leveraged to begin with. For example, if your company stock was last valued at $4 and you get options with a $1 strike price, that's only 1.33:1 leverage. If the stock price doubles to $8, now you only have 1.14:1 leverage. ↩︎

  16. In keeping with my previous work on leverage, I assume that borrowers must pay 1% plus the risk-free rate. ↩︎

  17. This is the average of the projections by AQR and Research Affiliates as of October 2021. ↩︎

  18. The utility function takes the geometric mean return as the utility and multiplies by 100,000 to make the numbers more readable. As a baseline, it calculates the expected utility of the EA portfolio without your investment, and then subtracts that from the total expected utility of the EA portfolio including your investment. ↩︎

  19. Still not impossible, because you could pick the data set or the time series that most closely matches the outcome you want. ↩︎

  20. For public equities, I used the historical return of the total US stock market, assuming zero fees or trading costs. To find the historical return of Val/Mom/Trend, I created a hypothetical portfolio that invested 80% in Alpha Architect's Value Momentum Trend Index and 20% in AQR's Managed Futures Index, which roughly reflects how I actually invest my money. Both indexes subtract estimated fees and trading costs. The historical returns are hypothetical, not actual. I didn't have data for 2018, so I calculated summary statistics over 1995–2017 instead. ↩︎

  21. Hall & Woodward (2012). The Burden of the Nondiversifiable Risk of Entrepreneurship. ↩︎

  22. Mulcahy, Weeks & Bradley (2012). We Have Met the Enemy...and He Is Us: Lessons from Twenty Years of the Kauffman Foundation's Investments in Venture Capital Funds and the Triumph of Hope over Experience. ↩︎


Linch @ 2021-11-13T01:29 (+27)

Thank you so much for this post! Though more research can be helpful (as you noted), under ideal conditions, between the analysis in this post and Bill Zito's more heuristic-y arguments about career capital, the case for tech earning-to-givers working in top startups rather than in BigTech is basically overdetermined.

Speaking for myself, if I was still working at Google and didn't see either  a clear path to much more specialized career capital at Google or a path to direct work elsewhere, I'd consider this analysis and surrounding arguments more-than-sufficient (85%?) to switch into startup work. (And I in fact did join a fast-growing startup in late 2019 based on much vaguer and lower-quality analyses than available now, which has empirically worked out fine for me^).  

I'd be very excited for more EAs to consider top startups, though as you've noted, there are a number of practical details which anybody reading this post and thinking of changing jobs should do well to read or internalize. 

I also think institutionally EA should figure out ways to make this transition smoother for people. I'd be excited for an entrepreneurial person who doesn't have better projects to do to take this on, though they should of course compare whether doing this is better than their counterfactuals.  

 (I think a reasonable critique of this whole line of research/inquiry is that a) EA writ large already has a lot of money, and b) earning-to-give is probably better in expectation if people made very high-variance bets like trying to start multi-billion dollar companies. I agree with this, but given that many EAs are trying to do tech earning-to-give anyway, helping them serve that need seems good.)

^ I turned out to be a bad fit for the startup, but after I left I got obsessed with working on covid stuff and the career capital worked out well. Also the money situation went unexpectedly well and I don't think I burned down too many bridges.

Jeff_Kaufman @ 2021-11-13T14:14 (+8)

Say a startup offers you an equity package that's worth $X per year at the current valuation. At the same time, a big company offers you a salary that's $X higher than your salary would be at the startup. Both compensation packages have the same face value.

Your post is very detailed so I could have missed where you expand on this, but I think this initial assumption is off. Face value compensation at the top tech companies is generally much higher than what you would get at a startup. Have a look at https://levels.fyi

Natália Mendonça @ 2021-11-13T16:32 (+6)

Face value compensation at the top tech companies is generally much higher than what you would get at a startup. Have a look at https://levels.fyi

It's more complicated than that. Some top startups (e.g. Stripe, Airtable, Databricks, Scale AI, ByteDance, Benchling, and several others) pay at least as much or a lot more than e.g. Google/Meta. Some of those (Stripe, Airtable, Scale AI) seem to offer new grads close to $100k more than Google does on average in the first year (counting signing bonuses, and assuming the valuation of all of those companies doesn't change). Also, Levels.fyi's 2020 report showed that a lot of the top-paying companies were startups. 

But it is probably the case that startups with more room for growth pay much less.

Jeff_Kaufman @ 2021-11-13T16:43 (+8)

Now I'm just confused about what companies were talking about! Stripe is 12 years old, has 4k people, and is worth ~$100B. Is it just that it is still private?

I would have counted it as a big company and not a startup in thinking about this post, but maybe that's not how the author intended it?

MichaelDickens @ 2021-11-13T18:03 (+11)

I would not consider Stripe a startup for the purposes of this post.

Natália Mendonça @ 2021-11-13T18:15 (+1)

I would have counted it as a big company and not a startup in thinking about this post, but maybe that's not how the author intended it?

I thought Stripe would be in the reference class of startups (since it still raises money from VCs), until I read Michael Dickens's reply to this comment. I agree that it was a supremely bad example, though. The other companies I mentioned probably count?

There are also a lot of smaller/newer companies that pay about as much as Google/Meta that I didn't mention in my first comment. They're mostly unicorns (though not all of them are), but I think they might be a substantial fraction of the set of companies people actively trying to work at startups might end up in -- they're large compared to the average startup, but the average startup is less likely to have the necessary infrastructure to absorb more people or recruit in a predictable way, and/or might hire exclusively from its personal networks. 

MichaelDickens @ 2021-11-13T18:01 (+4)

At a glance, I don't see startup salaries on levels.fyi. In my experience, most startups offer worse face-value compensation than large tech companies, but a significant minority offer competitive compensation. I was able to get a (slightly) higher offer from a startup than from Google.

Linch @ 2021-11-13T22:23 (+2)

For the record, this was true for me as well.

G Gordon Worley III @ 2021-11-13T22:15 (+4)

My own experience is that there's a sweet spot. Big tech companies only really offer high compensation for the most experienced and capable employees. If there's 10 levels and you're not at least at level 8, a big company is probably not, in my own informal analysis, like to offer you the best compensation in expectation. Some of this is simply because these folks have high opportunity costs, and the only way to get them as employees is to pay them enough that it balances off against what they would likely do instead: start a company.

If you're in the middle say 4-7, then a large, succeeding startup is probably the best bet. It offers better pay, more room for advancement and promotion, and decent equity.

If you're at the bottom, especially say because you're new to work, then early stage startups can provide really great returns in expectation. This works a couple ways. You won't make a lot of cash compensation, but you'll earn a lot of equity in expectation, possibly more than $10mm a year if the startup becomes a unicorn. Beyond that, you'll gain a lot of career capital by getting to do a bit of everything and having to operate fairly independently in ways that you won't get to do in a larger company, which means you'll be able to level up faster than you would in a more established place if you apply yourself.

This is all assuming you're best fit to be an employee rather than an entrepreneur, of course.

Peter Wildeford @ 2021-11-13T06:39 (+4)

Great and thorough analysis. Very cool!

One question:

EAs currently invest all their money in public equities, and none in startups. (The latter is obviously false, but it's also sort of true: on the margin, earners-to-give can consider working for startups that don't already have any EAs working for them. That set of startups has 0% EA investment.)

My understanding of the EA portfolio is that it is currently heavily invested in crypto, FTX, Facebook stock, and Asana stock. Does that change the expected returns? More importantly, does that change the correlations? I imagine a lot of tech startups of the kinds EAs might join likely have especially positive correlations with this part of the portfolio.

In fact, it might be interesting for more EAs to find start ups that are negatively correlated with these (e.g., an Asana competitor).

Linch @ 2021-11-13T22:33 (+6)

In fact, it might be interesting for more EAs to find start ups that are negatively correlated with these (e.g., an Asana competitor).

My intuition is that this will be positively rather than negatively correlated fwiw, in the same way that I don't think "make a different crypto exchange" is a good way to hedge against FTX risk.

There might be other ideas in this space however.

Peter Wildeford @ 2021-11-14T01:13 (+2)

Yeah I agree there are likely a lot of sector-based correlations where Asana and its competitors move together in response to broader outlooks about the sector.

MichaelDickens @ 2021-11-13T18:06 (+6)

Yes this is something worth considering. I did look at how much alpha the Cambridge Associates startup data had on top of US publicly-traded tech stocks vs. the US total market, and there wasn't much difference. EA money is in much more specific investments than just the tech sector, but that makes it harder to test the correlation.

Charles He @ 2021-11-13T02:50 (+3)

Another point in favor of *good* startups is that the hiring process is less artificial (e.g.  less likely to be surprised with inexplicably hard leetcode) and less abusive (e.g. less likely someone in the hiring loop presses you to do consulting for them).

At the very least, when either of these situations happens, it will come from the founder/manager or colleagues, which seems less dysfunctional.

Also, I’ve had much better access to interviews/founders at startups than FAANG/Uber/Lyft and genuine work experience seems to be respected more.

However, I know much less than someone like Linch, and would like to hear their opinion.

 

Also, personally, I still feel like this whole “do startups”/”make billion dollar company” sentiment on the EA Forum feels a bit “online”. I am worried it may give the wrong impression to some EAs, especially since I expect startups are the trendy thing already for everyone in the appropriate reference class.
 

MichaelDickens @ 2024-04-26T03:37 (+2)

I had an idea for a different way to evaluate meta-options. A meta-option behaves like a call option where the price equals the current value of the equity and the strike price equals the cash salary you'd be able to get instead.[1]

If I compare an equity package worth $100K per year versus a counterfactual cash salary of $100K and assume a volatility of 70% (my research suggests that small companies have a volatility around 70–100%), the call option for the equity that vests in the first year is worth $29K, and the call option for the equity that vests in the 4th year is worth $56K (which is equivalent to a 12% annual return). So on average, a meta-option on a 4-year equity package is worth somewhere in the ballpark of an 18% annual return.

(But if the equity has a lower face value than the counterfactual cash salary, it pretty quickly becomes not worth it.)

[1] This is kind of wrong because with a normal stock option you don't have to pay the strike until you exercise, but with an employee meta-option, you have to give up your counterfactual salary as soon as you start working, and you don't vest for the first year so you have to give up a full year of cash salary no matter what. If you have monthly vesting, the fact that you have to pay at the beginning of the month instead of the end doesn't matter much.

(edited to make the numbers make more sense)

Yonatan Cale @ 2021-11-15T23:09 (+2)

TL;DR: Personal fit matters a lot and I think we don't talk about this enough

I know this is out of scope for your post, but I thought it at least deserves a comment


Personal Fit is neglected

From lots of discussions with EA software developers, they often don't even think much about optimizing factors such as "how much will I enjoy this"

Personal fit also affects the financial aspect

A lot of the variance in income comes from skill, which comes from

  1. How much do I enjoy my work?
  2. How quickly do I improve at what I do? (This snowballs! A higher rate of improvement matters A LOT in the long term!)
    1. Improvement rate is affected a lot by how much I enjoy my work

I'm not saying you're wrong with your calculations

I'm just saying that other factors may be really important and I wish they got more "air time".

I still appreciate you doing these calculations, I couldn't do them myself, so still thank you

Yonatan Cale @ 2021-11-15T23:14 (+1)

(I'm considering writing more about "textbook optimizations", things that software developers can often do to optimize earnings and other good-things in our jobs, but this seems highly out of scope for your post)

Yonatan Cale @ 2021-11-16T00:01 (+1)

Here are my financial thoughts and questions from your post.

[I wasn't sure how to format this to make it easy to comment on, I'm open to changing it]

1. If I was only trying to optimize earnings for myself (regardless of EA), would you still recommend this? [not sure I understood this part]

Do you think that:

?

If so, then do I understand correctly that the value of working for a startup is:

1.1. Diversifying the EA portfolio, and

1.2. The value of money is not logarithmic for EA as a whole

?

If so,  1.1. seems like a small consideration to me, given the importance of this life decision for the individual, and 1.2. seems like it's maybe false, given EA currently has a ton of money. But anyway, I'd put this explicitly and let each person decide for themselves. (or maybe I didn't understand correctly? This is my biggest uncertainty)

2. Funding options exercising - this exists

There is a company that does this in Israel.

If this doesn't exist somewhere else, perhaps it can be started as a for-profit, I don't think it necessarily needs to be an EA org. (It can donate profits)

3. EAs picking stocks better than VCs

A main failure mode is, I think, developers who THINK they can pick good startups, but they're actually just hearing the same things that all the founders say. "We have the best people", "The market is billions of trillions", and so on. I've been there myself.

I think it's important to put some clear disclaimer here that most developers won't manage to beat VCs (in case they think of trying).

4. Opening an EA VC (because we have smart people)

My default answer would be "if it's possible, why doesn't it exist?", but my actual answer is "if some people think they can do it - I think that would be amazing".

btw it sounds like a dream job to me

5. Liquidation preference and other fishy terms

The real world problem here is that employees won't know about these fishy terms. Do you live in a place where if the founders sign an agreement giving the investors a 10x liquidation preference, the employees will surely hear about it? I think it's not the situation in Israel [but this is not my expertise]

6. Meta Options are worth an extra 16 percentage points

First of all, this is really interesting, I have been struggling to build some model around it and I can't wait to see what was proposed.

6.1. My push back: 

At least here, startup workers (and especially founders!!) work a lot more than big-tech workers. I hear people saying "it's hard so some people won't want it", but I don't think I saw anyone put a number on that "hard" figure.

Do you think it's reasonable to say that a startup employee works 20% more than a big-tech employee? And founders - way more. (I wonder if a better metric would be "count how much free time they still have" to express that if founders work 13 hours per day instead of 9, they might be working "double" in some sense)

6.1.1. Your push back (?)

I imagine you'd reply and say "those estimations for how much time people spend working at different companies varies a lot between companies, way more than it does between startups-as-a-whole and big-companies-as-a-whole".

I'd agree.

My point is "there might be even bigger considerations, like how much time you spend at work, which weren't counted", and I'm saying all this because I was surprised to see "16 percentage points" as the value that is so big that it seems to tip the scale.

Don't get me wrong - it is significant, and I'm especially happy to have it modeled explicitly. Maybe I could have done a better job setting up my titles here, sorry

7. Who are the top VCs?

If I understood correctly from you: "who is the top VC" changes over time, and the parameter that has predictive power is "who was the top VC last year". I'd find it useful to have an updating list of top VCs. Beating the market is always fun!

(Or is Y Combinator always on top?)

8. Transferring the risk / helping EAs who's startup failed

If the global EA org is more risk tolerant than individual EAs (who need money to live), and also startups are high-expected-return but also high-risk, then maybe we can use existing financial tools to transfer this risk: 

We could have the EA org pay money to the employee, and get the employee's options in return (early on, way before an IPO/exit). [Would this be legal? Why doesn't it already happen?]

This will also allow the EA org to run the complicated calculation of "which startup options to buy". Perhaps you could be part of this org, and use calculations such as you did here to find the most promising startups, and use EA employees to get access to their options. I think there's value in having some centralized org with expertise and skin in the game  run this calculation.

9. "Most people overvalue startups and underestimate risk"

This is absolutely my impression, I just wanted to say how much I agree.

I don't think we can learn much about the "free market" value of startup options by observing that individual developers want these options very much. 

I don't know if this seems obviously true to you: You do sometimes compare how much EAs would value startup options compared to other people. I agree the comparison holds, but I don't think "the other people" are a reasonable baseline. Let me know if we seem to disagree here, I could also elaborate

Linch @ 2021-11-16T15:01 (+2)

Do you think that:

  • Optimizing earnings for myself = work at big tech
  • Optimizing earnings for EA = work at startup

?

What does "optimizing" mean in this context? If I understand Michael's model correctly, from a monetary perspective, if you're trying to make as much expected money as possible, you should on average work at a top startup. (This is outside the scope of his model, but) if your utility function is fairly risk sensitive, you should on average work at bigtech. From a career capital perspective, it again depends on a lot of details (as I'm sure you know! :)). 

If so, then do I understand correctly that the value of working for a startup is:

1.1. Diversifying the EA portfolio, and

1.2. The value of money is not logarithmic for EA as a whole

?

If so,  1.1. seems like a small consideration to me, given the importance of this life decision for the individual, and 1.2. seems like it's maybe false, given EA currently has a ton of money

I think the diversification aspect was a relatively small fraction of the gains here (though Michael can feel free to disagree with me). I'm not sure what you mean by "The value of money is not logarithmic for EA as a whole" because there might be too many double negatives, but I think Michael is in fact assuming a logarithmic utility function of money for EA. (if it was linear, reducing correlation by diversifying the porfolio wouldn't be useful). However, the value of money to EA of your donations is close to linear, for the simple reason that EA already has a lot of money and we assume utility functions are approximately differentiable (i.e. locally linear).  

(It's possible this is what you're saying and we're just aggressively agreeing!)