Constant Proportion Debt Obligations, Zeno's Paradox, and the Spectacular Financial Crisis of 2008
CConstant Proportion Debt Obligations, Zeno’s Paradox,and the Spectacular Financial Crisis of 2008
Donald Richards ∗ and Hein Hundal † June 13, 2012
Abstract
We analyze a coin-tossing model used by a ratings agency to justify the inventionof constant proportion debt obligations (CPDOs), and prove that it was impossible forCPDOs to achieve in a finite lifetime the Cash-In event of doubling its capital. In thebest-case scenario in which the coin is two-headed, we show that the goal of attainingthe Cash-In event in a finite lifetime is precisely the goal, described more than twothousand years ago in Zeno’s Paradox of the Dichotomy, of obtaining the sum of aninfinite geometric series with only a finite number of terms. In the worst-case scenarioin which the coin is two-tailed, we prove that the Cash-Out event occurs in exactly tentosses.For the case in which the coin is fair, we show that if a CPDO were allowed to tossthe coin without regard for the Cash-Out rule then, eventually, the CPDO has a highprobability of attaining high net capital levels; however, hundreds of thousands of tossesmay be necessary to do so. Moreover, if after a large number of tosses the CPDO showsa loss then the probability is high that it will Cash-Out on the very next toss. In thecase of a CPDO that experiences a tail on the first toss, or on any early toss, we showthat, with high probability, the CPDO will have capital losses thereafter for hundreds oftosses, and we show that the sequence of net capital levels is a martingale. When theCash-Out rule is in force, we modify the Cash-In event to mean that the CPDO attainsa profit of 90% on its initial capital; then, we prove that the CPDO game, almost surely,will end in a finite number of tosses and the probability of Cash-Out is at least 89%.In light of these results, our fears about the durability of the on-going worldwidefinancial crisis are heightened by the existence of other financial derivatives more arcanethan CPDOs. In particular, we view askance all later-generation CPDOs that dependon an assumption of mean-reversion or utilize a betting strategy similar to that of theirfirst-generation counterparts. ∗ Department of Statistics, Penn State University, University Park, PA 16802. † Applied Research Laboratory, Penn State University, Science Park Road, State College, PA 16801. : Primary 91G20; Secondary 60K30.
Key words and phrases . AAA-rating; Binomial coefficient; Binomial distribution; Cash-In event;Cash-Out event; Central Limit Theorem; Collateralized debt obligation (CDO); Credit default swap;Credit derivative product company; Exotic synthetic CDO; Financial derivatives; Financial “engineer-ing”; Gambler’s fallacy; Geometric series; Golden Ratio; Law of Total Probability; LIBOR; Martingale;Mean-reversion; Pascal’s triangle; Ratings agency; Special purpose vehicle; Structured finance. a r X i v : . [ q -f i n . M F ] A p r Introduction
The tale is as old as the Eden Tree – and new as the new-cut tooth –For each man knows ere his lip-thatch grows he is master of Art and Truth;And each man hears as the twilight nears, to the beat of his dying heart,The Devil drum on the darkened pane: “You did it, but was it Art?”– Rudyard Kipling,
The Conundrum of the Workshops
Constant proportion collateralized debt obligations (CPDOs) were designed in 2006as a new product in the “synthetic” collateralized debt obligation market. CPDOswere developed by financial engineers (known colloquially as “quants”) at ABN AmroSecurities and first sold to the public in late 2006 with fanfare and AAA ratings [6].To create a CPDO, an investment bank forms a subsidiary corporation, called a“special purpose vehicle” (SPV). The SPV sells debt, called CPDO notes, to “investors”and deposits the proceeds in highly-rated collateral securities, such as government orhigh-grade commercial bonds, earning interest at a supposedly risk-free rate. The SPVnext uses the risk-free account as collateral to make a highly-leveraged sale of protectionon a widely-traded, investment-grade index of credit default swaps, such as the iTraxxEurope or Dow Jones CDX; at regular six-month intervals, the sale is wound down,any profits from the sale are paid into the deposit-account and any losses are paidout of the account. The SPV then repeats the process with a new highly leveragedsale of protection on the index, and the leverage factor is adjusted so that the SPV’sobligations can be met from the deposit-account. In general, the level of leveragingdepends on the difference between the current value of the SPV’s future liabilities andits net asset value [13, Section 1].CPDOs were designed to attract AAA ratings while paying a coupon rate substan-tially higher than similarly-rated instruments and without added risk. In particular,ABN Amro’s SURF notes supposedly represented AAA-rated debt, paying coupon rates2% higher than the London interbank offered rate, or LIBOR, and with no greater risk.Because of these purported properties, CPDOs received widespread acclaim.From the start, CPDOs were hailed as innovative. ABN Amro’s SURF was awardedthe “Deal of the Year” prize for 2006 by
Risk Magazine . The
International FinancialLaw Review , in giving SURF its “Award of 2006”, described it as the “Innovationof the Year ... [the] ultimate recognition for achievement in global capital markets.”SURF also received from
Euromoney magazine the accolade of “one of six ‘Deals of theYear 2006’ ” [2, p. 60, Appendix D]. Undoubtedly, the highest praise for SURF camefrom Bear, Stearns, a well-known investment bank, where analysts waxed religious inpraising CPDOs as the “Holy Grail of structured finance” [19]. Thus, by early 2007,estimates of CPDOs outstanding ranged as high as e that pay a risk-free 2% above CPDOs, then AAA-rated (CPDOs) that paya risk-free 2% above (CPDOs) , ad infinitum , a process that clearly is implausible.We were also intrigued by the remarkable, although not unprecedented, speed withwhich these novel financial entities floundered. In seeking to understand the basicreason for their rapid descent into insolvency, our curiosity was heightened by a newsarticle [24] whose last paragraph read as follows:“In a 2006 ‘primer’ on CPDOs written by a team of authors [...], Moody’sused an analogy that compared the product to a ‘coin-toss game.’ Thestrategy ‘is based on the notion that if ‘heads’ has appeared more frequentlythan expected, it is less likely to continue appearing,’ the report said.”In the study of random phenomena, such as the repeated tossing of a physical coin,it is well-known that the Law of Averages [12, p. 274] contradicts the “notion thatif ‘heads’ has appeared more frequently than expected, it is less likely to continueappearing.” Indeed, that notion is precisely the concept of gambler’s fallacy [11, 35].Consequently, we decided to analyze the probabilistic nature of CPDOs and, as weshall prove by means of elementary probability theory, there are serious flaws with thebasic structure of CPDOs defined by such a “coin-toss game.” As we investigate theprobabilistic behavior of such a game, what we shall discover is well-expressed by acomment made in another context by David Freedman [10, p. 102]:“At bottom, my critique is pretty simple-minded: Nobody pays muchattention to the assumptions, and the technology tends to overwhelmcommonsense.”In this paper, we shall use undergraduate-level methods to study the coin-tossingmodel for CPDOs provided in [34]. We begin by proving that, even if the coin is two-headed, the CPDO cannot achieve the Cash-In event of doubling the initial capitalwithin a finite lifetime, and we relate the case of the two-headed coin to a classicalparadox of Zeno, the Greek philosopher. If the coin is two-tailed, on the other hand,we prove that the Cash-Out event occurs in exactly ten tosses.3e consider the case in which the coin is fair, i.e., on any toss, heads and tails haveequal probability of occurring; however, all our calculations can be easily modified tocover the case of an unfair coin. We show that if a CPDO is allowed to toss the coinwithout regard to losses then, eventually, it has a high probability of amassing largenet capital; however, hundreds of thousands of tosses may be necessary to do so. Onthe other hand, if, after a large number of tosses the CPDO shows a loss then, withhigh probability, it will Cash-Out on the very next toss. In the case of a CPDO thatexperiences a tail on the first or on an early toss, we show that, with high probability,the CPDO will have capital losses after hundreds of additional tosses, and we obtainexplicit formulas for those probabilities. More generally, we show that the CPDObetting strategy does not have lack-of-memory: If at any given stage the CPDO showsa loss then, at all future stages, the probability of a loss, or of Cash-Out, generallyincreases. Further, we show that the sequence of net capital levels is a martingale.As we noted earlier, it is impossible for the CPDO to double its initial capital withina finite lifespan. Thus, for the case in which the CPDO is subject to the Cash-Out rule,we modify the Cash-In event to mean that the CPDO attains a profit of 90% on itsinitial capital. Then, we prove that the CPDO game, almost surely, will end in a finitenumber of tosses. Further, we prove that the probability that a CPDO will Cash-Outis at least 89%, and we provide in Section 7 figures that illustrate this phenomenon.On the basis of these results, we conclude that a CPDO was little more than arecipe for continual losses over lengthy periods, and with a high probability of default.Moreover, second-generation CPDOs [27] are likely to suffer the same fate if they employbetting strategies similar to those of their first-generation counterparts. We provide in this section a quotation taken from [34, pp. 3–4], a publication whichcame to our attention because of the newspaper article [24] (see also [7, 21]):“
An Analogy “To understand how the CPDO works, we can draw a parallel with a simplecoin-toss game. The outcome of the toss is either ‘heads’ or ‘tails’. Headsresults in a 100% return and tails a 100% loss.“The player has an initial stake of 1000, comprised of 100 from his ownpocket and 900 borrowed from a friend. At the outset his strategy is asfollows: if he succeeds in converting his stake into 2000 of winnings, he willstop and reimburse his friend, having thus multiplied his initial investmentby 11. This corresponds to the Cash-In Event.“At the same time, his friend is concerned about his stake and thus ifthe player loses more than 100, he will stop playing. This corresponds tothe Cash-Out Event. (For each toss, the player bets 1% of the difference4etween his current stake and 2000 – the simple rebalancing rule in ourexample.)“If he bets 10 from the initial stake on ‘heads’, there are two possible out-comes:“– ‘Heads’: the player’s stake rises to 1010, so at the next round he will betonly 9.90.“– ‘Tails’: he now has only 990, so at the next round he will bet 10.10.“Such a strategy is based on the notion that if ‘heads’ has appeared morefrequently than expected, it is less likely to continue appearing, and similarlyfor ‘tails.’ In other words, the strategy is based on the concept of ‘mean-reversion.’ ”
There are several problems with the strategy described in Section 2. We deal withthem in a number of subsections, as follows:
It is striking that the word “friend” is used to describe the entity from which “900” is“borrowed”, especially so when, in the real world, “an initial stake of 1000” correspondsto hundreds of millions of U.S. dollars. It is plausible that the use of the word “friend”may have encouraged among CPDO participants a subconscious lowering-of-the-guardas to possible negative outcomes of the “game.”In our view, CPDO participants would have been better served had the word“banker” been substituted for “friend.” Indeed, to judge by the way in which banks,broker-dealers, and hedge funds co-exist in the highly competitive financial world, itwould have been better had the word “friend” been replaced by “foe” throughout [34,pp. 3–4]. (Here, we note that some hedge funds have complained that their bankersor broker-dealers were overly abrupt in closing margin accounts after a short string of“tails,” thereby forcing the funds into insolvency.)Similar remarks apply to the use of the words “player” and “game.” Perhaps “bat-tle” would have been a more realistic term than “game”; similarly, substitute “warrior,”“gambler,” or “speculator” for “player.”In case a reader views our concerns about the language of the quotation as an over-occupation with mere semantics, we note that analysts at an investment bank alsoposed the question, “CPDO: Friend or foe?” [25, p. 572].
We have greater concerns with the “concept of ‘mean-reversion’ ” referred to in [34].5s we observed before, in the repeated tossing of a physical coin, the “notion thatif ‘heads’ has appeared more frequently than expected, it is less likely to continueappearing” is precisely the concept of gambler’s fallacy , the misbelief that the long-runrelative frequency of a random event will apply even in the short run. Classic examplesof gambler’s fallacy are well-known in casino gambling, where players often believe thata run of good luck is highly likely to follow a run of bad luck. Because individualoutcomes of the tosses of a coin are independent random events, however, a long stringof bad luck may well be followed by another long string of the same. (Indeed, it is well-known in probability theory that, in the indefinite tossing of a fair coin, the probabilitythat a given string of heads or tails will occur infinitely often is 100% [3, p. 61].)Tversky and Kahneman [35, p. 106] describe gambler’s fallacy as the phenomenonwherein“ . . . subjects [in a probability learning experiment] act as if every segmentof the random sequence must reflect the true proportion: if the sequencehas strayed from the population proportion, a corrective bias in the otherdirection is expected. . . . [The] heart of the gambler’s fallacy is a miscon-ception of the fairness of the laws of chance. The gambler feels that thefairness of the coin entitles him to expect that deviation in one directionwill soon be cancelled by a corresponding deviation in the other. Even thefairest of coins, however, given the limitations of its memory and mortalsense, cannot be as fair as the gambler expects it to be.”Freedman [11] has described connections between misconceptions in elementary prob-ability theory and gambler’s fallacy, which he describes as “a cognitive illusion whosepower is demonstrated by Cohen” [5].Although mean-reversion may have been observed in the past behavior of a financialtime series, it is not clear to us how it can be predicted that future observations of thesame time series will continue to exhibit mean-reversion. We sense an overuse of theconcept among financial analysts in the same way that na¨ıve casino gamblers hope formean-reversion away from runs of bad luck (but never from runs of good luck).
The betting strategy in the CPDO coin-toss model is curious in that it leads to a largerbet after a loss and a smaller bet after a win. Such a strategy is opposite to Kelly’sformula [22, 33] and to Whitworth’s formula [31, 32, 36], for those two formulas reducethe size of the bet following a loss and increase it after a win. The Kelly and Whitworthformulas are well-known to satisfy optimality properties with regard to maximizationof expected logarithmic utility or conservation of capital [29, 33], so the CPDO strategyis unlikely to enjoy similar properties.The CPDO strategy differs from standard methods in the theory of gambling sys-tems [3, 8, 26]. The strategy is not timid (i.e, betting a fixed amount on each toss),which may be commendable given the dangers of timidity [3, p. 108]. Also, the strategy6s not bold (i.e., betting either the amount needed to double the stake in one win, or thetotal remaining stake, whichever is smaller), which may be unwise given that boldnessis optimal in the sense that it maximizes the probability of attaining Cash-In [3, 26]. Wealso recall that, regardless of strategy, the probability of doubling the stake in a subfaircasino is at most 50% [8, p. 1]; thus, after borrowing 900 from a “friend,” and ignoringall costs of operations, a real-world CPDO had at most a 50% chance of attaining theCash-In event. In the case of timid play, bounds on the probability of Cash-In can beobtained by means of Chernoff’s inequality [28, Example 5f, pp. 452–453].
It will be simple to prove the following result.
Theorem 1.
In the best-case scenario in which the coin is two-headed, it is impossiblefor the CPDO to double its stake in a finite number of tosses. In the worst-case scenarioin which the coin is two-tailed, the CPDO will Cash-Out in exactly tosses. In the case of a two-headed coin, a CPDO will be ensnared in
The Dichotomy [39],a paradox of motion formulated by Zeno of Elea, the Greek philosopher (b. ∼
490 BC).In such a paradox the mythical Atalanta, a fast runner, runs 1 mile in the first minute,1/2 mile in the next minute, 1/4 mile in the next minute, and so on, in perpetuity.Atalanta would require an infinite amount of time to run two miles, for the distancescovered in successive minutes follow a geometric progression whose sum is less than 2 forany finite number of minutes. In the case of a CPDO equipped with a two-headed coin,we shall prove that its stake increases with each toss by successively smaller incrementsthat form a geometric progression; hence, similar to Atalanta, the CPDO will requirean infinite number of tosses to double its starting capital.Given the enormous amounts of time and financial resources committed to real-world CPDOs, and the enormity of any problems that would arise from flaws in thedesign of CPDOs, it is curious that the connection with Zeno’s paradox was overlookedentirely by CPDO inventors, purchasers, ratings agencies, and financial regulators. Let us turn now to a probabilistic analysis of the coin-toss CPDO model in [34]. Al-though such an analysis can be done using the advanced methods of Bilingsley [3,Section 7], Dubins and Savage [8], or Maitra and Sudderth [26], we will utilize onlyelementary probabilistic concepts familiar to undergraduate students, generally refer-ring to the undergraduate-level textbook [28]. Admittedly, we want to underscore that Admittedly, we ourselves overlooked it, and we are grateful to Moulinath Banerjee, of the Univer-sity of Michigan, for pointing out the connection to us. We can, at least, offer the excuses that we arenot quants, regulators, or ratings analysts, and in comparison with real-world CPDO operators, wehave scant financial resources to conduct full-time research on the subject. C , C , C , . . . represent the outcomes of successive coin tosses, where C j = (cid:40) +1 , if the j th toss results in heads − , if the j th toss results in tails. (4.1)Throughout the paper, the constants γ = 0 . , (4.2)and δ = 0 . , (4.3)will play a crucial role, and we display these equations to emphasize their repeated use.Readers may change the values of these constants to suit their individual needs, andthe calculations appearing in the paper can be modified in a straightforward manner.For k = 0 , , , . . . , denote by Y k the size of the stake held by the CPDO on the k thtoss. We take Y , the size of the initial stake, to be 1 unit, consisting of γ = 0 . − γ = 0 . to or below γ , the cut-off level. Hence, if the CPDO’s capital falls to γ , the bank will Cash-Outthe account rather than permit another coin toss and risk any loss on its loan.From the strategy of betting the proportion δ , of the difference between the goal of2 and the current stake, it follows that Y , Y , Y , . . . satisfy the recurrence relation Y k = Y k − + δ (2 − Y k − ) C k , (4.4)where k = 1 , , , . . . . To simplify the algebraic calculations in the ensuing analysis, itis convenient to work with X k = Y k −
1, the CPDO’s net profit (or loss) at the k thstage. It is easy to see from (4.4) that X = 0 and, for k = 1 , , , . . . , X k = X k − + δ (1 − X k − ) C k . (4.5)It is an undergraduate-level exercise to solve the recurrence relation (4.5), and then weobtain the following explicit formula for X k . Theorem 2.
For k = 0 , , , . . . , X k = 1 − k (cid:89) j =1 (1 − δC j ) . (4.6)8e can now verify the best- and worst-case scenarios described in Theorem 1. Proof of Theorem 1 . If the coin is two-headed then C k = +1 for all k and then, by (4.6),we have X k = 1 − (1 − δ ) k . Since δ < < (1 − δ ) k <
1, therefore 0 < X k < k , and this proves that Cash-In cannot occur in a finite number of tosses.If the coin is two-tailed then C k = − k ; hence, by (4.6), X k = 1 − (1 + δ ) k ,and X k decreases as k increases. Solving for k the equation 1 − (1 + δ ) k = − γ , weobtain k = ln(1 + γ ) / ln(1 + δ ) = 9 .
58; hence, Cash-Out occurs at the tenth toss.Henceforth, we assume in our analysis that the coins referred to in [34] are fair,i.e., that the probability of heads is 50%, and that individual coin tosses are mutuallyindependent, so that each toss provides no information whatsoever about the outcomeof other tosses. In probabilistic terminology, we assume that the random variables C , C , C , . . . are mutually independent, and that, for all j = 1 , , , . . . , P ( C j = +1) = P ( C j = −
1) = 1 / . Then, each C j has a Bernoulli distribution with probability of success 1 / long-term relative frequency, he would be assuming that,on average, he will outperform a majority of market participants in perpetuity , and thatthose opponents will continue to compete with him despite his noticeable advantage.On the other hand, if his coin is such that the probability of heads is less than 50%then there is no financial advantage to tossing given that, on each toss, his chances oflosing are greater than that of winning. Indeed, if the probability of heads is less than50% then, almost surely, the CPDO will Cash-Out in finitely many tosses.In any case, the ensuing analysis can be easily extended to apply to any choice ofthe probability of heads.We now have the following results for the mean, E ( X k ), and variance Var( X k ), ofthe sequence of net capital levels. Theorem 3.
For k = 1 , , , . . . , E ( X k ) = 0 and Var( X k ) = (1 + δ ) k − . (4.7) In particular,
Var( X k ) increases exponentially with k . − γ , the Cash-Outlevel, than to 1 so it is inauspicious that Var( X k ) grows exponentially with k while, atthe same time, E ( X k ) = 0 for all k . Moreover, that growth remains exponential in k , regardless of the size of δ ; so, even a more cautious CPDO which bets at each stage δ/
2, i.e., 50 basis points, of the difference between X k and the goal of 1, also will haveexponential growth in Var( X k ) while having E ( X k ) = 0 for all k .These observations alone would cause us to decline to commit billions of dollars tothe CPDO coin-toss model. Let us now consider the situation in which the CPDO’s banker is patient and will allowthe CPDO to toss the coin indefinitely without regard for the Cash-Out rule.Suppose that k , the number of tosses, is very large. Because the coin is fair, the C j in (4.1) will be approximately equally divided between +1 and −
1. Then, by (4.6), X k (cid:39) − (1 − δ ) k/ (1 + δ ) k/ = 1 − (1 − δ ) k/ . Intuitively, for large values of k , X k > X k → k → ∞ . Thus, if the bankeris sufficiently patient then, with high probability, the CPDO eventually will achievenear-Cash-In, in particular, will attain positive net capital levels. However, in makingthis statement more precise, we discover bad news hidden within this good news. Theorem 4.
For any t ∈ (0 , , lim k →∞ P ( X k ≥ t ) = 1 . (4.8) However, lim k →∞ P ( X k +1 ≤ − γ | X k <
0) = 1 . (4.9)The proof of this result is provided in Section 6.The result (4.8) shows that, eventually, the CPDO will, with high probability, attaina net capital level greater than any specified level t . Thus, a patient banker is highlyunlikely to have a loss on the CPDO loan. And yet, the banker must be prepared to beextremely patient for the CPDO to amass high net capital; for instance, if the CPDOdesires a 95% probability in order that X k ≥ − γ = 0 . k th toss then the probability is high that it will Cash-Outon the next toss. 10ecause P ( X k ≤ − γ ) ≤ P ( X k < k →∞ P ( X k ≤ − γ ) = 0 , i.e., the probability of Cash-Out eventually approaches 0. Nevertheless, the CPDOshould take little comfort from this result. Indeed, let µ = ln(1 − δ ) , σ = (cid:16) ln 1 + δ − δ (cid:17) , and denote by Φ( · ) the cumulative distribution function of the standard normal distri-bution; then, we shall prove later that, for large k , P ( X k ≤ − γ ) (cid:39) Φ (cid:16) − ln(1 + γ ) + kµ √ kσ (cid:17) . (4.10)On calculating the right-hand side of (4.10) for k ≥
25, we find that it increases steadilyfrom 2.7% (at k = 25) to 33.1% (at k = 1 , “Beware of such sirens, young man! ... above all avoid a martingale, ...” – W. M. Thackeray, The Newcomes: Memoirs Of A Most Respectable Family
The sequence X k , k = 1 , , , . . . turns out to have a property even more dangerousthan those described in Theorem 4. Theorem 5.
The sequence X , X , X , . . . is a martingale: E ( X k | X , . . . , X k − ) = X k − , almost surely, for all k ≥ . This result is, in fact, a special case of an example given by Williams [38, p. 95,Example (b)], who shows that if M = 1 and, for k ≥ M k is the product of k inde-pendent, non-negative random variables, each with mean 1, then M k is a martingale.The martingale nature of M k appears to have been first recognized by Abraham Wald[37] in his fundamental work in statistical inference.The connection between CPDOs and martingales was also noted by Jones [17] andKatsaros [21], albeit in the layman’s sense. Still, it is well-known in the popular liter-ature that martingale strategies are dangerous over protracted periods of time.The martingale property in Theorem 5 entails that, for a CPDO which has observedthe outcomes X , . . . , X k − , the expected outcome of the next toss is X k − , the lastobserved outcome . Such a result should be disquieting to the CPDO because, once it11egisters a net capital loss then, on average, the outcome of the next toss also is likelyto result in negative net capital. The martingale property does imply that if X k − > X k > − γ , the Cash-Out level, than it is to 1, the Cash-In level.Therefore, unless a CPDO expects a long run of heads from the outset , the martin-gale property of X k suggests that the betting strategy will be troublesome. To quantifythis statement, we now calculate the probabilities of net losses for small values of k .On the first toss, the CPDO will have a net capital loss if and only if the toss resultsin tails; hence, P ( X <
0) = 1 /
2. At the second toss, it follows from (4.6) that X < C = C = −
1, hence P ( X <
0) = 1 /
4. As for P ( X < X < C , C , and C are equal to −
1; therefore P ( X <
0) = 1 /
2. Also, by similar arguments, we obtain P ( X <
0) = 5 /
16 and P ( X <
0) = 1 / k th Toss k P ( X k < P ( X k < | X < P ( X k < k ≤
10. In all ten cases, thecorresponding probability of Cash-Out is 0, a result that surely greatly relieves theCPDO, but we shall expand on that point later – to the CPDO’s added discomfiture.Extending Table 1, we now present a closed-form expression for P ( X k <
0) formoderate values of k ; this result will explain the regular appearance of the value 50%in the table and show that it prevails for a moderately large number of tosses. Theorem 6.
For m = 1 , , . . . , , P ( X m <
0) = 12 − m +1 (cid:18) mm (cid:19) , (4.11) and, for m = 0 , , , . . . , , P ( X m +1 <
0) = 12 . (4.12)12 proof of this result is provided in Section 6.Another disconcerting property of the CPDO is its behavior conditional on having aloss on the first, or on an early toss. Consider, say, P ( X < | X < C from C , . . . , C , we obtain P ( X < | X <
0) = P (cid:16) (cid:89) j =1 (1 − δC j ) > | C = − (cid:17) = P (cid:16) (1 + δ ) (cid:89) j =2 (1 − δC j ) > (cid:17) . (4.13)Noting that C , C , C , C are uniformly and independently distributed over {− , +1 } ,and calculating the number of times that the inequality in (4.13) is satisfied, we obtain P ( X < | X <
0) = 11 /
16, or 68.8%. In like manner, we complete the last columnof Table 1, and we also extend those calculations by deriving explicit formulas for theconditional probabilities P ( X k < | X <
0) for higher values of k . Theorem 7.
For m = 1 , , . . . , , P ( X m < | X <
0) = 12 , (4.14) and, for m = 1 , , . . . , , P ( X m +1 < | X <
0) = 12 + 12 m +1 (cid:18) mm (cid:19) . (4.15)It is also the case that a loss at the second, or any early toss, leads to resultssimilar to Theorem 7. For instance, we shall derive explicit formulas for the conditionalprobabilities, P ( X k < | X < k . Theorem 8.
For m = 1 , , . . . , , P ( X m < | X <
0) = 12 + 12 m − (cid:18) m − m − (cid:19) ; (4.16) and for m = 1 , , . . . , , P ( X m +1 < | X <
0) = 12 . (4.17)Thus, the probabilities of net capital losses are substantially large for early tosses, asare the probabilities of future losses conditional on early losses, and these probabilitiesremain large for a considerable number of tosses.13 related problem is the evaluation of P ( X k +1 < | X k < k + 1)th toss, given that ithas a loss at the k th toss. Here, the results are startling and are nicely illustrated bythe example of P ( X < | X < X < C = C = −
1, i.e.,the first two tosses resulted in tails, then P ( X < | X <
0) = P ((1 − δC )(1 − δC )(1 − δC ) > | C = − , C = − P ((1 + δ ) (1 − δC ) > , because, with δ = 0 .
01, (1 + δ ) (1 − δC ) > C = ±
1. It is interesting that thephenomenon P ( X < | X <
0) = 1 remains valid for all δ such that (2 + δ ) / (1 + δ ) >
1, i.e., 0 < δ < ( √ − /
2, the reciprocal of the famous Golden Ratio.Generalizing the above example, we shall establish the following result that demon-strates astoundingly the lengthy effect of an early loss on the CPDO’s net capitallevels. By this result, if the CPDO shows a net loss on an early toss then, for nearlytwo hundred additional tosses, the probability of a successive loss is 1 or close to 1.
Theorem 9.
For m = 1 , , , . . . , , P ( X m +1 < | X m <
0) = 1 , (4.18) and for m = 0 , , , . . . , , P ( X m +2 < | X m +1 <
0) = 1 − m +1 (cid:18) m + 1 m + 1 (cid:19) . (4.19)We have studied, so far, the dependence between X k + l and X k for small values of l only. In the case of general l , we have the following result. Theorem 10.
For k, l = 1 , , , . . . , P ( X k + l < | X k < ≥ P ( X l <
0) (4.20) and P ( X k + l ≤ − γ | X k < ≥ P ( X l ≤ − γ ) . (4.21) More generally, if c ∈ R , c > − , and P ( X k < − c ) > then P ( X k + l ≤ − c | X k < − c ) ≥ P (cid:0) X l < c − c c (cid:1) . (4.22)Here again, the consequences for a CPDO are ominous. The inequality (4.20) meansthat given that the CPDO has a loss at the k th toss, the probability that it has a lossat l th additional tosses can never be smaller than the (unconditional) probability ofa loss at the l th toss. Hence, the CPDO is punished in perpetuity following an earlyloss. Similar remarks apply to the inequality (4.21) which asserts that, conditional ona loss, the probability of Cash-Out at any subsequent stage is not smaller than thecorresponding unconditional probability. 14 .3 Cash-Out: The case of the impatient banker “That which is painful, instructs.” – Benjamin Franklin Because not all bankers are infinitely patient, we now consider the case in whichthe CPDO receives from the banker a prompt Cash-Out call if its net assets falls to orbelow − γ .As we have shown in Section 3.4, it is impossible for the Cash-In event to occur ina finite number of tosses. Consequently, we modify the Cash-In rule throughout thissection of the paper to mean that the CPDO attains a net capital level of at least 1 − γ .Noting that the initial net capital is 0, if the CPDO attains modified Cash-In with X k ≥ − γ then the net return on personal capital will be at least (1 − γ ) /γ ; with γ = 0 .
1, this will result in a net return of at least 1 , α k := P ( − γ < X < − γ, . . . , − γ < X k < − γ ) , (4.23)the probability that the game lasts for at least k tosses, i.e., the CPDO will avoidCash-Out and Cash-In over the first k tosses. Then, we shall prove the following result. Theorem 11.
Almost surely, the CPDO game terminates in a finite number of tosses.That is, lim k →∞ α k = 0 . Finally, we obtain an upper bound for the probability that the CPDO will attainedCash-In. The result is likely to be painful, and hence instructive, for real-world CPDOsand so we present its proof immediately.
Theorem 12.
The probability that the CPDO attains Cash-In is at most γ + δ + γδ = 11 . . Proof . By Theorem 11, the game terminates in a finite number of tosses, almost surely.Then, we define P w to be the probability that the CPDO wins, i.e., attains Cash-In.Conditional on the event that the CPDO wins, let E w be the CPDO’s expected netcapital at that time; and, similarly, conditional on the event that the CPDO loses, i.e.,attains Cash-Out, let E l be the CPDO’s expected net capital at that time.The expected net capital for the CPDO after every toss is equal to the expectednet capital before the toss. The bankroll of the CPDO being bounded at all times, the15xpected net capital at the beginning and at the end of the game must be the same.Therefore, by the Law of Total Probability,0 = P w E w + (1 − P w ) E l . Solving this equation for P w , we obtain P w = − E l E w − E l . (4.24)If the CPDO loses, his last wager could not have exceeded (1 + γ ) δ ; thus, − γ ≥ E l ≥ − γ − (1 + γ ) δ = − ( γ + δ + γδ ) , equivalently, γ ≤ − E l ≤ γ + δ + γδ. On the other hand, it is obvious that E w ≥ − γ , hence E w − E l ≥ − γ + γ = 1.Combining these bounds on E l and E w − E l with equation (4.24) gives P w = − E l E w − E l ≤ γ + δ + γδ = 0 . . (cid:3) Simply put, even with the less stringent modified Cash-In rule, the probability thata CPDO will Cash-Out is at least 88.9%. We remark that the bound of 88.9% likelyis very close to the exact value of that probability; indeed, the results in Section 7 of1,000 CPDOs, each tossing their coin 50,000 times, starting with total capital of $1,000,and subject to the modified Cash-In rule, indicates that the proportion of CPDOs thatCash-Out is close to 89%.
The tale is far older than the Buttonwood Tree – and new as IPO veneer –For each trader knows ere his bonus grows he is master of Greed and Fear;And each trader hears as midnight nears, and his bedsheets tug and chafe,The Devil drum on the darkened pane: “You’ve bought it, but is it Safe?” – with apologies to Rudyard Kipling
Real-world CPDOs, invented at a time when the U.S. economy was near a zenith,were sold in August, 2006 as the economy weakened and began its descent into recession.As credit default indices and “bespoke” securities weakened, Lady Luck thus handedCPDO buyers a string of tails from the start. In an attempt to recover their losses,CPDO managers may have accelerated their dynamic hedging, tossing their coins faster,which gave them more tails. It is even possible that the process of dynamic hedging via rapid and forced asset sales may have increased the probability of subsequent tails.16he highly leveraged nature of CPDO notes, with notional amounts as high asfifteen times their nominal amounts, caused more grief. At 15X leverage, CPDOs have γ = 0 . − γ = 0 . δ , the betting proportion, in a futile attempt to returnto profits. By (4.7), an increase in δ simply increases fluctuations in X k and, at a timewhen credit-default indices were falling, would have accelerated the CPDOs’ losses.And so, down went the CPDOs. The obituaries provided in [4, 20, 23] are especiallyinstructful.In retrospect, we suspect that the problems encountered by Moody’s were causedless by software errors and more by the flawed nature of the CPDOs’ basic bettingstrategy. And yet, there are other prominent factors that, in our view, contributedgreatly to the CPDO debacle. We describe four of them as follows: First , the assessments of CPDOs provided by the Bank of England [1], the Bank forInternational Settlements [2], and the European Central Bank [9] seemed to have cometoo late to be of help to early purchasers of CPDO notes. (Moreover, we can sympa-thize with a purchaser who interpreted as laudatory the statement [1, p. 5], “Majorinnovations in securitisation can also be observed in the field of synthetic processes.To be mentioned are, for example, Constant Proportion Debt Obligations (CPDOs),products that have received particular interest as they are highly rated and at the sametime promise high interest.”)And yet, CPDO purchasers would have benefited from a close reading of the deeperanalyses provided in [1, p. 199–200] where CPDOs were described, ominously, as“apparently a ‘free lunch’ for investors”; in [9, p. 87]; and especially [2, Appendix D,pp. 60–72] and [27], where the warnings were severe.Indeed, sales of CPDOs could have been curtailed had the comprehensive assess-ment of CPDO risks in [2, loc. cit. ] been available in 2006. These risks, includingmanagement fees and mark-to-market accounting, surely would have given potentialbuyers pause. The Bank for International Settlements [2, p. 62] noted that an “ob-vious flaw in the first-generation CPDO design is its vulnerability to a legal form offront-running”, and opined that the CPDO market in 2007-2008 did not then “appearto be large enough for front-running to be a problem.” And yet, as any quant wouldagree, the efficient-market hypothesis would be violated were any legal form of front-running to remain unexploited by market participants. We think that wide knowledgeof the risk of front-running could have reduced the number of potential CPDO buyers.
Second , we did not incorporate into our analysis of CPDOs risks such as capitalgains taxes, commissions on sales and purchases, management fees, mark-to-marketaccounting, potential front-running, or a coin that is slightly biased towards tails. Hadwe done so, the outcome would have been considerably worse than those illustratedby our calculations. We encourage our readers to calculate the various probabilities inSection 4 for the case in which p = 18 /
38, the probability of success at the roulette-wheel game of red-and-black. It is well-known [3, p. 107] that, with timid or bold play,17 casino gambler’s probability of success decreases as p changes from 1 / /
38, andreaders who perform similar calculations for the CPDO coin-tossing model will findthat the probabilities of Cash-In also decrease.
Third , as we explained in the Introduction, the SPVs used the proceeds from thesales of their CPDO notes to make a highly leveraged sale of an index of credit defaultswaps. As it turns out, such an action in late 2006 was most unwise and reckless. Toexplain, we first provide a concise, down-to-earth explanation of the nature of creditdefault swaps, taken from Herzog [15, p. 50]:“Credit default swaps are essentially unregulated insurance policies coveringthe losses on securities in case a triggering event occurs. Such an event couldbe a fire, a plane crash, or a mortgage foreclosure. Financial institutionsbuy credit default swaps to protect themselves against the adverse effects ofsuch events. In this sense default swaps are similar to fire insurance in thesense that a homeowner buys fire insurance to protect his investment in casehis house burns down. However, unlike fire insurance, credit default swapscan also be used as a purely speculative ‘investment.’ In this case, creditdefault swaps are like buying insurance against the risk that my neighbor’shouse burns down. Whereas with my own house, I have what insuranceprofessionals call an ‘insurable interest,’ with my neighbor’s house I do not.The situation with credit default swaps is similar to bookies trading bets,with banks and hedge funds gambling on whether an investment (say, acollection of subprime mortgages bundled into a security) will succeed orfail.”Thus, in late 2006, the buyers of CPDO notes were undertaking the risks that gowith being a highly leveraged, i.e., highly undercapitalized, insurance company. More-over, in purchasing an index of credit default swaps in late 2006, the CPDO note holderswere insuring not simply one house in one city but entire and adjoining neighborhoodsin the same city. With subprime mortgage lenders making loans frenziedly, huge num-bers of homeowners had become accustomed to playing with (financial) fireworks intheir living rooms in late 2006. Consequently, housing fires were not unlikely to oc-cur. Given that a fire in one house was guaranteed to ignite fires in adjacent houses,thence throughout the neighborhood, and eventually across the entire city, a CPDOnote holder, in his role as an unwise insurance company or as a very na¨ıve bookie, wasdoomed to bankruptcy.
Fourth , the quants created in the case of a CPDO a structure that violated theefficient-market hypothesis, an axiom of the field of “financial engineering.” It surely isnot a good omen when researchers violate their own axioms. Moreover, it strengthensthe well-known complaints that quants, under pressure from investment bankers todevise structured products to be sold to the public, are incapable of maintaining evenquasi-scientific standards; and that the executives and boards of directors of ratingsagencies are not competent to assess the quants’ activities.18n light of the results that we have derived, we fear that the on-going worldwidefinancial crisis will be prolonged by other arcane financial derivatives with structures asbizarre as CPDOs. Here, we refer to credit derivative product companies, some of whichcarry leverage of up to 80-to-1; CDOs-squared and -cubed; and exotic synthetic CDOs.Frankly, it is difficult to escape the conclusion [4] that some of those instruments arelittle more than “dead men walking.” We expect that second-generation CPDOs [27]also will likely cause grief if they employ a betting strategy similar to that of their first-generation counterparts, or depend crucially on an assumption of mean-reversion, orin any way represent a “free lunch.” With negative outcomes from these instrumentslikely to impact the financial markets broadly, we expect that the markets’ currentdistrust of ratings and regulatory agencies is likely to prevail for years to come.The “don’t-blame-the-quants” opinion of [30] notwithstanding, we believe that thequants indeed should shoulder the blame for the CPDO debacle and its unsettlingeffects on the financial markets. Insofar as the losses on CPDOs are concerned, itwould appear that investors were beguiled by the quants and their sirenical financialdevices in much the same way that boaters on the Rhine were mesmerized by Lorelei and her captivating voice. Proof of Theorem 2 . By the recurrence formula (4.5) we see that, for k = 1 , , , . . . ,1 − X k = 1 − X k − − δ (1 − X k − ) C k = (1 − X k − )(1 − δC k ) , On applying this formula with k replaced by k − , k − , . . . ,
1, we obtain1 − X k = (1 − X k − )(1 − δC k )= (1 − X k − )(1 − δC k − )(1 − δC k )...= (1 − X )(1 − δC )(1 − δC ) · · · (1 − δC k − )(1 − δC k ) . Substituting X = 0 and solving for X k , we obtain (4.6). Proof of Theorem 3 . Noting that C , . . . , C k are mutually independent and, by (4.1),that E ( C j ) = 0 for all j , it follows from (4.6) that E ( X k ) = 1 − k (cid:89) j =1 E (1 − δC j ) = 1 − k (cid:89) j =1 (cid:0) − δE ( C j ) (cid:1) = 0 . The mythical Germanic siren who, sitting on a high cliff while combing her beautiful hair andsinging sensually, tempted boaters on the River Rhine to their deaths. Webster’s Online Dictionarynotes that “ ‘Lorelei’ is a name that signifies or is derived from: ‘an ambush cliff’.” X k ) = Var k (cid:89) j =1 (1 − δC j )= E k (cid:89) j =1 (1 − δC j ) − (cid:0) E k (cid:89) j =1 (1 − δC j ) (cid:1) = (cid:0) k (cid:89) j =1 E (1 − δC j ) (cid:1) − . Since C j ≡ E ( C j ) = 0, we have E (1 − δC j ) = E (1 − δC j + δ C j ) = 1 + δ , andthen (4.7) follows immediately.Finally, (4.7) implies that for large k , Var( X k ) ∼ exp( k ln(1 + δ )), so Var( X k )increases exponentially with k . Proof of Theorem 4 . By (4.6), the inequality X k ≥ t is equivalent to k (cid:89) j =1 (1 − δC j ) ≤ − t. (6.1)Let W j = ln(1 − δC j ), where j = 1 , . . . , k ; by taking natural logarithms in the inequality(6.1), we see that (6.1) is equivalent to (cid:80) kj =1 W j ≤ ln(1 − t ). Therefore, P ( X k ≥ t ) = P (cid:16) k (cid:88) j =1 W j ≤ ln(1 − t ) (cid:17) . (6.2)Since C , . . . , C k are independent, identically distributed random variables, then so are W , . . . , W k . Further, it is simple to show that the mean and variance of each W j are µ = ln(1 − δ ) and σ = (cid:16) ln 1 + δ − δ (cid:17) , respectively, and µ < δ < · ) denote the cumulative distribution function of the standard normal distri-bution. On applying the Central Limit Theorem [28, p. 434] to (cid:80) kj =1 W j , we obtainlim k →∞ P (cid:16) k (cid:88) j =1 W j ≤ ln(1 − t ) (cid:17) = lim k →∞ Φ (cid:16) ln(1 − t ) − kµ √ kσ (cid:17) = 1 . (6.3)At this stage we can verify the remark, made below the statement of Theorem 4,regarding the number of tosses necessary for the CPDO to attain X k ≥ − γ withprobability 95%. Indeed, on applying the approximation, P ( X k ≥ t ) (cid:39) Φ (cid:16) ln(1 − t ) − kµ √ kσ (cid:17) , k , we obtain k (cid:39) , . { C = − , . . . , C k = − } implies { X k < } . Therefore, P ( X k < ≥ P ( C = − , . . . , C k = −
1) = 2 − k > , so any conditional probability that is conditioned on the event { X k < } is well-defined.By definition, P ( X k +1 < | X k <
0) = P ( X k +1 < , X k < P ( X k < . (6.4)By the Law of Total Probability [28, p. 72], P (cid:0) X k +1 < , X k < (cid:1) = P (cid:0) X k +1 < , X k < | C k +1 = − (cid:1) P ( C k +1 = − P (cid:0) X k +1 < , X k < | C k +1 = +1 (cid:1) P ( C k +1 = +1) . (6.5)By (4.5), P (cid:0) X k +1 < , X k < | C k +1 = − (cid:1) = P (cid:0) X k − δ (1 − X k ) < , X k < (cid:1) = P ( X k < , and also, P (cid:0) X k +1 < , X k < | C k +1 = +1 (cid:1) = P (cid:0) X k + δ (1 − X k ) < , X k < (cid:1) = P (cid:0) X k < − δ − δ (cid:1) . Therefore, by equations (6.4) and (6.5), P ( X k +1 < | X k <
0) = P (cid:0) X k < (cid:1) + P (cid:0) X k < − δ − δ (cid:1) P ( X k < (cid:104) P (cid:0) X k < − δ − δ (cid:1) P ( X k < (cid:105) . (6.6)For large values of k , it follows from (6.3) that P (cid:0) X k < − δ − δ (cid:1) (cid:39) Φ (cid:16) ln(1 − δ ) + kµ √ kσ (cid:17) (6.7)and P ( X k < (cid:39) Φ (cid:16) kµ √ kσ (cid:17) . Therefore, lim k →∞ P (cid:0) X k < − δ − δ (cid:1) P ( X k <
0) = lim k →∞ Φ (cid:16)(cid:0) ln(1 − δ ) + kµ (cid:1) / √ kσ (cid:17) Φ (cid:0) kµ/ √ kσ (cid:1) = 1 , k →∞ P ( X k +1 < | X k <
0) = 1.The proof that lim k →∞ P ( X k +1 ≤ − γ | X k <
0) = 1 proceeds by a similar argument.By the same approach that led to (6.6), we obtain P ( X k +1 ≤ − γ | X k <
0) = P (cid:0) X k ≤ − γ + δ δ (cid:1) + P (cid:0) X k < − γ − δ − δ (cid:1) P ( X k < . On approximating each term on the right-hand side as in (6.7), and then applyingL’Hospital’s rule, we obtain lim k →∞ P ( X k +1 ≤ − γ | X k <
0) = 1.
Proof of Theorem 5 . In calculating X k from earlier outcomes { X , . . . , X k − } , it followsfrom equation (4.5) that only X k − and C k are germane. Because C k is independent of X k − and E ( C k ) = 0, we obtain E ( X k | X , . . . , X k − ) = E ( X k | X k − )= X k − + δ (1 − X k − ) E ( C k ) = X k − . (cid:3) Proof of Theorem 6 . By (4.6), P ( X m <
0) = P (cid:0) m (cid:89) j =1 (1 − δC j ) > (cid:1) . Suppose that i of the random variables C , . . . , C m are equal to − C j are equal to +1. Then we need to determine all values of i such that the product (cid:81) mj =1 (1 − δC j ) ≡ (1 − δ ) m − i (1 + δ ) i >
1. To that end, it is simple to verify that thesequence a i = (1 − δ ) m − i (1 + δ ) i , where i = 0 , . . . , m , is strictly increasing as i increases; in fact, a i +1 a i = 1 + δ − δ > , for all i = 0 , . . . , m −
1. Note also that a m = (1 − δ ) m (1+ δ ) m = (1 − δ ) m <
1, so a i < i = 0 , . . . , m , and therefore the first value of i for which it is possible to have a i > i = m + 1. By solving the inequality a m +1 >
1, i.e., (1 − δ ) m − (1 + δ ) m +1 > m ≤ m ,(1 − δ ) m − i (1 + δ ) i (cid:40) < , i = 0 , . . . , m,> , i = m + 1 , . . . , m. Next, for each i ≥ m + 1, we count the number of cases in which i of C , . . . , C m areequal to − m − i variables C j are equal to +1. This is a standard22alculation [28, p. 6] in an undergraduate course on probability theory; the answer isthe binomial coefficient, (cid:18) mi (cid:19) = (2 m )!(2 m − i )! i ! . Because there are 2 m equally likely possible choices for the sequence C , . . . , C m , itfollows that, for m ≤ P ( X m <
0) = 12 m m (cid:88) i = m +1 (cid:18) mi (cid:19) . (6.8)This sum is well-known; once we recognize that it is the sum of nearly the last half ofall entries in the 2 m -th row of Pascal’s triangle, we deduce that it equals2 m − − (cid:18) mm (cid:19) . On substituting this result into (6.8), we obtain (4.11).For the case in which k is odd, say, k = 2 m + 1, the calculations are similar. Byapplying the same arguments as before we deduce that, for all m ≤ P ( X m +1 <
0) = 12 m +1 2 m +1 (cid:88) i = m +1 (cid:18) m + 1 i (cid:19) ;noting that this sum is precisely the sum of the last half of all numbers in the (2 m +1)-throw of Pascal’s triangle, we find that it equals 2 m +1 / m . Therefore, for m ≤ P ( X m +1 <
0) = 2 m / m +1 = 1 / Proof of Theorem 7 . Let k = 2 m , where m is a positive integer. By the same argumentas at (4.13), P ( X m < | X <
0) = P (cid:16) (1 + δ ) m (cid:89) j =2 (1 − δC j ) > (cid:17) . (6.9)To calculate this probability, we need to enumerate the number of cases in which theinequality on the right-hand side of (6.9) is satisfied as each of C , . . . , C m variesuniformly and independently over the set {− , +1 } .Suppose that i of the variables C , . . . , C m are equal to − C j are equal to +1; then we need to determine all values of i for which the product(1 + δ ) (cid:81) mj =2 (1 − δC j ) ≡ (1 − δ ) m − − i (1 + δ ) i +1 >
1. Now, it is simple to verify thatthe sequence a i = (1 − δ ) m − − i (1 + δ ) i +1 , where i = 0 , . . . , m −
1, is strictly increasing as i increases; in fact, a i +1 a i = 1 + δ − δ > , i = 0 , . . . , m −
1. Also, note that a m − = (1 − δ ) m (1 + δ ) m = (1 − δ ) m <
1, so a i < i = 0 , . . . , m −
1, and therefore the first value of i for which it is possible tohave a i > i = m . By solving the inequality a m >
1, i.e., (1 − δ ) m − (1 + δ ) m +1 > m ≤ − δ ) m − − i (1 + δ ) i +1 (cid:40) < , i = 0 , . . . , m − ,> , i = m, . . . , m − . Next, for each i ≥ m , we count the number of cases in which i of C , . . . , C m areequal to − m − − i variables C j are equal to +1. Similar to theforegoing, the answer is the binomial coefficient, (cid:18) m − i (cid:19) = (2 m − m − − i )! i ! . Because there are 2 m − equally likely possible choices for the sequence C , . . . , C m , itfollows that, for m ≤ P ( X m < | X <
0) = 12 m − m − (cid:88) i = m (cid:18) m − i (cid:19) . This sum is well-known, being the sum of the last half of all entries in the (2 m − m − / m − . Consequently, for m ≤ P ( X m < | X <
0) = 2 m − / m − = 1 / k is odd, say, k = 2 m + 1, the calculations are similar. Wededuce, first, that for all m ≤ P ( X m +1 < | X <
0) = 12 m m (cid:88) i = m (cid:18) mi (cid:19) ;second, by comparing this sum to the total of the last half of all numbers in the 2 m -throw of Pascal’s triangle, we obtain m (cid:88) i = m (cid:18) mi (cid:19) = 2 m − + 12 (cid:18) mm (cid:19) ; (6.10)and this leads to the stated result. Proof of Theorem 8 . The proof of this result is similar to the proof of Theorem 7.Nevertheless, we provide complete details in order that the paper be self-contained.Let k = 2 m , where m is a positive integer, m ≥
2. By the same argument as at(4.13), P ( X m < | X <
0) = P (cid:16) (1 + δ ) m (cid:89) j =3 (1 − δC j ) > (cid:17) . (6.11)24o calculate this probability, we need to enumerate the number of cases in which theinequality on the right-hand side of (6.11) is satisfied as each of C , . . . , C m variesuniformly and independently over the set {− , +1 } .Suppose that i of the variables C , . . . , C m are equal to − C j are equal to +1; then we need to determine all values of i for which the product(1 + δ ) (cid:81) mj =3 (1 − δC j ) ≡ (1 − δ ) m − − i (1 + δ ) i +2 >
1. Now, it is simple to verify thatthe sequence a i = (1 − δ ) m − − i (1 + δ ) i +2 , where i = 0 , . . . , m −
2, is strictly increasing as i increases; in fact, a i +1 a i = 1 + δ − δ > , for all i = 0 , . . . , m −
2. Also, note that a m − = (1 − δ ) m (1 + δ ) m = (1 − δ ) m < a i < i = 0 , . . . , m −
2, and therefore the first value of i for which it ispossible to have a i > i = m −
1. By solving the inequality a m − >
1, i.e.,(1 − δ ) m − (1 + δ ) m +1 >
1, we obtain m ≤ m ,(1 − δ ) m − − i (1 + δ ) i +2 (cid:40) < , i = 0 , . . . , m − ,> , i = m − , . . . , m − . Next, for each i ≥ m −
1, the number of cases in which i of C , . . . , C m are equal to − m − − i variables C j all are equal to +1 is the binomial coefficient, (cid:18) m − i (cid:19) = (2 m − m − − i )! i ! . Because there are 2 m − equally likely possible choices for the sequence C , . . . , C m , itfollows that P ( X m < | X <
0) = 12 m − m − (cid:88) i = m − (cid:18) m − i (cid:19) . On evaluating this sum by means of (6.10) and substituting its value into the previousequation, we obtain the desired result.For the case in which k is odd, say, k = 2 m + 1 with m ≥
1, we deduce by a similarargument that, for m ≤ P ( X m +1 < | X <
0) = 12 m − m − (cid:88) i = m − (cid:18) m − i (cid:19) = 12 . (cid:3) Proof of Theorem 9 . By the Law of Total Probability, P ( X k +1 < | X k <
0) = P ( X k +1 < | X k < , C k +1 = − P ( C k +1 = − P ( X k +1 < | X k < , C k +1 = +1) P ( C k +1 = +1)= [ P ( X k +1 < | X k < , C k +1 = − P ( X k +1 < | X k < , C k +1 = +1)] .
25t is clear that a net capital loss at the k th stage followed by tails at the next tossguarantees a net loss at the ( k + 1)th stage. Therefore, P ( X k +1 < | X k < , C k +1 = −
1) = 1 , and so we have P ( X k +1 < | X k <
0) = 12 (cid:16) P ( X k +1 < | X k < , C k +1 = +1) (cid:17) . (6.12)Next, by the definition of conditional probability, P ( X k +1 < | X k < , C k +1 = +1) = P ( X k +1 < , X k < , C k +1 = +1) P ( X k < , C k +1 = +1) , If C k +1 = +1 then, obviously, X k +1 > X k ; we see now that the events { X k +1 < , X k < , C k +1 = +1 } and { X k +1 < , C k +1 = +1 } are equivalent, and hence have the sameprobability. Therefore, P ( X k +1 < | X k < , C k +1 = +1) = P ( X k +1 < , C k +1 = +1) P ( X k < , C k +1 = +1) . (6.13)We proceed now to evaluate the numerator and denominator above. Because X k and C k +1 are mutually independent, it follows from Theorem 6 that P ( X k < , C k +1 = +1) = 12 P ( X k < (cid:40) (cid:16) − m +1 (cid:0) mm (cid:1)(cid:17) , k = 2 m, ≤ m ≤ , , k = 2 m + 1 , ≤ m ≤ . As for the numerator in (6.13), it follows by (4.6) and the mutual independence of the C j that P ( X k +1 < , C k +1 = +1) = P (cid:0) (1 − δ ) k (cid:89) j =1 (1 − δC j ) > , C k +1 = +1 (cid:1) = P (cid:0) (1 − δ ) k (cid:89) j =1 (1 − δC j ) > (cid:1) . We calculate this probability by proceeding as in Theorem 7, thereby obtaining P ( X k +1 < , C k +1 = +1) = (cid:16) − m +1 (cid:0) mm (cid:1)(cid:17) , k = 2 m, ≤ m ≤ , (cid:16) − m +1 (cid:0) m +1 m +1 (cid:1)(cid:17) , k = 2 m + 1 , ≤ m ≤ . For k = 2 m with 1 ≤ m ≤
99, the numerator and denominator of (6.13) are equal;then, by (6.12) and (6.13), P ( X k +1 < | X k <
0) = 1.26or k = 2 m + 1 where 1 ≤ m ≤
98, we use the above results to obtain P ( X m +2 < | X m +1 < , C m +2 = +1) = 1 − m (cid:18) m + 1 m + 1 (cid:19) , and then it follows from (6.12) that P ( X m +2 < | X m +1 <
0) = 1 − m +1 (cid:18) m + 1 m + 1 (cid:19) . (cid:3) Proof of Theorem 10 . Because (4.20) and (4.21) are special cases of (4.22), it sufficesto prove the latter only.If (cid:81) kj =1 (1 − δC j ) > c and (cid:81) k + lj = k +1 (1 − δC j ) ≥ (1 + c ) / (1 + c ) then, clearly, (cid:81) k + lj =1 (1 − δC j ) ≥ c . Therefore, P (cid:16) k + l (cid:89) j =1 (1 − δC j ) ≥ c , k (cid:89) j =1 (1 − δC j ) > c (cid:17) ≥ P (cid:16) k + l (cid:89) j = k +1 (1 − δC j ) ≥ c c , k (cid:89) j =1 (1 − δC j ) > c (cid:17) = P (cid:16) k + l (cid:89) j = k +1 (1 − δC j ) ≥ c c (cid:17) P (cid:16) k (cid:89) j =1 (1 − δC j ) > c (cid:17) = P (cid:16) l (cid:89) j =1 (1 − δC j ) ≥ c c (cid:17) P (cid:16) k (cid:89) j =1 (1 − δC j ) > c (cid:17) , where the last two equalities are due to C , . . . , C k + l being independent, identicallydistributed random variables. Applying (4.6) to express the above inequalities in termsof the X k , we obtain P ( X k + l ≤ − c , X k < − c ) ≤ P (cid:0) − X l ≥ c c (cid:1) P ( X k < − c )= P (cid:0) X l ≤ c − c c (cid:1) P ( X k < − c ) , and this result is equivalent to (4.22).Finally, we consider Theorem 11. To prove that result, we require a preliminarycalculation. Lemma 13.
If k is an even positive integer then, for j = 0 , . . . , k , (cid:18) kj (cid:19) < k (cid:16) πk (cid:17) / exp (cid:16) k (cid:17) . roof . By Stirling’s inequalities [28, p. 43], we have(2 πk ) / k k exp( − k ) < k ! < (2 πk ) / k k exp (cid:16) − k + 112 k (cid:17) . Also, it is well-known [28, p. 156] thatmax j =0 , ,...,k (cid:18) kj (cid:19) = (cid:18) kk/ (cid:19) = k !(( k/ . Therefore, for j = 0 , . . . , k , (cid:18) kj (cid:19) ≤ k !(( k/ < (2 πk ) / k k exp( − k + k ) (cid:0) (2 πk/ / ( k/ k/ exp( − k/ (cid:1) = 2 k (cid:16) πk (cid:17) / exp (cid:16) k (cid:17) . (cid:3) Proof of Theorem 11 . Let N k be the number of heads obtained among the first k tossesof the coin. Then X k = 1 − k (cid:89) i =1 (1 − δC i )= 1 − (1 − δ ) N k (1 + δ ) k − N k = 1 − (1 + δ ) k (cid:16) − δ δ (cid:17) N k . Therefore, − γ < X k < − γ if and only if − γ < − (1 + δ ) k (cid:16) − δ δ (cid:17) N k < − γ. Solving this inequality for N k yields k ln(1 + δ ) − ln(1 + γ )ln(1 + δ ) − ln(1 − δ ) < N k < k ln(1 + δ ) − ln γ ln(1 + δ ) − ln(1 − δ ) , and the length of this interval range for N k isln(1 + γ ) − ln γ ln(1 + δ ) − ln(1 − δ ) , (6.14)a finite number that we denote by β . Hence, P ( − γ < X k < − γ ) ≤ β max j =0 ,...,k P ( N k = j ) . (6.15)28ecause N k has a binomial distribution, N k ∼ B ( k, / P ( N k = j ) = 2 − k (cid:18) kj (cid:19) < (cid:16) πk (cid:17) / exp (cid:16) k (cid:17) , (6.16)where the inequality in (6.16) follows from Lemma 13 when k is even. From inequalities(6.15) and (6.16) we conclude that P ( − γ < X k < − γ ) < β (cid:16) πk (cid:17) / exp (cid:16) k (cid:17) (6.17)when k is even. By (4.23), α k is monotonic decreasing and0 ≤ α k ≤ P ( − γ < X k < − γ );therefore, by (6.17), α k → k → ∞ . In the following figures, we present the graphical results of the simulations of 1,000CPDOs, each tossing their coin 50,000 times, starting with total capital of $1,000, andsubject to the modified Cash-In rule. As the simulations indicate, the proportion ofCPDOs that Cash-Out is close to 89%. 290123456789 eferences [1] Bank of England.
Quarterly Bulletin.
Volume , No. 2, 2007.[2] Basel Committee on Banking Supervision. Credit Risk Transfer: Developmentsfrom 2005 to 2007.
Bank for International Settlements, Basel, Switzerland, 2008.[3] P. Billingsley.
Probability and Measure , third edition. Wiley, New York, 1995.[4] A. Capon. Obituary: the CPDO, 2006–2008. No flowers.
Euromoney , November,2008.[5] L. J. Cohen. Can human irrationality be experimentally demonstrated? (withdiscussion).
Behavioral and Brain Sciences , (1981), 317–370.[6] P. J. Davies. Questions lie behind CPDO hype. Financial Times . November 13,2006.[7] J. Dorn. Modeling of CPDOs identifying optimal and implied leverage, Universit´eParis, Panth´eon-Sorbonne, November, 2007.[8] L. E. Dubins and L. J. Savage.
Inequalities for Stochastic Processes: How toGamble If You Must . Dover Publications, 1976.[9] European Central Bank.
Financial Stability Review.
Frankfurt, Germany, June,2007.[10] D. A. Freedman. As others see us: A case study in path analysis.
Journal ofEducational Statistics , (1987), 101–128.[11] D. A. Freedman. [Bayes for beginners? Some reasons to hesitate]: Discussion. TheAmerican Statistician , (1997), 266–268.[12] D. Freedman, R. Pisani, and R. Purves. Statistics , third edition. Norton & Co.,New York, 1998.[13] M. B. Gordy and S. Willeman. Constant proportion debt obligations: A post-mortem analysis of rating models. Finance and Economics Discussion Series, Paper2010-05, The Federal Reserve Board, January 2010.[14] J. Glover. CPDOs rated AAA may risk default, CreditSights says.
Bloomberg ,September 6, 2007.[15] T. N. Herzog. History of mortgage finance with an emphasis on mortgage insur-ance. In:
Housing in Retirement Monograph (A. Rappaport and S. Siegel, editors),Society of Actuaries, 2009.[16] International Monetary Fund.
Global Financial Stability Report: Market Develop-ments and Issues.
Washington, DC, 2007.4017] S. Jones, Rated AAA, but riskier than junk– think CPDOs.
Financial Times ,September 6, 2007.[18] S. Jones. CPDOs approach cash-out events.
Financial Times , September 18, 2008.[19] S. Jones. When junk was gold – Part 2.
Financial Times , October 17, 2008.[20] S. Jones. Requiem for the CPDO.
Financial Times , October 17, 2008.[21] G. Katsaros. New CDO-related products. R. W. Pressprich & Co., April, 2007.[22] J. L. Kelly. A new interpretation of information rate.
Bell System Technical Jour-nal , (1956), 917–926.[23] R. Khasawneh. Complex credit vehicles get downgrade to ‘default’. The Wall StreetJournal . New York, NY: October 15, 2008.[24] A. Lucchetti and K. Scannell. New debt products test Moody’s methods.
The WallStreet Journal (Eastern edition). New York, N.Y.: May 22, 2008, p. C.12.[25] S. Mahadevan, P. Naraparaju, and A. Sheets. CPDO – Friend or foe? In:
Struc-tured Credit Insights: Instruments, Valuation and Strategies (Chapter 84), thirdedition. Morgan Stanley, March 16, 2007.[26] A. P. Maitra and W. D. Sudderth.
Discrete Gambling and Stochastic Games .Springer, New York, 1996.[27] A. Morris and L. Schmidt. Avoiding a design flaw of early CPDOs: Why earlyCPDOs are not safe and how to improve on them. UBS Investment Research. July20, 2007.[28] S. Ross.
A First Course in Probability , seventh edition. Prentice Hall, New Jersey,2007.[29] C. M. Rump. Capital growth and the St. Petersburg game.
The American Statis-tician , (2007), 213–217.[30] S. Shreve. Don’t blame the quants. Forbes Magazine , October 8, 2008.[31] G. Sz´ekely and D. St. P. Richards. The St. Petersburg paradox and the crash ofhigh-tech stocks in 2000.
The American Statistician , (2004), 225–231.[32] G. Sz´ekely and D. St. P. Richards. Remain steadfast with the St. Petersburg para-dox to quantify irrational exuberance. The American Statistician , (2005), 235–239.[33] E. O. Thorp. The Kelly criterion in blackjack, sports betting, and the stock mar-ket. In: Handbook of Asset and Liability Management, Volume 1: Theory andMethodology (S. A. Zenios and W. Ziemba, eds.), pp. 385–428. North-Holland,Amsterdam, 2006. 4134] O. Toutain, M. Kheloufi-Trabaud, and B. Marjolin. CPDOs: A primer. Moody’sInvestors Service, November 8, 2006.[35] A. Tversky and D. Kahneman. Belief in the law of small numbers.
PsychologicalBulletin , (1971), 105–110 [reprinted in A Handbook for Data Analysis in theBehavioral Sciences, Methodological Issues (G. Keren and C. Lewis, eds.), pp.340–350].[36] W. A. Whitworth.
Choice and Chance , fifth edition. Deighton, Bell, & Co., Cam-bridge, 1901.[37] A. Wald. Sequential tests of statistical hypotheses.