TLDR: This book is recommended for anyone who uses math to make money, whether in finance, or (in my case) the areas of the tech sector that rely on modeling customer behavior to make a buck.
This is not a how-to manual, or even a book that tries to make a monolithic argument (although if you take it seriously, one seems to emerge). It’s more of a conversation with Aaron Brown, a prominent (I’m told) risk-manager, old-school quant, and all-around interesting guy in the financial field. It’s easy for these kinds of conversational “what I’ve learned, what I think” books to come out sideways - a barely-edited transcript of an “author” and their ghostwriter (the plague of political memoirs), a platform for the airing of grievances, a journey into the depths of self-ego-stroking, or a collection of platitudes. Fortunately, Brown is better than this.
The book starts to heat up with an interesting interpretation of the Dutch “tulip mania”, which he contends was not a classical mania at all, but a somewhat rational investment in a commodity with characteristics making it a good money substitute. Tulip bulbs (the commodity actually invested in, not tulips per se) are portable, somewhat durable, have predictable inflation rates via an annuity in the form of more bulbs, are impossible to counterfeit or debase, and have an aesthetically pleasing, branded output. It makes at least as much sense to use them for money as it does wampum, and possibly even more than gold or silver, especially when you have a massive inflationary influx of bullion from the New World and a government that debases coins. Their exchange rate for classical coin is somewhat explainable by purely economic and legal factors.
After some digression into the nature of “money” per se, this is used as an entry to a discussion of “money” vs. options / futures / derivatives. In Brown’s formulation, the functional purpose of commodity futures, like an agreement to buy or sell a quantity of wheat for a certain price at a certain time, is not a way to “lock in a price” in the sense of hedging price risk. A miller typically is not “long wheat” in the futures market to guard against price spikes - a price spike impacts him in unpredictable ways (for instance, it may be due to either an increase in demand, or a fall in supply, which would have opposite effects on his business). Instead he contracts for purchase on a more flexible retail basis (not in the futures market), and is short wheat in the futures market (using the proceeds, in fact, to buy the actual wheat he is processing). These offsetting transactions, one long, one short, have the effect of borrowing wheat directly, without the necessary use of money (the contracts themselves can be used as collateral for each other since they nearly offset). When he has sold his flour, he buys out his short with the proceeds and repeats the process. Historically, money itself was a scarce and volatile commodity on the frontier, and eliminating the use of money eliminated a substantial source of risk. Instead of an interest rate linked to public debt levels, bank stability, foreign currency flows, etc., one has an interest rate more closely linked to the inherent properties of the markets actually transacted in.
As a digression of my own, it is plainly inaccurate to say that “continually borrowing with no intention of ever paying anyone back is a totally modern development”. If the miller is doing his job right, his debt will be greater each year, and he may very well be borrowing from the same person each time - if it made sense then, why not now? The question is whether he is doing something productive with the proceeds, and what the ratio of his debt to the value of his business is (in fact, if he can borrow more easily and does so, this in and of itself causes his business to increase in value). It gets dramatically more complicated, and the analogy rather breaks down, if the miller’s business is incredibly difficult to value accurately, and he owes most of the debt to his own subsidiaries, heirs, tenants, people legally obliged to hold it & never redeem… In any case, if one must analogize, it’s a far more suitable analogy than grotesque “country as household” rhetoric.
The final, and most generalizable, part of the book focuses on the notion of risk and probabilities itself, how it relates to models, and how the impact of these risks (like the unfortunate “going bankrupt” thing) manifests and can be controlled in the real world. The Kelly criterion is introduced as a way to think about investment strategy, and the warring camps of frequentists and Bayesians are reconciled into a neat package by explicitly considering what a machine learning practitioner would call a loss function and Brown considers as a choice of numeraire (this synthesizes nicely with the question of borrowing bushels of wheat vs. cash, and it is only poor editing that leaves this connection to the discovery of the reader). Dollars are an easy choice, that usually has a neat quasi-Bayesian interpretation - you may consider it in terms of conversion rates or eyeballs, but be careful when the relationship between those and the almightly dollar breaks down. Dollars aren’t always appropriate either, especially when there is no free market setting prices for the underlying events - if you’re trying to build an accurate speech-to-text engine, it’s foolish to try to put a dollar price on each error.
When models break down, Brown has an engaging explanation of the concepts of value-at-risk and risk management in general. Being not an expert in the field, it’s difficult for me to judge his approaches to risk, and many of them seem inapplicable to my line of work. The technology sector doesn’t have “traders” to manage at quite the level he describes, but the notions of rigorously evaluating performance, taking appropriate levels of risk and planning for what happens when you fail, is universal.
Ultimately, reading this book has convinced me that there is a massive mismatch in technical sophistication between models in the financial and technology sectors. The high-end predictive models being developed by the technology sector, for image processing, natural language processing, collaborative filtering, click modeling, fraud detection, etc., cover much more ground and seem vastly more sophisticated than those of the financial sector. They incorporate more data, make higher-dimensional predictions, and generally use heavier machinery. But the superstructure, the way models are actually used, thought of, and evaluated on a meta level, lags badly behind. Most of this is probably due to the decision time horizon - it would be a different story if the financial sector didn’t require sub-millisecond latencies for their predictions, or if a slight increase in face recognition was worth the billions of dollars a consistent edge in the financial markets is worth. It may be, with time, that we will see the financialization of the technology sector, securitizing and selling derivatives on click rates or ad impressions in the same way we securitize timber reserves or the profits from Facebook in toto. Already the startup sector leads the way - the only way to understand something like Instagram is as a purchase, not of a revenue stream per se, but of a nevertheless valuable commodity, financed by transforming monetary capital into users, engagment, a software platform, or whatever you wish to call their end result.
The unfortunate aspect of this book, which is not cleanly separable from the thing that makes it interesting, is that it’s clearly a very personal work. The syncretic aspects sometimes diverge into tangential areas, some of the main ideas and interesting connections are scattered, and it could generally use a better editing job. Fortunately, no one will actually force you to read the whole thing - unless they intrigue you, skip the comics, the social history of Wall Street, and the disclaimers, and enjoy a very fresh look at the interface between predictive models, risk, and decision-making.