Paul Graham is right (using AVC’s data) | Hacker News Reader

Paul Graham is right (using AVC’s data)

zachaysan.tumblr.com

69 points3pt1415926 comments

I don't really understand what the debate is about. Even if the market is totally 'bimodal' valuations matter the same way as in a non bimodal market. Expected return is orthogonal to expected Risk.

Let's say the market is so bimodal that a startup fails or becomes google. Then the expected return is:

p * G

where 'p' is the chance of becoming Google, 'G' is Google's value.

Basically investors' bid for valuation is their bid for what p is.

The story is only that if they say that the company's value is X while the startup says it is 2 * X that means that they say p = p1 while the startup thinks p = 2 * p1

Of course valuations matter. It matters linearly. In any market. Even in the most risky markets. I can decrease my risk by diversifying my investments into lots of different high risk assets and my expected return remain the same, but I can only increase my expected return by buying in at low valuation. This is exactly what YCombinator does. This is just very basic math.

TL;DR: The fact that investors care about valuations does not mean they don't recognize that the success distribution is bimodal.

Yes, exactly right, except with two caveats.

Doubling the valuations means you get to make half the number of bets--thereby halving your expected returns--except...

(1) ...if the price of finding a given start-up to invest in is significant. You've already paid the search cost before you decide on the valuation.

Extreme example: It costs $100 to find each start-up, and you are arguing over a valuation of $1 or $2. Paying $1 instead of $2 won't let save you enough cash to find another start-up.

(2)...if the market cannot be approximated as infinite because there are only a couple of Googles. In other words, when valuations drop to half, you can only double the number of bets if there are other similar bets to be made.

So when Paul Graham says

>Did they not understand that the big returns come from a few big successes, and that it therefore mattered far more which startups you picked than how much you paid for them?

He's either mistaken or (more likely) he's implicitly suggesting that one or both of the above factors are very strong. But if so, why not be explicit about it in his original piece?

I think even Kleiner Perkins would rather have invested in Google at a $7.5 million valuation than a $75 million valuation as they did. However, negotiating hard on valuation can cause you to lose the opportunity to invest, and the people who argue "bimodal" are saying that this is a poor strategy since it may lock you out of the best deal of the decade just for a few quibbling valuation points.

If p is bimodal, then p = 2*p1 is not possible. Either p = 0 or p = 1. The values in between do not occur.

Then the only thing to quibble over is G. But if you believe p = 1 for a company, what's the point of quibbling over G?

No. The event being bimodal only implies that the probability is centered around two values, not that p =1 or p = 0 for those values. For example, take a biased coin that turns up head once in a million. This is bimodal with p(head) = 10^-6.

PG's argument appears to be that the error in your estimate of p is almost always larger than the value of p and one should invest as long as 2*p is also within the estimation error.

> PG's argument appears to be that the error in your estimate of p is almost always larger than the value of p and one should invest as long as 2*p is also within the estimation error.

I don't think this is correct under canonical Bayesian reasoning. All uncertainties (whether because of ignorance or "objective uncertainties") can be bundled into your probability assessment. There is no "error" on your probability.

I'm not expert though. I'm only....80% sure.

Interesting chart, thanks for presenting. A few issues with your methodology:

1. I'm sure compounded returns would be nice to do, but you don't know the start date for any of those investments. A 2004 vintage fund does not mean the investments are from 2004. Instead, it means the fund made it's first investment in 2004 and will continue deploying money over ~7-8 years.

2. In the VC industry, a wipeout is a wipeout. Call it Zero, not 0.5x.

3. LPs care more about cash-on-cash multiples, not compounded returns (aka IRR). There's an expression that will help explain this point: "You can't eat IRR." VC is a 10-year horizon business and GPs call do special moves to improve the IRR (calling down capital as needed, and then returned capital upon exit even if it's going to be called down again and reinvested). But, at the end of the day, the best measure of the success of any fund is the gross cash-on-cash multiple of the entire fund.

These types of returns and their distribution are surprisingly seen across every industry, no matter the scale you're looking at. For example, if you had invested in all the top car companies that existed in the early 1900's, you'd have lost money on most and only gained money in Ford, GM, Dodge, and overall you would only make a reasonable about of money. Warren Buffett points out that this type of behavior can be seen across every revolutionary business in the past 100 years - automobiles, airlines, radio, television, computers, and now we're seeing another example of this.

On a larger macro-scale, this type of return distribution can be seen in the S&P 500. Over the past 20 years (1990 - 2009), the S&P 500 returned 8.2%. However, if you had pulled your money out of the market on the 10 best single days of those 20 years, your return would drop to 4.5%. If you took out the best 30 days from the past 20 years, you'd have a 0% return! Imagine that, 30 days out of 20 years, and it would cost you all your gains. The same pattern can be seen over any time period for the S&P.

The lesson in all of this is that diversification is the key to consistent returns.

> However, if you had pulled your money out of the market on the 10 best single days of those 20 years, your return would drop to 4.5%. If you took out the best 30 days from the past 20 years, you'd have a 0% return! Imagine that, 30 days out of 20 years, and it would cost you all your gains.

I suspect that the 10 and 30 worst days had comparable effect, albeit in the other direction.

You can't win if you don't play, but neither that nor the dominance of "big win days" implies that you should play every day.

Exactly right. You remove the 10 worst or 30 worst days, your returns would be way above the 8.2% average. The point is, no one knows ahead of time which days will be one of the 10 best, and which will one of the 10 worst. Same thing with angel investors - you don't know which startups will be home runs and which ones will be strike outs. The data suggests you should play every day.

  >> you'd have lost money on most and only gained money in Ford, GM, Dodge, and overall you would only make a reasonable about of money.

That is an argument in favor of the efficient market theory. The car companies were correctly priced, based on the market size that resulted and not knowing which car companies were going to make it.

The efficient market theory certainly breaks down. For example, at the height of the dot-com bubble, many of the dot-com companies were valued as if it were likely they were going to take a very large share (if not most) of their intended market.

The efficient market theory doesn't say that the market predicts winners. It says that the market reflects all known (or knowable, depending on version) information. Actually, it doesn't say exactly that either, it says that prices reflect valuation (which is affected by knowledge) and resources.

Sometimes "we" don't know, or the people who do know don't have enough resources to push the rest of us to the appropriate conclusion.

If you double the valuations, the # of bets halve. This affects your likelyhood of getting ANY wins.

My own angel portfolio: since 2006: 46 deals, three possible winners, four dead, one weak exit.

If you double the valuations, the # of bets halve. This affects your likelyhood of getting ANY wins.

I would expect this is only true if one "must" have a certain percentage of a startup. It makes some amount of sense for traditional VCs who insist on taking a board seat, but I though this was a differentiating characteristic of angels.

you are typically limited by the % the company is willing to sell.

higher valuation deals tend to raise more.

> you don’t let the really good ones get away because they are asking twice as much as you were expecting

The problem is that to actually apply this strategy you have to overpay for everything, because you don't know which ones are the really good ones until after the fact.

Ah, but you don't overpay for the failures, because they don't take more of your money. Instead, you invest the same $500,000 and get a smaller percentage of ownership. But in all of the 0x and 1x return cases, it doesn't matter what percentage of the company you own.

For the outright losses this is true. But overpaying can transform the moderate wins into losses. Buy in at 1M sell for 10 you get a 10x return. Buy in at 20, sell for 10, you get a 50% loss. So if you're playing the middle (which is where the reliable returns are to be found) valuation matters. A lot.

Of course this data is presented in hindsight. If expected value of a return on a company is the probability it's the next G times the percent you own:

E[Return] ~= E[Company] X %Owned

And

E[Company] = P(Company=NextGoogle) X Value(NextGoogle)

If you invest in a portfolio of companies, then you'd try to control the things you can:

1) P(NextGoogle): Impossible to estimate, look for good founders

2) Value(NextGoogle): maximize this (look for big markets)

3) %Owned: maximize this

Note that YC does (1) and (3). I would argue they don't do (2) at all. Noone thought AirBnb would be as big as it is, but they're crushing it.

I disagree with the conclusion that the return distribution is bimodal. My own analysis* is that it is instead lognormal. This shape makes sense to me - a lognormal distribution is the result of the product of random variables (whereas a normal dist is the sum of randomly distributed variables). The success factors in a startup tend to be multiplicative, not additive (e.g. great market, great product, customer adoption, great strategic partners, etc).

* based on observed portfolio performance which is unfortunately not public information, but also on published data from mckinsey

Beat me to it: http://news.ycombinator.com/item?id=1847466

Your graphs don't back up your headline.

Fred thinks his returns are trimodal, and that he plays mainly in the middle. Looking at your chart, he seems to be 100% correct. He's got one exit at 70% annual returns, but he makes the bulk of his money in the middle of the range.

that's right

i can't comment on that post, so i will do it here

if we toss out our two biggest winners, we will still have a terrific fund

of course, i am thrilled to have them

but they are not required to have a good fund

But what if the venture capital model creates a situation where outcomes must be bi-modal? Consider the case where a CEO goes to a VC (either on the board or in the fund raising process) with 2 plans

Plan A has a ~90% chance of slow growth Plan B has a ~10% chance of explosive growth

The VC will point you towards plan B, because plan A isn't a win given their investment structure.

If this is true, then you cannot conclude that investments are always bimodal, just that in the past they have been due to structural reasons. And this is relevant because investors could create a model where they foster and profit from non bi-modal companies.

Would the amount invested strengthen the chance of a good outcome?

a compounded 20x return does not validate the following:

"The expected value of a startup is the percentage chance it’s Google"

Reply on news.ycombinator.com