1. Introduction
Francis X. Diebold, Neil A. Doherty, and Richard J. Herring
Successful financial risk management requires constant grappling with the known, the unknown and the unknowable (âKuUâ). But think of KuU as more than simply an acronym for âthe known, the unknown, and the unknowableâ; indeed, we think of it as a conceptual framework. We believe that âKuU thinkingâ can promote improved decision makingâhelping us to recognize situations of K and u and U and their differences, using different tools in different situations, while maintaining awareness that the boundaries are fuzzy and subject to change.
Perhaps the broadest lesson is recognition of the wide applicability of KuU thinking, and the importance of each of K and u and U. KuU thinking spans all types of financial risk, with the proportion of uU increasing steadily as one moves through market, credit, and operational risks. In addition, KuU thinking spans risk measurement and management in all segments of the financial services industry, including investing, asset management, banking, insurance, and real estate. Finally, KuU thinking spans both the regulated and the regulators: regulatorsâ concerns largely match those of the regulated (risk measurement and management), but with an extra layer of concern for systemic risk.
1.1. KNOWLEDGE AS MEASUREMENT, AND KNOWLEDGE AS THEORY
Knowledge is both measurement and theory. Observed or measured facts about our world have no meaning for us outside our ability to relate them to a conceptual model. For example, the numerous stones we find with what appear to be reverse images of animals and plants would be unremarkable if it were not for their place in our intellectual model of the world we live in. Without the evolutionary theories associated with Darwin, the fossil record would be no more than a collection of pretty stones. And, indeed, without the pretty stones, Darwin may not have conceived his theory.
When we speak of knowledge, there is no bright line that separates our measurements from our theories. Though we may see the deficit at one, or the other, end of the spectrum, knowledge joins phenomenological observations with conceptual structures that organize them in a manner meaningful to our wider human experience. We would argue, for example, that both of the following assertions are true:
When you can measure what you are speaking about, and express it in numbers, you know something about it; but when you cannot measure it, when you cannot express it in numbers, your knowledge is of a meager and unsatisfactory kind: it may be the beginning of knowledge, but you have scarcely, in your thoughts, advanced to the stage of science. Lord Kelvin (Popular Lectures and Addresses, 1891â1894)
The whole machinery of our intelligence, our general ideas and laws, fixed and external object, principles, persons and gods, are so many symbolic, algebraic expressions. They stand for experience, experience which we are incapable of retaining and surveying in its multitudinous immediacy. We should flounder hopelessly, like the animals, did we not keep ourselves afloat and direct our course by these intellectual devices. Theory helps us to bear our ignorance of fact. George Santayana (The Sense of Beauty, 1896).
Thus, if we talk of what is known and what is unknown, we may be referring to the presence or absence of data to corroborate our theories, or to the inability of our theories to provide meaning to the curious phenomena we observe and measure.
For this volume, we have adopted the taxonomy of knowledge used in a famous article by Ralph Gomory (1995). Gomory classifies knowledge into the known, the unknown, and the unknowable, for which we adopt the acronym KuU. As applied to knowledge-as-measurement and knowledge-as-theory, we envision the KuU paradigm roughly as follows.
Knowledge as Measurement. The knowledge-as-measurement approach focuses on measuring possible outcomes with associated probabilities.
1. K refers to a situation where the probability distribution is completely specified. For example, the distribution of automobile or life insurance claims for an insurance company is more or less known. This is Frank Knightâs (1921) definition of riskâboth outcomes and probabilities are known.
2. u refers to a situation where probabilities cannot be assigned to at least some events. The systemic risk to financial systems and terrorism risk might fall into this category. This is Knightâs definition of uncertaintyâevents are known but probabilities are not.
3. U refers to a situation where even the events cannot be identified in advanceâneither events nor probabilities are known. Once they occur, they enter the domain of u. Retrospectively, the surge of asbestos claims for long-standing injury (real or imagined) is an example, as, indeed, are many legal actions caused by innovative legal theories.
Knowledge as Theory. The knowledge-as-theory approach focuses on the conceptual model that helps us to understand the underlying structure of the phenomenon of interest.
1. K refers to a situation where the underlying model is well understood. We may refer to this as a paradigm. This is not to say that the model is correct, only that experts are in broad agreement. For example, scientific models of evolution based on Darwin refer to a situation of scientific knowledge. We may not agree on all the details, but there is almost universal agreement among scientists on the broad structure. We might say there is âknowledgeâ on the broad principles of corporate governance, or risk-neutral options pricing. Thus, in short, K refers to successful theory.
2. u refers to a situation where there are competing models, none of which has ascended to the status of a paradigm. Credit risk management and operations risk management fall into this category. Other examples might include the performance of markets and financial institutions in emerging economies. If K refers to theory, then u refers to hypothesis or, more weakly, conjecture.
3. U refers to a situation with no underlying model (or no model with scientific credibility). This does not mean that we cannot conceivably form hypotheses, and even theory, in the future. But until some conceptual model is constructed, we cannot understand certain phenomena that we observe. Indeed, we may not even to be able to identify the phenomena because, in the absence of hypotheses or theory, it never occurs to us to look! For example, we would never look for black holes unless we had a theory about how matter behaves under extreme gravitational forces.
The two taxonomies are complementary. For example, the inability to specify the tail of a distribution might be due both to the absence of data and to deficiencies of statistical theory. Thus, innovations such as extreme value theory can lead to upgrading of knowledge (from U to u or from u to K) under both taxonomies. As another illustration, the absence of a theory for a yet-to-be-identified phenomenon is hardly surprising and the emergence of such events will generate an interest in both measurement and theory.
The various authors in this volume generally adopt the KuU framework (not surprisingly, as we did bully them gently toward a common terminology), though most use it to address knowledge-as-measurement issues, and some modify the framework. For example, Richard Zeckhauser notes that, as regards measurement, we could otherwise describe KuU as risk, uncertainty, and ignorance. Similarly, Howard Kunreuther and Mark Pauly use the alternative ambiguity in a similar manner to our u and Knightâs uncertainty. However, the most common chomping at the KuU bit was in insisting that we look at informational asymmetries. For example, Ken Scott looks at corporate governance in KuU terms, driven partly by informational (and skill) differences between managers and owners. Similarly, Zeckhauser observes that some investors have informational and skill advantages over others and then examines how uninformed investors, who recognize their inferior status, form strategies to benefit from the higher returns that can be earned from the knowledge and skills they lack.
A related theme that arises in some of the chapters is that the language used by different stakeholders depends on what is important to them. Clive Granger in particular notes that risk means different things to different people. Most particularly, many people think mostly of the downside of risk because that is what worries them. Thus, he emphasizes downside measures of risk, many of which (such as the various value at risk measures) have become increasingly important in risk management. Similarly, Scott notes that the conflict of interest that lies behind corporate governance is partly due to the fact that different stakeholders emphasize different parts of the distribution; undiversified managers may be more focused on downside risk than diversified shareholders.
1.2. KuU LESSONS FOR FINANCIAL MARKETS AND INSTITUTIONS
Here we highlight several practical prescriptions that emerge from KuU thinking, distilling themes that run through subsequent chapters. That we will treat K first is hardly surprising. Indeed, the existing risk management literature focuses almost exclusively on K, as summarized, for example, in the well-known texts of Jorion (1997), Doherty (2000), and Christoffersen (2003), and emphasized in the Basel II capital adequacy framework, which employs probabilistic methods to set minimum capital requirements.
Perhaps surprisingly in light of the literatureâs focus on K, however, we ultimately focus more on situations of u and U here and throughout. The reason is simple enough: reflection (and much of this volume) makes clear that a large fraction of real-world risk management challenges falls largely in the domain of u and U. Indeed, a cynic might assert that, by focusing on K, the existing literature has made us expert at the least-relevant aspects of financial risk management. We believe that K situations are often of relevance, but we also believe that u and U are of equal or greater relevance, particularly insofar as many of the âkiller risksâ that can bring firms down lurk there.
1.2.1. INVEST IN KNOWLEDGE
Although life is not easy in the world of K, it is easier in K than in u, and easier in u than in U. Hence, one gains by moving leftward through KuU toward knowledge, that is, from U to u to K. The question, then, is how to do it: How to invest in knowledge? Not surprisingly given our taxonomy of knowledge as measurement and knowledge as theory, two routes emerge: better measurement and better theory. The two are mutually reinforcing, moreover, as better measurement provides grist for the theory mill, and better theory stimulates improved measurement.
Better Measurement. Better measurement in part means better data, and data can get better in several ways. One way is more precise and timely measurement of previously measured phenomena, as, for example, with increased survey coverage when moving from a preliminary GDP release through to the âfinalâ revised value.
Second, better data can correspond to intrinsically new data about phenomena that previously did not exist. For example, exchange-traded house price futures contracts have recently begun trading. Many now collect and examine those futures prices, which contain valuable clues regarding the marketâs view on the likely evolution of house prices. But such data could not have been collected beforeâthey did not exist. Chapters like Bardhan and Edelsteinâs sweeping chronicle of KuU in real estate markets call to mind many similar such scenarios. Who, for example, could collect and analyze mortgage prepayment data before the development of mortgage markets and associated prepayment options?
Third, better data can arise via technological advances in data capture, transmission, and organization. A prime example is the emergence and increasingly widespread availability of ultra-high-frequency (trade-by-trade) data on financial asset prices, as emphasized in Andersen et al. (2006). In principle, such data existed whenever trades occurred and could have been collected, but it was the advent and growth of electronic financial marketsâwhich themselves require huge computing and database resourcesâthat made these data available.
Finally, and perhaps most importantly, better financial data can result from new insights regarding the determinants of risks and returns. It may have always been possible to collect such data, but until the conceptual breakthrough, it seemed pointless to do so. For example, traditional Markowitz risk-return thinking emphasizes only return mean and variance. But that approach (and its extension, Sharpeâs celebrated CAPM) assumes that returns are Gaussian with constant variances. Subsequent generations of theory naturally began to explore asset pricing under more general conditions, which stimulated new measurement that could have been done earlier, but wasnât. The resulting explosion of new measurement makes clear that asset returnsâparticularly at high frequenciesâare highly non-Gaussian and have nonconstant variances, and that both important pitfalls and important opportunities are embedded in the new worldview. Mandelbrot and Taleb, for example, stress the pitfalls of assuming normality when return distributions are in fact highly fat-tailed (badly miscalibrated risk assessments), while Colacito and Engle stress the opportunities associated with exploiting forecastable volatility (enhanced portfolio performance fuelled by volatility timing).
Thus far, we have treated better measurement as better data, but what of better tools with which to summarize and ultimately understand that data? If better measurement sometimes means better data, it also sometimes means better statistical/econometric modelsâthe two are obviously not mutually exclusive. Volatility measurement, for example, requires not only data but also models. Crude early proxies for volatility, such as squared returns, have been replaced with much more precise estimates, such as those based on ARCH models. This allows much more nuanced modeling, as, for example, in the chapter by Colacito and Engle, who measure the entire term structure of volatility. They construct a powerful new model of time-varying volatility that incorporates nonstationarity and hence changing distributions, nudging statistical volatility modeling clo...