Averages are addictive. Whatever the quantity of interest (heights, weights, financial returns, interest rates, rainfall levels, temperatures, economic growth rates, budget deficits, sales, costs, income, wealth per capita, speed...) it is first summarized by its average. When it comes to predicting the value of a random variable from past observations (like the stock market return next year) averages come again into play. But, as we all (should) know: "The expected value is not to be expected". This is a casual way of saying that computing an average inevitably loses information about the structure of the data it aims at synthesizing. In other words, averaging yields a trade-off between data compression and information loss. Information compression makes it handy to deal with large samples of data. Compression however has a cost. Indeed, knowing the average height of students in a class is of no help to the tailor in charge of sewing the individual school uniforms! If he were to sew an average size uniform, some students would end up with trousers way too short while others would wear jackets too big for them.
The information loss is well monitored in the case of the famous bell curve, also known as the normal distribution. In this popular distribution data are symmetrically spread around their mean. Their dispersion around the mean is measured by a parameter called the variance which happens to also quantify the information loss if one were to squeeze the whole distribution to its mean. The bell curve is appealing because although data are dispersed around their mean they remain in a tight neighborhood to it. Student heights is a good example of this behavior: For a given age class, say 18 year old, there are no student measuring 3 meters or 90 centimeters. There are tall students as well as small ones. However they do not deviate that much from the average height. As a result, the mean and the variance are enough to capture the full distribution.
Is it always that simple? Of course not. It all depends on the context from which the data are drawn. If the data relate to weight, height and the like the bell curve is a good approximation, and the average conveys a valuable information that can be benchmarked against the variance. Now, what if we look at wealth per capita levels? Assume hundred people in a room. Their average wealth is, say, $180 000. Enters Bill Gates. The average is computed again, and now it is worth, say, $300 millions. This figure makes no sense at all. One single individual out of 101 affects the whole result. Hundred individual are well below the mean while one is well above it. The bell curve does not work, and the average wealth does not make any sense. Under this scenario the average is mean indeed. The same holds true for floods. The average flood is of no interest to the bridge builder. If he were relying on it, the bridge would not last long for sure. Rivers like wealth processes are too complex to be bell-behaved! Although the bell curve is called the normal curve it is not the normal state of affairs. Things that matter most to mankind are not normal. Think of how connected the economy has become: one small change somewhere may trigger a sequence of events that lead to a monstrous event. Think of the weather and its nonlinear intricacies. The string of factors, be they meteorological, social or economic, will magnify the initial change in ways that we are more often than not unable to monitor, not to say understand. Averages can no longer be trusted.
How come then we spend so much time and resources to estimate averages, dispersion measures, probability distributions while we simply fool ourselves? We become preys to our own inevitable errors, errors that can be lethal in hyper-connected times. This is sad: we should rather strive, as Nassim Taleb puts it, to "live happily in a world we do not understand." When we devote significant resources to model a given variable, say, x, and to estimate its average we forget that what matters is not x itself but how we are exposed to it, say, f(x). The function f describes our exposure to x, for instance the gains and/or losses associated with various values of x. Suppose x stands for floods. The function f may describe the damages to our home that floods could trigger. Whatever the depth of our knowledge of floods our house will not be spared if it sits next to the river. What matters is not the knowledge of x (namely floods) but f(x), namely whether our home is vulnerable or not to floods. Nassim Taleb calls the confusion between x and f(x) the conflation error. Whatever the accuracy with which we measure flood levels, one day our home will be flooded if we build it in an exposed area. That very day we will be hit by Nassim Taleb's famous black swan.
There is an easy way to live without the fear of the flood black swan: build our home away from rivers or seaside, namely change our exposure. Indeed,some battles do not deserve to be fought! But even if we are careful enough to avoid the conflation error, we are not done yet. Our effort should now be directed toward the understanding of the shape of function f(x). Is it linear? Is it convex? Is it concave? In other words how does it behave in response to changes in x. When x changes, does f(x) change more or less than x? This is a major question that we cannot escape especially in times of growing inequalities, winner-take-all dynamics and fast automation. To understand why let us look at the fate of the middle-class. The middle class that was once the symbol of prosperous economies is hollowed out. Shooting for the average position has become a very hazardous stance. Over the last few years, jobs have been more and more polarized. At the upper end the cognitive jobs that benefit from automation fare rather well both in terms of offerings and wages. Through automation they become more productive. They can do more while delegating the ancillary tasks to automation. At the lower end, employment in manual-task intensive jobs that machines cannot replicate (like say cleaning, services to the persons etc...) have risen. These jobs however command lower wages as a lot of people are chasing them including the hollowed out middle-class people. In the meantime jobs in the middle are eaten by software at a fast pace.
It really does not pay to be or to stay in the middle. Jobs prospects are better at both ends than at the middle. This hollowing out of the middle-class is stronger in an economy framed by increasing returns to scale and winner-take-all effects. The upper end captures most of the gains while the lower and middle ends have to share the leftover. Digital automation means that a growing share of income is currently going to capital owners. This effect is usually known as "capital-biased technological change" although I fear, if things do not change, we may have soon to call it robber robot lords biased technical change. To use Nassim Taleb's words, we may have entered into a barbell economy. Is this is good news or bad news?
The obvious and immediate bad news is that the formerly praised middle class is being "stolen" from us. Robots, software and machines take over the jobs that once made a significant portion of the society well-off. This middle Golden Age may not return, and for many it is indeed awful news. The good news rings like a salutary wake-up call. It helps understand why the average can be truly mean, why average may translate into fragile. Indeed, the hollowing out of the middle class provides clear evidence that it is fragile: it is vulnerable to external shocks such as technology driven labour market shocks. It loses from them more than it may gain. Automation lower production costs. Middle-class workers gain through lower consumption prices, but this gain is illusory when at the same time they lose their jobs. This observation brings us back to the exposure issue, to the conflation error in which x is often confused with exposure f(x). Assume that x stands for income. At the risk of oversimplifying being a middle-class member means that one earns the mean of x. Let us call it mean(x). The function f(x) describes the gains and losses associated with earning income x. Function f may include all sorts of things including well-being, satisfaction related benefits, accomplishment, status, prestige, etc... Let us take the example of an ultra simplified economy. In this economy the middle-class people are subject to random shocks that may depress their income x or increase it. Let us assume that a sudden technological shock occurs: A piece of software has been invented that automates middle-class routine jobs. The product that used to be manufactured by the middle-class worker can now be sold for free. What is the outcome? Holding the "average" job makes the middle-class worker vulnerable to automation duplication. As a result, he loses more (his job) than he gains (his purchasing power). As a thought experiment, imagine a fictitious individual who over a given period of time swings between times of low income and times of high income. Let us compare this fictitious individual with the middle-class earner (who constantly earns the mean income). As the middle-class worker has more downside than upside his "satisfaction" from earning the average is higher than the average "satisfaction" he would feel from swinging back and forth, namely if he were the fictitious worker. This spread is a valuable signal. It shows that middle-class workers overestimate the true state of their condition given their (concave) exposure to technological disorder. In other words randomness and harm are both underestimated. Once the exposure has been properly taken into account, it turns out that the seemingly stable and robust middle class is an illusion. Middle-class people are closer to the fictitious swinging individual than they think, and this is something they should be aware of. The average is indeed a liar. Northwestern University Professor of Economics and History Joël Mokyr summarizes the situation very nicely:
"Modern technology often leads to winner-take-all outcomes, and the inequality implications in terms of income – though not in terms of access to the good itself – are worrisome. What we gain as consumers, citizens, viewers and patients we may lose as workers. The demand for labour 'hollows out' and the demand for medium-skilled labour declines unless and until new jobs are created to absorb those replaced by automatons and robots."
The situation of cognitive well-paid workers is opposite. They gain from digital disorder, at least as long as the disorder outcome is software that complement their skills. They for sure gain more than they may eventually lose especially in presence of increasing returns to scale and winner-take-all effects. Their skills get augmented by the digital leverage of software. Here again the average is a liar. It underestimates the true situation of cognitive workers. Indeed, while middle-class workers are concave to external shocks cognitive workers are convex to these shocks2. As Joël Mokyr, Chris Vickers and Nicholas L. Ziebarth observe:
"routine tasks with little unpredictable variability are more likely to be mechanized, while jobs that require continuous adjustment to new information and new physical settings along with fine sensory motor-coordination are more difficult to automate."
Their observation is valuable on two accounts. First, it seems that jobs that themselves contain an intrinsic dose of disorder are those that are more difficult to hand over to machines. Moreover machines or software may help people holding these jobs become even better at them by relieving them from ancillary time consuming tasks. Second, lower-skill workers are now competing head-to-head with middle-skill workers who have no other job choices. Wages decrease as a result. These job market movements entail a growing bipolarization in which a larger percentage of the population chases lower paid jobs while a minority enjoys the riches. As a result, and as we noted earlier, the economy is now entering a barbell mode with strong dynamics at both its low end and high end and vanishing ones at its middle. The inevitable question is whether or not this "barbellization" is sustainable. In other words, the average is mean, but is there life beyond the average?
I think it is no coincidence that the barbellization takes place at the same time of what I call the Faustian "free against your data" swap occurs. Software, machines and artificial intelligence require tons of data. The easy way for robber robot lords to siphon them is to offer free services in exchange of them. Software then eats the middle because it has indeed been fed by the middle. A strong (data) currency has been traded by the middle against a rather weak (free) one. This is why Joel Mokyr, Chris Vickers and Nicholas L. Ziebarth observe that
"in a world of cheap goods, while inequality in terms of wealth or income may rise, inequality in the form of access to "primary" resources would greatly be diminished."
In other words, the lower end and the middle gets poorer in absolute terms, but they should not complain too much as they now enjoy free or low cost prices. What they lose on one hand, they regain it on the other hand. It remains however to be proven that the balance is (or will be) fair in relative terms. The massive use of data has another consequence which the optimists call the sharing economy while the pessimists call it the on demand (precarious) economy. One metaphor is often used these days to describe where this on-demand economy is heading, namely the Uberization of the economy. In short, old corporate fortresses with their legions of full time employees are torn apart to give place to data driven light structures with flexible on-demand workforces. As a result it is not only (middle) jobs that are disappearing. It is the very nature of work that is changing. This change takes place not only through self-employment but also and above all through a new breed of firms that match online in real time tasks and workers. Uber, Amazon's Mechanical Turk, BlaBlacar, AirBnB, etc... are the best known. Lawnlove (on-demand gardening), KitChit (on-demand cook), GoldenShine (on-demand domestic cleaning) are other less known examples of the same trend.
As always this buoyant activity carry good news and bad news. The bad news is that this on-demand workforce may indeed become a precarious workforce. It will have to face significant uncertainty as to when and where it will be called for a task and how much it will receive while at the same time having to cope with regular and fixed living expenses. Moreover modern work conditions are stress conducive as we all know. Stress levels may become worse. They may even outweigh the benefits of working from home, namely achieving a better balance between work and family life. There is also the obvious risk that employers may take advantage of their position especially with lower income workers that are easily substitutable. The on-demand flexibility associated with random revenues also begs the question of social coverage, of taxes, of access to housing... It is quite clear that we have entered into unchartered territories that the visible hand will have to thoroughly investigate to foster and preserve accountability and sustainability. The current fiscal, social and regulatory frame designed for other times is ill-suited for the new issues at stake. The new framework will have in to ensure that robber robot lords do not exhibit predatory behaviors. While this issue is being taken care of, the new framework shall provide the settings for a reshaped work landscape.
Indeed, if we manage to keep the robber robot lords at (political) bay, if we succeed in regaining legitimate power on our data, good news may be in store. Things went bad for the middle class workers because of the fragility of their concave technological exposure. It does not have to be so. Digital capital is cheap, and the cost of production is getting closer to the cost of reproduction. Convex exposure to digital capital is easy to get into, much easier than convex exposure to physical capital. In other words, it has become easy for people to own and manage their digital capital which in turn means that we may see cottage businesses blossoming. Obviously this does not imply that everybody is a born cottage entrepreneur. It means that people with ideas to, say, better service their local community can experiment them, tinker with them with a small downside and a truly rewarding upside. This is the best way to put oneself into a convex position where there is more to gain than to lose, where the feeling of having achieved something is at its peak. At last, it will happen provided the visible hand has worked hard to make it easy to ignite.
Despite the good news one shall not minimize the unavoidable adjustments that we will have to weather in the short and medium run because "average" is no longer good. A lot of people fell victim of the illusory comfort of reaching and belonging to the middle class. They did not realize the dangerous exposure (the other side of the coin) that being average implies. By the same token they also bought the miraculous digital bounty of what I call the Faustian and unfair "free against your data" swap. Sadly enough, wealth and income inequalities powered by increasing returns to scale and winner-take-all dynamics have simultaneously soared yielding a bipolar economy. For this economy to be sustainable, for the upper end not to divorce in unbearable ways from the lower end will require more than the charity of the upper end. It will require smart visible hands (an oxymoron for a lot of scholars) to make the two sides of the barbell accountable to each other. Otherwise, we will end up with a more and more concave lower end and a more and more convex upper end. The lower end will be enslaved to robber robot lords devouring the lion's share of productivity gains.
To put in (too short) a nutshell, the middle was stolen from us. While the middle was no panacea to say the least let us make sure, thanks to the cheapness of digital capital, that convexity is truly ours, that it will not be stolen from us. Digital winds shall fill all sails ready to venture afar. As I write nothing is less sure!
Comments