Uncertainty Wednesday: Valuations and Inflation Risk

Today’s Uncertainty Wednesday is about the risk of inflation. I spend much of my time thinking about company valuations, both as we make new investments and also as we try to mark valuations in existing companies. Of course valuations for private companies are influenced by available public comps and those have been trading at high multiples, including the current crop of IPOs. Now there are two countervailing views in the market. One view is that valuations will return to lower historical multiples and the other view is that current multiples are the new normal.

Why could multiples have expanded permanently? Well there are two possible explanations (which are not mutually exclusive). First, that due to pervasive network effects successful companies can grow faster and have stronger ultimate positions in their respective markets. This is to say the higher multiples are justified on fundamental grounds. Second, that there is so much money sloshing around causing low interest rates and thus driving up present value of growth companies. This is to say the higher multiples reflect a changed monetary landscape. The place to look here is the Fed’s balance sheet which expanded massively during the 2008 financial crisis and again in 2014 as can be seen on the following graph.

Now what that graph also shows is that for about a year now the Fed has had some success in bringing that money back in by shortening its balance sheet to the tune of a few hundred billion dollars. That’s a huge amount of money but still pales compared to the roughly $3.5 trillion expansion.

What does all of this have to do with inflation you may ask. Well, the multi trillion dollar question is where we go from here. Can the Fed successfully continue to shorten its balance sheet, or is the following projection from JP Morgan Asset Management correct?

This basically suggests settling in with an extra $3 trillion or so and growing from there. It reminds me of what someone said as the Fed started with quantitative easing: “The Fed’s balance sheet is like toothpaste – easy to squeeze out, impossible to put back in the tube.” (I can’t find the quote right now to attribute).

If that’s indeed the case then it is true that asset valuations may stay as high as they are (to the extent that the monetary explanation for high multiples applies), but it also means that there is the potential for an inflationary shock across all prices, including everyday consumption items. I have written a lot about inflation before where I keep making the point that technology is deflationary. But that doesn’t rule out inflation from a loss of faith in purchasing power.

There is a small but nonzero chance that’s where we could wind up, especially if the collective belief suddenly shifts to concluding that we will use inflation exclusively to finance government spending. Now I should be clear that I am actually a fan of using the money supply to finance a Universal Basic Income, but that requires a shift to full reserve banking and the addition of some kind of wealth tax to rebalance the system. It could all be done, but only if we are willing to change a lot of things instead of just printing more money.

Uncertainty Wednesday: Pascal’s Wager and Impeachment

I have been enjoying writing about current events as part of Uncertainty Wednesdays, such as the Notre Dame fire and the Boeing crashes. Today’s edition is about impeachment following the Mueller report. One of the great early thinkers about uncertainty was Blaise Pascal who made many important contributions to mathematics (among them Pascal’s triangle which is a staple of learning math).

In a posthumously published work Pascal sets out a simple argument for believing in God which has become known as Pascal’s wager: a belief in God imposes a finite cost during one’s lifetime but with some small probability offers an infinite benefit of an eternal heavenly afterlife (he was thinking of a Christian God but the argument applies to many other religions that have a similar concept). Conversely if you do not belief in God you have a finite gain versus a small probability of an infinite loss from an eternity in hell. With this setup the rational choice is to believe in God.

Pascal’s wager is an important contribution to thinking about living under uncertainty because it separates out the probability of states of the world (God exists, God does not exist) from the payoffs of different decisions when those states are realized. In doing so it makes clear that when payoffs can be very large in either direction (here infinite gain or infinite loss), then even small probabilities have a huge impact on decisions. Importantly no amount of prior research or forecasting can meaningfully resolve this analysis. Pascal’s wager can thus be seen as an early (maybe the earliest?) contribution to a line of thinking about the importance of extreme payoffs that today is being carried on prominently by Nassim Taleb.

What does any of this have to do with the question of impeachment? Well, one analysis of whether or not the Democrats should move to impeach Trump in the wake of the Mueller report has centered around the impact on the 2020 presidential elections. Some make the argument that impeaching now reduces the Democrats chances of winning in 2020. But if you look at Pascal’s wager, let’s say you find some evidence that give you a small reduction in the probability that God exists, that still does not change the decision because the payoff is so extreme.

I believe a similar logic applies here if you consider the payoffs to be the longterm consequences for limits on the behavior of all future presidents. You can either impeach or not and then the president either wins re-election or does not. If you impeach there is a small cost if Trump is reelected but a massive gain if he is not. Conversely if you do not impeach there is a massive cost if Trump is reelected and a small gain if he is not. Why? Because impeachment is about congressional oversight of the behavior of a president. It is a crucial part of the checks and balances of the system of government in the United States.

The crucial part here is that this logic is independent of whether you think impeachment will succeed in removing the president. In fact, I assumed that it will not succeed and that Trump will be able to run in 2020 in any case. Another way of phrasing the analysis is as follows: do not give up on a fundamental principle for tactical gain (here the fundamental principle being congressional oversight). The fundamental principle has huge (infinite) upside/downside whereas tactical gains are always small by comparison.

Uncertainty Wednesday: Learning from Notre Dame

It was upsetting to see Notre Dame burn on Monday, as the building is a beloved landmark to the many millions who have visited Paris and enjoyed a stroll near its majestic presence. In addition it is also a major symbol for many French and an ongoing place of religious worship. It now appears that reconstruction is not only possible but will also be well funded (prompting some interesting debates about fairness).

In today’s Uncertainty Wednesday, I want to focus on an important lesson from the disaster. The apparent cause for the fire was an ongoing renovation project. Notre Dame was in pretty good shape for a building of such an advanced age. That means interventions have limited upside (marginal improvement) but, as we have seen, dramatic downside. Whenever the payoff structure of any activity has this asymmetry of limited upside, unlimited downside it is advisable considering not engaging in it at all, or proceeding with extreme caution.

As it turns out, this structure is pervasive and yet we don’t pay nearly enough attention to it. For a second example, one need look no further than the recent crashes of two Boeing aircraft (see post from a few weeks ago). Any modification to a working aircraft has limited upside and massive downside. In the case of the Boeing 737 MAX 8, the improvement was supposed to provide better fuel efficiency. That’s a worthy goal for sure, but a marginal improvement. Given the downside potential it means either not to do it at all or to be incredibly cautious.

And here is a quick example from the startup world. Adding debt to the financing mix can improve your cost of capital. But it also increases the risk of going out of business entirely. So again we encounter the structure of marginal improvement combined with existential downside risk.

Teaser: next Uncertainty Wednesday we will look at how the nearly same intervention can have a very different payoff structure when the circumstances are different.

Uncertainty Wednesday: Are Our Phones Listening To Us?

From time to time, I see people online posting how they simply talked about something only to be presented an ad about that very thing soon thereafter. This quickly leads into speculation that some software on the phone was using the microphone to listen in on conversations and used that information for ad targeting. There has been extensive discussion online of the technical feasibility and the likely legal consequences. The rough consensus appears to be that it is theoretically possible, has not actually been observed using tools that listen in on traffic and would be a huge legal risk for whoever does it.

In today’s Uncertainty Wednesday I want to take a different approach though. And that’s asking if we should necessarily be surprised by such seemingly eerie ad targeting. Consider a similar coincidence: you meet a stranger, say on an airplane flight (writing this in the air) and you discover after some conversation that you both really like a rather obscure arthouse movie. “What a coincidence” you both say. But is it?

Here we are in one of the weirder aspects of probability that has to do with the frame of reference. Something can be simultaneously highly unlikely and certain. Come again? This seems like a perfect contradiction. Well, consider the type of lottery that one person wins every time there is a drawing. If millions of people participate in a drawing, then your chances of winning are super low (1 in, well, millions). But the chances of someone winning are 100%. For a lottery the distinction between the frame of reference of you, the individual (a sample of size 1), and the lottery as a whole (the population) is self evident.

But your conversation with the stranger on the airplane is similarly just one of tens of millions of people are having with a stranger today. And among that larger number there are bound to be people who discover that they share some obscure interest with the person they are talking. So once you zoom out to the level of all people it is not clear that anything has been learned. Just like there is no information in someone winning the lottery of the kind I described above. We already knew before the drawing that this would be true.

Now let’s come back to internet advertising. Billions of ads are served to hundreds of millions of people every day. Even if there were no targeting of any kind, we should, at the population level expect to see quite a few dead on hits, such as hearing aids being advertised to someone who never before searched for hearing aids but recently mentioned them in conversation. Put differently, the base rate of coincidences in the population is meaningfully above zero. But you only observe what is happening to you. And so for you it feels like an incredible coincidence. Too incredible in fact, which is why you search for some kind of explanation.

Just to be clear: please don’t take this as my saying that “phones listening” is definitively a tin foil hat paranoia. It may well be real. All I am saying is that there are statistical reasons for why people may feel like they have been targeted even though they have not.

PS Apologies for the lack of links, but I wrote this on a flight with super slow internet connection

Uncertainty Wednesday: Working for a Startup

I got the idea for today’s Uncertainty Wednesday from seeing quite a few tweets in my timeline to the effect that startups don’t really have meaningful upside for employees. Tracing these tweets back, led me to this post titled “working for a startups makes increasingly less sense” which has the following quote in it

That startup sold for 200 million dollars and, as the 10th engineer, I made … 15000 dollars from that exit.

First, let me do some math here: 0.015 / 200 = 0.000075 which is 0.0075%. So one of two things have to be true. Either this person got a terrible equity package, or the company had a ton of preferred overhang so that very little of the exit value went to the employees (this can and does happen).

The 10th engineer at a startup might have 0.1% (possibly more, I have seen 0,25% or more). Let’s assume a bunch of dilution along the way to 0.05% at exit. That would still be 200 * 0.0005 = 0.1 million. Now 100,000 dollars isn’t a fortune but it is certainly more meaningful than 15,000 dollars. 

Now you can easily see that these numbers of course move around a lot with the size of the grant, the amount of dilution along the way and the value of the exit. We had a very capital efficient company in our portfolio which only ever raised $5 million of primary capital and exited for $1 billion. So here the math for an early engineer would be 0.001 * 1000 = 1, or $1 million and could be as much as $2.5 million or more (if the initial grant was say 0.25%).

With that out of the way, let me still suggest though that the right way to approach early stage startup equity is to assume that it is worthless. You should not take a job at an early stage startup for any reason other than that you are excited about (1) what the company is working on and (2) the people you will be working with and (3) anything else, such as the specific technologies being used, the opportunity to take a lot of responsibility and so on.

I do a lot of recruiting for the companies that we have invested in at USV. And this point comes up a lot. Especially when people have competing offers. Engineers in particular will try to compare the offers based on some estimate of NPV. I always tell them that this fundamentally the wrong way to think about this choice. We spend the bulk of our waking hours with work, and to make this about the financial upside is ignoring the factors that matter most to long term personal satisfaction.

In the most extreme form people have some kind of scoring system where they assign values to different aspects of a startup, including the potential for financial upside, so that they can rank order their choices. This too is futile. I recently was talking to somebody who was considering an opportunity at one of our portfolio companies and at Facebook who had come up with such a system. Whenever I encounter this now, I point people at this wonderful talk by Ruth Chang, who is a philosopher at Rutgers:

I highly recommend Ruth’s perspective on making the decision whether or not to work for a startup. You are the author of your life: what do you want the next chapter to be about.

Uncertainty Wednesday: Mueller Report Edition (Hindsight Bias)

In the last few Uncertainty Wednesdays, we have been looking at various fallacies arising from our lack of appreciation of uncertainty, such as imperfect correlation (narrative fallacy) and the baserate fallacy. Today we will look at hindsight bias, for which the Mueller Report provides as perfect example. Hindsight bias is the tendency after an event occurs to conclude that we knew the event would occur all along. It is an ex-post reduction of the uncertainty that actually existed prior to the event.

Leaving aside the pretty important fact that very few people have actually read the Mueller report, it is clear that at a headline level it did not find sufficient evidence for a criminal charge of collusion. This has led lots of pundits to loudly proclaim how they were right all along and some of them to go further by saying that this proves the investigation wasn’t needed. These statement are of course fueled by political motives but they are also hindsight bias at work.

Consider a different situation for a moment. You are buying a house that you really like. It is from a time period in which asbestos was used a lot and your spouse pushes for an asbestos inspection, but you are reluctant to spend the money. After a contentious debate you go ahead with the asbestos inspection. The inspection finds no asbestos. You: “I told you all along this is a great house and this inspection was a total waste of money.”

This example, because it is not political, makes the hindsight bias fallacy quite obvious. There was real uncertainty about whether there is asbestos or not in the house. When your opinion of “no asbestos” was confirmed in hindsight you take this as evidence that you were definitely right, that there was no uncertainty to begin with and hence the inspection was a waste. I have a feeling that the comments on this post will be full of hindsight bias on the Mueller report, presented even after having read this post.

The arguments will go something like, “Albert, in your example there was real reason to suspect asbestos based on when the house was built, but there was no reason to suspect collusion.” For this, I simply leave you with 

And the meeting at Trump tower. And the pro Russia change in the GOP platform following Trump’s nomination. But please do go ahead, because I will be happy to have the comments confirm the point of the post.

Uncertainty Wednesday: When Small Numbers Matter (Aircraft Safety)

Last week’s Uncertainty Wednesday was titled “Fooled by Small Numbers” and discussed why drawing conclusions from small samples is dangerous. A reader asked how that relates to grounding the Boeing 737 Max 8 airplane after two crashes? Isn’t that also drawing a conclusion from a small sample? No. Because here we are actually looking not at a sample but at the population of all flights.

Globally there are nearly 40 million commercial flights annually. That is over 100,000 flights every day! Each and everyone one of those flights is tracked. And in an amazing accomplishment of engineering, process and public safety regulations we have had years with zero  crashes. When people complain that we are going backwards on commercial flight because we used to have supersonic flight with the Concorde, what they are ignoring is that the real demand was not for speed but for safety.

With such a large number of successful flights, a crash is a strong indicator that something went wrong. At a high level there are two possible explanations: equipment failure or operator error (of course the two can interact with each other and compound). On any single crash it would be hard to have two much of a view as to which one caused it and one might rightly suggest that grounding a whole fleet of planes is an overreaction. But two crashes with the same type of equipment point strongly at the aircraft itself.

That argument is of course further strengthened when there is significant additional evidence pointing at the plane, such as complaints by pilots along with known changes to the plane. In summary then: when something normally goes right more than 100,000 times a day every day, and then goes wrong twice in a row with the same equipment, that’s a very strong signal.

Uncertainty Wednesday: Fooled by Small Numbers

The last couple of Uncertainty Wednesdays examined fallacies based on statistical phenomena such as imperfect correlation and the base rate. Another one of these has to do with small numbers. A famous example of this fallacy occurred when a study showed that among the best performing schools in the country there was a high percentage of small schools. This resulted in a push for making schools smaller.

A lot of money and effort was expended on this push for smaller schools. That is, until someone decided to also study the worst performing schools. And again, it turned out that there was a high percentage of small schools among them. Ouch! So apparently being a small school either gives you really good or really bad performance.

In reality much of this is simply was the result of small numbers. When you have a small number of students, you have an even smaller number of teachers. And so it is much easier to wind up either with really bad teachers or with really good teachers. As a school gets larger, it is much more likely to have some good and some bad teachers. Hence it is much more likely to be of average performance.

This small numbers effect is, like the previous fallacies, incredibly widespread. In companies think about the performance of departments (small ones versus large ones). For governments consider small cities or countries compared to large ones. And so on.

Uncertainty Wednesday: The Base Rate Fallacy and Why Hiring is Hard

Last Uncertainty Wednesday considered how imperfect correlation is related to the narrative fallacy which in turn underlies observations such as the extravagant headquarters effect. Today I will examine why hiring is so ridiculously hard and even organizations which are very good at it will make their fair share of mistakes.

The underlying reason for why hiring is hard is the base rate fallacy, which occurs when you update your beliefs too much on the basis of flimsy information. What do I mean by that? The classic example that Kahneman provides in Thinking Fast and Thinking Slow is guessing which job someone has based on a description of their characteristics: “Steven wears glasses and has a meek demeanor. Is Steven more likely to be a librarian or a truck driver?”

Our story telling brain wants to jump to the conclusion that Steven seems to fit our stereotype of a librarian much more than that of a truck driver. But in the US there are 3.5 million truck drivers and only 170 thousand librarians. So the base rate is that Steven is 20x as likely to be a truck driver than a librarian! It is therefore quite unlikely that the information about glasses and his demeanor is enough to suggest that Steven is actually more likely to be a librarian.

How does this relate to hiring? When I first got into business, I used to think that most people are good at their jobs and it was just a question of avoiding people who are outright bad. Now several decades in, I am convinced it is the exact opposite. The base rate heavily tilts towards people *not* being good at their job! This may seem overly dramatic but if you look at the realm of sports for a moment and ask yourself the following: if I randomly pick someone who knows how to play soccer, how likely are they to be a good soccer player? The answer quite obviously is: not likely at all (soccer is the most popular sport globally).

So now consider hiring someone in say marketing. How likely is someone, even someone who has done a lot of marketing, to be good at marketing? The answer is: quite unlikely. Very few people are actually good at marketing. Very few people are really good at anything. So in trying to find these people, the base rate works heavily against you! This means that you need to massively dial up your level of skepticism and the information required to move someone from the base rate to the “this person is good at marketing” needs to be incredibly strong.

There is tons of great advice out today on what kind of interview questions to ask. Unfortunately that is of course a bit like a game of spy versus spy because as the questions become known, people tend to be better prepared for them. At the end of the day there is no substitute for meeting a lot of people to fill any one role, to really dig into what they have actually done and if at all possible to have them do some of what they will be doing as part of the process (and of course talk to people who have worked with them). Throughout it all always consider that the base rate is working against you in hiring.

Uncertainty Wednesday: Imperfect Correlation and the Narrative Fallacy

Last Uncertainty Wednesday I wrote about how regression to the mean at least partially explains the “curse” of building extravagant headquarters (which often signals an impending downturn in the performance of a company). Whenever we come up with a detailed causal explanation for something that is simply a statistical effect we are caught up in what is known as the narrative fallacy. As Kahneman documents so well in his book Thinking Fast and Thinking Slow, the brain has long evolved to want to see causal connections and tell stories about them which is why we are so prone to this fallacy.

Regression to the mean can be explained by considering that skill and performance are imperfectly correlated due to the role of luck. Imagine for a moment the converse: a hypothetical sport in which outcomes are strictly based on skill every single encounter. Such a sport would be incredibly boring because skills don’t change very quickly and so for many encounters the results would be pre-determined. There would also be no regression to the mean in such a sport. If your team had a great season last season it will have a great season again next season, because all of last season’s success was due to skill, none of it due to luck.

As it turns out imperfect correlation is also at the heart of another powerful narrative fallacy: inverse correlation at the extremes. Because imperfect correlation is everywhere (perfect correlation being rare), this narrative fallacy is pervasive. How does it work? Here is a graph of two imperfectly correlated variables

Now ask yourself the following question: why is it that countries with the highest GDP per capita do not have the highest well being? And watch your brain immediately go into narrative overdrive. “Super high GDP per capita means people are working all the time, so they have less time for their friends and family.” Voila. A perfect explanation. Done and moving on.

Except you have just committed a case of the narrative fallacy. Whenever two variables are imperfectly correlated there is negative correlation at the extremes and it simply must be that way. The only time that’s not the case is when the two variables are perfectly correlated. To see this better visually consider the following version of the chart

Now look at the points to the right of the green line. Those are countries with extremely high GDP per capita. You can easily see that these countries are unlikely to have the highest well being – the only way they could would be if well being and GDP per capita were perfectly correlated. This is also true of course at the other extreme, when you look at points to the left of the blue line. These are countries with the lowest GDP per capita. Again they are unlikely to have the lowest well being because the only way that could happen is with perfect correlation.

Since correlation is symmetric you could go through the same exercise drawing horizontal lines and looking only at countries with the highest well being (or the lowest well being). And since this shape is the same for all imperfectly correlated variables (just remove the axis labels) this has absolutely nothing to do with the measures being GDP per capita and well being but applies to any two imperfectly correlated variables. Kahneman gives the example of the intelligence of spouses – which can make for a very controversial dinner party conversation (given that it is imperfectly correlated). People will come up with all sorts of amazing explanations for that one!

So next time you find yourself coming up with a story about anything at the extremes of imperfectly correlated variable (e.g. the highest performing startup entrepreneurs and some habit) just stop yourself right there to see whether you are about to be sucked into the narrative fallacy.