Book Review—Probability and Finance: It’s Only a Game!
Update: This post has been substantially revised in response to feedback and incorporating answers to good questions. If something seems ridiculous or nonsensical it is definitely best to assume that is my error rather than the authors’.
Probability and Finance: It’s Only a Game by Glenn Shafer and Vladimir Vovk is about math and a new field within math. It is a book-length argument for an extraordinary claim about probability theory. In the course of making this claim they do a lot of what has become my favorite thing to read in math and science—carefully revisit the philosophy and assumptions that underly things we have mostly taken for granted. The claim itself is this: they have used game theory to replace measure theory as the basis for probability.
I’ll allow you to digest that for a moment.
The website for the book is here: http://www.probabilityandfinance.com/. The authors have provided a number of reviews, and their own responses; they are diligent in offering corrections and clarifications even in the case of reviews which are completely favorable. This is the first time I followed the back-and-forth in book reviews in any detail, and I found it very helpful to see the criticisms and confusions of better-educated, more-qualified reviewers voiced and addressed. A word on the authors: Vovk was a student of Kolmogorov directly, and Shafer has previously published on mathematical theories of evidence. They also do work in machine learning, on Conformal Prediction. They appear to take seriously the magnitude of their claim, and consistently highlight what is different about their approach and where the advantages lay. The book argues for a game-theoretic basis to probability theory in Part I, and goes on in Part II to apply this same game-theoretic treatment to achieve finance theory.
Of particular interest is the section of the book on historical background, which is mostly Chapter 2. What strikes me is that there is not much overlap between their history and the one we hear about Bayesian probability, particularly from Jaynes’ book. They begin in the same place as usual with Pascal and Fermat, but hew strongly to formal mathematics throughout—which is to say that Bayes is not mentioned at all. They do discuss some of the different interpretations of probability, including the interpretation of personal choice; this is the closest I can see to what we understand to be the Bayesian interpretation, but they term it neo-subjectivism and trace it to Bruno de Finetti. They relate a lot of interesting context for the development of measure theory, including some details about the relationship between frequentism and computability. Shafer and Vovk’s work relies rather on Richard von Mises and Jean Ville, and in particular Ville’s work with martingales.
The martingale is one of their core concepts, and it has come to my attention the term has an established meaning in probability and game theory which they do not employ, so it is worth disambiguating. In the book they explicitly rely on a generalization of the martingale betting system, which was originally a strategy of doubling one’s bets in order to guarantee that the next bet would make back all the money lost so far. It is from this the name of the stochastic process in probability comes. They privilege the historical origination in many cases, most of which were new to me; this does introduce term confusion sometimes, but they apply it consistently in order to capture (what they claim to be) important subtleties.
They pick up the use of the martingale directly from Jean Ville, who applied it to prove the difference between a frequency which converges via oscillation above and below vs. one which converges only from above—it turns out that if it converges from above, a martingale betting system can make a gambler infinitely rich. This was an important point against von Mises’ collectives, with respect to Kolmogorov’s measure. Ville generalized the martingale to be any system of varying bets, which provided a stronger proof of the impossibility of becoming infinitely rich. The authors argue that the use of the martingale betting system is a universal test for randmoness—Ville showed that for any event with probability 1, there is a non-negative martingale which diverges to infinity if that event fails. They extend Ville to say that the probability of an event is the smallest possible initial value for a non-negative martingale that eventually reaches or exceeds 1 if the event happens. They start from that to build their game-theoretic notion of the martingale, which they then rely on to build the rest of their results.
A point which took me by surprise: this book makes little reference to the established field of game theory. In the first this is because they are firmly working in the field of math rather than economics, and in the second this is because they have reached back further than game theory’s solidification as a field to work effectively from first principles. This puts the work into an interesting grey zone, where they neither have a strong battery of current results to draw from or well-established axioms, but as they state plainly in the introduction they are arguing for the establishment of a new field this does not seem unreasonable to me (the lay reader). The authors make arguments from common sense and appeals to intuition in a way that reminds me heavily of Probability Theory: The Logic of Science.
Their basic probability game consists of two players, Skeptic and World: Skeptic is an imaginary scientist, betting imaginary capital, in order to test a theory. Both players have perfect information, and alternate moves. Skeptic bets on what will happen, and World decides what will happen—the theory governs what the payoffs are, and if Skeptic wins too much the theory is cast into doubt. In the simple case, this amounts to Player 1 betting on what Player 2′s move will be.
I want to take a moment to dwell on something Shafer and Vovk do right out of the gate which blew my mind, but was so obvious in retrospect I am hurt by not having thought of it before: they decompose World into multiple players. My initial interest in game theory came through the wargaming and military strategy aspects of the field, which traditionally suffers from the weakness of modelling whole governments as a single rational agent. I had always reasoned that a more detailed model would make each of a government’s moves the outcome of a series of games played among its sub-agents, but dividing it up into different players is a considerably simpler way of getting at the same intuition.
Some of the results they develop for probability include:
Strong law of large numbers
Weak law of large numbers
Law of the iterated logarithm
Central limit theorem
The second part of the book deals with applying this game theoretic method to problems of finance. My understanding of the background for financial theories is limited, but the impression I have is that they consist of being generated directly against the market data, or sometimes based on probability concepts. That is to say, there isn’t a general mathematical basis for finance theories. This treatment changes that, but importantly it they develop finance theories directly rather than going through probability first. So if we take a theory in finance which was developed from probability previously, and compare it to this book, we go from this:
Measure → Probability → Finance
To this:
Games → Finance
As a result of not going through probability to get to finance, it looks like the probabilistic elements are native to finance theory under this construction—in fact they title this section ‘Finance without Probability’. As there was not a unified mathematical grounding before, I am extremely interested to see if the later research in this vein shows the same kind of benefits for putting finance on a coherent mathematical footing as happened with other applied fields. My expectations are not high however; finance is already an area of maximum civilizational effort.
A direct benefit of game-theoretic finance is relaxing the assumption of stochastic price behavior. The authors take aim at stochastic assumptions generally—they argue that the game theoretic approach does not entail a deterministic or randomness assumption, and so they can accomodate both events in the same analysis without issue. This is motivated by meta-theoretical concerns: instead of specifying a full probability measure and making the measure larger and more complicated when contradicted by evidence, they can simply withdraw the wrong predictions. They explicitly tout the minimalism of the approach.
A brief aside: in this same section Shafer and Vovk raise the question of using game theory, which has discrete steps, to model the markets, which take place in continuous time. I assumed this would be child’s play—or at least undergraduate’s play—as I have a degree in electrical engineering and we do that very thing all the time. Apparently I was wrong in this assumption: they say they solve the problem with Nonstandard Analysis, and go on to say some mathematicians are uncomfortable with this but it is really okay. Perhaps this is because it was only developed in the 1960s?
Some of the results for finance:
Black-Scholes formula (discrete, continuous, with interest)
Diffusion processes
Central limit theorem
Efficient market hypothesis
This is a math book about math theories but it is heavier on the history, nuance and philosophy than it is the formalism. If you already know measure theory, by the author’s own admission you do not gain much utility from this treatment, although you may find the background information interesting and helpful. The finance section, and particularly the discussion of assumptions and information efficiency, seem relevant to prediction markets. The book was published in 2001 and as of yet has no second edition; if you would like to eyeball the papers published afterward to get a better sense of relevance and/or sanity, a list is here. A final caveat—I have not finished the book, so corrections may be forthcoming as needed.
Interesting. Are there cases where the game theory foundation leads to a different probability theory than the measure theory based one? That is, are there any results in standard probability theory that are proved based on measure theory that work out different if you base it on game theory?
Also, I guess I should read the book but what the heck does it even mean for probability theory to be based on game theory since game theory is not generally presented in a way that seems conductive to being the foundation for other parts of mathematics.
Based on the contents of this book, no—they only argue that they get the same results measure theory does. They argue that it is a better choice as a theoretical foundation on the grounds of simplicity, intuitive appeal, and occasionally having broader implications than the measure theory version. This point is addressed frequently in the reviews listed on the website, but in general if you already know measure theory there is not much advantage to learning this, except to refine the occasional subtle philosophical point. However, the book was published in 2001 and they have had continuous research since then; there may be new results of which I am unaware.
That being said, the direct deprivation of finance theory does offer some specific advantages. I don’t know how finance theory was actually developed, but my impression is it comes from probability theory—so the process changes from this:
Measure → Probability → Finance
To this:
Games → Finance
The pitch is that the probabilistic elements are in fact native to finance theory under this construction. Their central example in the book is that there are things in finance we assume to be stochastic but are not, and they argue people accept this assumption because they believe measure theory justifies it. They claim they can accommodate a mixture of stochastic and deterministic behavior without changing any assumptions. As far as I can tell this won’t enable us to calculate things we could not before, but it does seem very much in the vein of knowing where your assumptions lie, and making the model match reality as closely as possible—which hypothetically would lead to making fewer mistakes about how to calculate things.
I really should have signposted this up front, but when they are talking about game theory they really mean more in the John Conway sense them the economics sense. They actually rewind as far as the 1890s and then pick up the thread from there. Since they don’t provide historical context for mathematical or economic game theory, I think it would be fair to say that they have re-derived game theory from base assumptions. They seem to view games and probability as conceptually inseparable, because Probability arose from betting strategies in games in the first place. They argue that the prices are the probabilities; this is very similar to the way that futarchy works via the efficient market hypothesis.
I will make an update to clarify some of this!
Update: as of 2019 they have a new book out, Game-Theoretic Foundations for Probability and Finance.
They have added a game-theoretic derivation of Ito stochastic calculus, the capital asset pricing model, the equity premium, and portfolio theory.
You mention that a Martingale is a betting strategy where the player doubles their bet each time.
A Martingale is a fair game (i.e. the expected outcome is zero). If your outcome is given by a coin toss, and you receive only what you bet, then that is a Martingale game (you win X £ with probability 12 and lose X £ with probability 12 too ).
Then you could say that doubling your bet is a betting strategy on a Martingale game, BUT not that a Martingale game is a betting strategy where the player doubles their bet each time (in the same way that a dog is an animal but an animal is not a dog).
Does that make sense?
Other than that I’m very intrigued by the claim made. Definitely worth reading, but my hopes for something worthwhile are few :P
That does make sense, although it is not how they use the term in the book. In the book they explicitly rely on the betting system, for several reasons:
The name of the stochastic process in probability comes from the betting system, and they privilege the historical origination. This does introduce term confusion sometimes, but they apply it consistently in order to capture (what they claim to be) important subtleties.
I utterly failed to communicate this, but they pick up the use of the martingale directly from Jean Ville, who applied it to prove the difference between a frequency which converges via oscillation above and below vs. one which converges only from above—it turns out that if it converges from above, a martingale betting system can make a gambler infinitely rich. This was an important point against von Mises’ collectives, with respect to Kolmogorov’s measure.
They argue that Ville’s (generalized) use of the martingale betting system is a universal test for randmoness. Ville showed (they say) that for any event with probability 1, there is a non-negative martingale which diverges to infinity if that event fails. They extend Ville to say that the probability of an event is the smallest possible initial value for a non-negative martingale that eventually reaches or exceeds 1 if the event happens.
They start from that to build their game-theoretic notion of the martingale, which they then rely on to build the rest of their results.
This still leaves me in the wrong though, because once generalized it no longer is the doubling of bets—it is any system of varying bets. I will make an update to include some of this!
Big props for posting a book review—that’s always great and in demand. However, some points on (what I think is) good form while doing these:
a review on LW is not an advertisement; try to write reviews in a way that is useful to people who decide to not read the book
I also don’t care to see explicit encouragement to read a book—if what you relate about its content is tempting enough, I expect that I will have the idea to go and read it on my own
Update completed. Do you find it improved along those dimensions?
Yes! Not just improved, but leading by stellar example :)
This is very helpful, thank you. Between these points and the first two responses being things I really should have anticipated up front, I will make an update.