The Technocrats' Recovery

Plus! Big Tech Sees Like a State Update; Why Don't Car Subscriptions Work; Taylorism-as-a-Service; The Semiconductor Boom; R&D and the Genius Problem

Welcome back to The Diff. Here are the subscribers-only posts you missed this week:

This is the once-a-week free edition of The Diff, the newsletter about inflections in finance and technology. The free edition goes out to 18,954 subscribers, up 195 since the last edition.

In this issue:

The Technocrats' Recovery

Every recession follows the same template, but only in retrospect. At first, times are good, the economy is growing, and everyone is happy. But returns drop—the converse of "multiples go up" is "yields go down"—so people borrow to keep making the returns they're used to. The economy gets more forward-looking, which also makes it less flexible: investing for the long term means having less liquidity in the short term. At some point, there's an adverse shock, and its side effects ripple through the economy in unpredictable ways, leading to an unwinding of leverage, slow or negative growth, layoffs, bankruptcies, and all the usual symptoms of a recession.

The trouble with that model is the "there's an adverse shock" part. It's very easy to see in retrospect that the bursting of the housing bubble, or the Covid-19 pandemic, or the end of the Y2K capital expenditures spree could end a bubble. But zoom in on any long-term chart of the market's performance, and you'll see plenty of cases where stocks declined on plausible evidence of an incipient recession, and then rose to new records when that recession didn't materialize. In 2011, Europe was a mess, the US's credit rating had just gotten downgraded—and we muddled through. In 2015-16, declining oil prices and contractionary policies in China led to a steep drop in US investment—for manufacturers, that period was the end of the post-crisis expansion, and a second one started soon after. And yet, the market went on to hit record highs.

Since predicting recessions is hard, reasonable policy responses to them are harder still. One reason the dot-com bubble got so heated in 1999 was what happened in 1998: a series of financial crises in East Asia and elsewhere led to panic-selling and the near-collapse of the US financial system. In response, the Federal Reserve cut rates sharply, among other interventions. The slow feedback loop between interest rates and the real economy meant that money was fairly easy to come by at a time when the economy was already doing quite well, which led to massive increases in investment in the information technology sector.

This is forgivable; predicting the future course of the economy is challenging, and even the best get caught flat-footed. When inputs and outputs have variable lags and uncertain quality, control systems are hard to get right.

That's changing, in two directions.

First, recessions are increasingly driven by finance, and especially by the kinds of assets with quoted prices. Historically, banks played a bigger role in crises, and it was always hard to measure their performance in great detail. Japan's lost decade, for example, was partly a function of banks keeping their loans marked at 100 cents on the dollar despite their borrowers' dubious ability to pay. Instead of a short, severe recession with lots of writedowns, the economy limped along for years.

Markets do not react gradually to central bank policy. Even the difference in latency between an audio-only and video feed of the same speech can be exploited. An economy where more assets are publicly quoted and instantly marked-to-market is in one sense a more volatile one; the people who would white-knuckle their way through temporary insolvency will get margin calls and liquidate instead. But, paradoxically, it's an easier economy to stabilize, because central banks can just turn the "intervention" dial until prices start behaving the way they'd like.

The interconnectedness of different markets also makes the job easier; it was not very obvious in August of 2007 that the reason momentum stocks and value stocks were dropping at the same time was that diversified hedge funds had made money-losing subprime trades. And it wasn't easy to see why Russia's default in 1998 was so bad for Long-Term Capital Management, a company that didn't own any Russian assets. If everything is connected, and leverage is the single point of failure, then liquidity becomes the single source of success.

This is a self-fulfilling prophecy: as crises get more finance-driven, the financial sector becomes a more important mechanism for solving them, so the crisis-intervention cycle tends to make the financial sector more important and easier to intervene with—which makes the next round of emergency central bank measures more effective (at the cost of making them more necessary).

The other source for optimism is this fascinating paper from a few months ago, on the macro effects of Covid-19. It's full of interesting details, and very much validates the V/L recovery thesis ($). For the highest income earners, employment and spending plummeted in March 2020, but their recession was more or less over by the end of May. For the lowest earners, employment remained 20% lower than its peak as recently as October. Higher earners cut spending more, while low earners used unemployment and relief checks to substitute for missing jobs, and kept their spending fairly steady.

All this is interesting, but raises the question—how do we know? As it turns out, the paper is a great exercise in real-time econometrics, using data from consumer credit card panels, a company that pays consumers to upload photos of receipts, a small business credit card data provider, payroll data from Paychex and Intuit, numbers around how many workers clocked into which shifts from an hours-tracking product, scraped job postings, and usage data from an education software company. This creates a very three-dimensional look at what's happening with the economy—who's earning, who's working, and what they're spending on. Public data sources added more information on where people were going, and how many were getting sick.

Covid-19 had financial aftereffects, but unlike the last three US recessions, it was not primarily driven by finance. So broad interventions to affect aggregates were not the answer. Coming up with targeted policies is hard, but with datasets like this, it's now possible. The ideal economic response to Covid was a combination of:

  1. Paying people not to do things that raised infection risk, like going to restaurants, bars, and salons, and
  2. Providing a cash buffer to offset the drop in consumption, such that anyone who could have paid their debts had Covid not happened can still do so.

That's an unachievably high standard, but we can get much closer now than we could have even a few years ago, as the volume and timeliness of data grows. The report shows some of the strengths of the Covid response: unemployment is up among low-income workers, but their consumption hasn't collapsed accordingly. And it points to some mistakes; the policy response redistributed a great deal of wealth to asset owners who can work from home. It also points to some interesting future problems: spending growth was higher for durable goods (if spending doesn't drop much, but some forms of pure consumption are gone, that's arithmetically true). If those durable goods are long-term substitutes for future consumption—kitchen gear replacing meals out, better TVs replacing movie theater trips, video games as a substitute for leaving the house—then the next big economic problem is a supply glut that's shifted from manufacturing-centric economies in East Asia and Europe to households in the US.

Financialization and better data are two broad cases of the world getting more legible. A legible world has its upsides and downsides, but one of the upsides—the whole reason states and companies try to impose legibility—is that they're easier to govern. In the next recession, there will be far more precise tools available to policymakers who want to make the it as painless as possible.

A Word From Our Sponsors

As economic activity comes online, new data trails are left behind. Over 200 innovative investment firms leverage this data for alpha-generating insights. Thinknum’s alternative datasets power investment decisions across more than 500,000 public and private companies around the world.

For example, track job listings to identify which companies are expanding in real time. Or track the individual tenants by commercial mall to analyze churn on a daily basis.

Interested? Come learn about all of our 35 datasets here.


Two interviews this week: I talked to Nathan Barry about the business of writing newsletters, and spoke to Taylor Pearson about bubbles, S-curves, optionality, and other classic Diff topics.

Big Tech Sees Like a State Update

A consortium of tech companies including Microsoft, Salesforce, and Oracle are working with healthcare companies to create a proof-of-vaccination service. The rollout of vaccines has coincided with several new and much more contagious strains of Covid-19 ($, FT), so in one sense the deficient argument for March's lockdowns finally applies: normal behavior for the unvaccinated is much more dangerous than it was before, but that behavior only has to be constrained for a limited amount of time before everything can go back to normal. Having a single trusted source for vaccination status is a much better solution than rolling back restrictions based on the share of people vaccinated in a given location—healthy twentysomethings are towards the back of the line for vaccines, but they're also the demographic most likely to behave in high-R(t) ways.

In other vaccine news: Instacart has joined Dollar General and Trader Joe's in paying employees to get vaccinated. This is not too far off from the outcome a Coasian utilitarian would expect if vaccine access were rationed by price: any business that relies on frequent in-person contact faces large internal costs from workers getting sick, and imposes externalities if they sicken others. Expanding the scope of behavioral green zones is an important policy goal.

Why Don't Car Subscriptions Work?

BMW has cancelled its monthly car subscription service, which let users pay a fixed monthly price to get access to a range of vehicles, along with insurance and maintenance. In hardware-as-a-service, I argued that the subscription model helps manufacturers of durable goods get around several weaknesses of the hardware model. But it hasn't worked here. One possibility is that the car industry is relatively mature, so incorrect estimates for the demand of a new generation of cars are off by tens of percentage points, not orders of magnitude. Another possibility is that, given the many options for financing a car purchase, or leasing a car, subscription models aren't differentiated enough to appeal.


The Economist highlights advances in manufacturing automation ($) as another gradually-then-suddenly trend accelerated by Covid-19. One detail in particular stood out:

Drishti, an American startup, has come up with a way to apply artificial intelligence (ai) and computer vision to analyse busy video streams of workers on assembly lines. Marco Marinucci of Hella, a big German car-parts supplier, says his firm used Drishti’s kit to analyse and fix problems at a high-volume assembly line. This allowed its throughput to rise by 7% last year.

This is a very positive development. In the very long term, capital equipment has been a complement to labor in the abstract, but a substitute in most specific cases—tractors eliminate farming jobs, but most of us found jobs off the farm in response. When software raises productivity at existing jobs, it tends to lock in both the software and the reliance on human workers; they both adapt to what the other can't do.

The Semiconductor Boom

Going back to the 90s, US investors have loved asset-light businesses. Any company that could move factories and operational complexity off its balance sheet, and live purely on royalties, was accorded a high multiple. This trend is starting to reverse, and one piece of evidence is Taiwan Semi's blowout quarter, where they missed revenue estimates but wildly exceeded capital expenditures estimates. TSM plans to spend $25-$28bn on capital expenditures this year, compared to consensus of $21bn. Capital intensity is a drag on growth when companies need to constantly finance new investments to add revenue. But it's also a moat, if they don't have many competitors who can effectively spend as much as they can.

R&D and the Genius Problem

Nintil assembles evidence that scientific productivity is heavily skewed to the most productive researchers. (Much of the writeup consists of raising good objections to this thesis, and then looking at data that shoots those objections down.) This is a very positive story, if it's true; smart people have many career opportunities, most of which pay a lot better than academic research. If most good research comes from a handful of people in each field, then outbidding the private sector for their efforts would be a good deal.

On the other hand, there's an optimistic but status quo-based interpretation: as tech companies get bigger and more diversified, they're better able to monetize highly abstract and theoretical R&D. AI looked overhyped a few years ago, but turned out to be underhyped, in part because a handful of companies threw big budgets and vast datasets at it. As these companies get bigger, their demand for pure research will grow; Google can afford to converge with MIT faster than MIT can adapt to converge with Google.