Centre for Practical Quantitative Finance

CPQF Working Paper Series

No. 25 On the Calibration of the Cheyette Interest Rate Model

Ingo Beyna, Uwe Wystup

June 2010


Ingo Beyna
PhD Student CPQF Frankfurt School of Finance & Management Frankfurt/Main

Uwe Uwe Wystup
Professor of Quantitative Finance Frankfurt School of Finance & Management Frankfurt/Main


Frankfurt School of Finance & Management Phone: +49 (0) 69 154 008-0 Fax: +49 (0) 69 154 008-728 Sonnemannstr. 9-11 D-60314 Frankfurt/M. Germany

On the Calibration of the Cheyette Interest Rate Model
Ingo Beyna

ingo.beyna@web.de Frankfurt School of Finance & Management Centre for Practical Quantitative Finance Sonnemannstraÿe 9-11, 60314 Frankfurt, Germany
Uwe Wystup

uwe.wystup@mathnance.com Frankfurt School of Finance & Management Centre for Practical Quantitative Finance Sonnemannstraÿe 9-11, 60314 Frankfurt, Germany
June 28, 2010

We investigate the robustness of existing methods to calibrate the Cheyette interest rate model to at-the-money swaption, caps and oors. Existing algorithms may fail, because they suer from numerical instability of derivatives. Therefore, we apply derivative-free techniques and nd that they stabilize the calibration. Furthermore, we identify auspicious volatility parametrizations determining the Cheyette model. In combination with the established calibration techniques the results imply an accurate market reproduction and stay robust against changes in the initial values. In contrast to existing approaches that use approximations, we apply exact semi-close-form pricing formulas.


Keywords: Cheyette Model, Calibration, Optimization without derivatives, Genetic Optimization

We would like to thank Prof. Dr. P. Roÿbach (Frankfurt School of Finance & Management), Prof. Dr. H. Mittelmann (Arizona State University) and Prof. Dr. W. Schmidt (Frankfurt School of Finance & Management) for their support and advice on this paper.


Beyna and Wystup - On the Calibration of the Cheyette Interest Rate Model


1 2 3 Introduction Literature The 3

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4 5

Cheyette Interest Rate Models 3.1 Derivation of the Model . . . . . . . . . . . 3.1.1 The Heath-Jarrow-Morton framework 3.1.2 The Cheyette Model . . . . . . . . . 3.2 Volatility Parametrization . . . . . . . . . . 3.3 Pricing Formulas . . . . . . . . . . . . . . . Calibration Problem Formulation . . . . . . . . . . . . . . . . . Constraints . . . . . . . . . . . . . . . . . Characterization of the optimization space Quality Check . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10 13

5 7



4.1 4.2 4.3 4.4


16 18 18

5 Optimization Methods 5.1 Overview of methods . . . . . . . . 5.2 Assessment of methods . . . . . . . 5.2.1 Newton Algorithm . . . . . Description . . . . Results . . . . . . 5.2.2 Powell Algorithm . . . . . . Description . . . . Results . . . . . . 5.2.3 Downhill Simplex Algorithm Description . . . . Results . . . . . . 5.2.4 Simulated Annealing . . . . Description . . . . Results . . . . . . 5.2.5 Genetic Optimization . . . . Description . . . . Results . . . . . . 5.3 Conclusions . . . . . . . . . . . . .
A Appendix

23 24 24 24 24 24 24 25 26 26 27 30 30 31 34 34 36 37

41 42


Beyna and Wystup - On the Calibration of the Cheyette Interest Rate Model




In 1992, D. Heath, R. Jarrow and A. Morton (HJM) [Mor92] have standardized a valuation approach on the basis of mainly two assumptions: the rst one postulates, that it is not possible to gain riskless prot (No-arbitrage condition), and the second one assumes the completeness of the nancial market. The HJM model, or strictly speaking the HJM framework, is a general model environment and incorporates many previously developed models like the Vasicek model ( 1977) [Vas77] or the Hull-White model ( 1990) [Whi90]. The general setting mainly suers from two disadvantages: rst of all the diculty to apply the model in market practice and second, the extensive computational complexity caused by the high-dimensional stochastic process of the underlying. The rst disadvantage was improved by the development of the LIBOR Market Model ( 1997) introduced by [Mus97], [Jam97] and [San97], which combines the general risk-neutral yield curve model with market standards. The second disadvantage can be improved by restricting the general HJM model to a subset of models with a similar specication of the volatility structure. The resulting system of Stochastic Dierential Equations (SDE) describing the yield curve dynamic breaks down from a high-dimensional process into a low-dimensional structure of Markovian processes. Furthermore, the dependence on the current state of the process allows the valuation by a certain Partial Dierential Equation (PDE). This approach was developed by O. Cheyette in 1994 [Che94].

The framework proposed by Cheyette for modeling interest rates leaves the user with a wide choice of alternatives. The model is completely determined by the selection of (separable) volatility functions and given a parametrization, one might end up in known models like Ho-Lee (constant volatility) or Hull-White (exponential volatility structure). Each model bears dierent advantages and the selection of a model depends on its application. In order to use one model in practice it needs to be calibrated to liquidly traded interest rate derivatives. The aim of the calibration is minimizing the dierences in prices between the market quotes and the model implied prices. The minimization is performed with respect to the free parameters of the model in case of the Cheyette framework to the coecients of the volatility parametrization. Alternatively, the calibration can be implemented to minimize the dierences in implied (Black-Scholes) volatility. The comparison of the volatilities is more standardized, because it is independent of the notional and the maturity. Furthermore, the quotation in the market of prices and implied volatility coincides for plain products, thus the calibration problem stays unchanged.

Beyna and Wystup - On the Calibration of the Cheyette Interest Rate Model


The calibration of an interest rate model is one of the most important steps for pricing exotic products. Surprisingly, very little literature is known dealing with the applicability of optimization algorithms. The special structure of the Cheyette model inuences the performance of the minimization methods. The pricing formulas are given semi-explicitly [Hen03] and their derivations are not given in closed-form. One should thus apply derivativefree minimization algorithms like the Downhill Simplex algorithm or Genetic Optimization techniques. In our work we have applied several methods to the calibration problem and in this paper we discuss their behavior in detail. The results are quite dierent for the analyzed techniques and the choice of an adequate optimization algorithm plays an important role. As the calibration problem seeks a global minimum, we have incorporated some stochastic algorithms which do not guarantee the detection of the global minimum, but reach it with high probability. Furthermore, its application to the calibration of interest rate models is not well explored yet and we give a starting point in this paper.




The calibration of interest rate models to plain derivatives can be formulated theoretically quite easy as a minimization of the dierences in prices between the model and the market quotes. Taking the specic pricing formulas into account, the characteristics of the optimization problem appear. Brigo and Mercurio [Mer05] formulate the calibration problem for the general HJM framework in an abstract way and concentrate on swaptions and caps. Unfortunately, they neither discuss the practicability nor analyze the numerical tractability. Andersen and Andreasen [And02] identify several problems concerning the numerical tractability of the general calibration problem for the Gaussian HJM model. Especially the computation time of pricing vanilla caps and swaptions complicates the accurate calibration. They therefore suggest an approximation by asymptotic expansions and apply it to low-dimensional HJM models. They receive accurate results for small maturities, but for long maturities these worsen sometimes. A discussion of the used optimization algorithms or the applicability of known techniques is missing as well. A slightly dierent approach is presented by Andreasen [And05], who approximates the underlying stochastic dierential equation by a time-homogeneous model. Thus he receives (quasi) closed-form pricing formulas. This type of asymptotic solution just takes simple volatility structures into account and does not deliver convincing results. Again Andreasen

3.T ] are m-dimensional progressively measurable stochastic processes satisfying certain conditions on the regularity: . in contrast to short rate models or market models. F. The uncertainty in the economy is characterized by the probability space (Ω.Beyna and Wystup . 3 3. In the second step we will limit the class of models to a specic process structure of the yield curve dynamic. All the presented approaches incorporate an approximation of the pricing formula to simplify the minimization problem. In this paper we use the exact pricing formulas developed by [Hen03] and analyze several optimization techniques to receive accurate results in reasonable computation time.T ] and the volatility of the forward rate σ = (σ(t. The algorithm is really fast.1. T )dt + σ(t. and Q is a probability measure. we do not simply adjust the problem. T ))t∈[t0 . The dynamic of the forward rate is given by an Itô process df (t. where Ω is the state space. F is the σ -algebra representing measurable events. which makes the bootstrapping of the volatility possible. T ) = µ(t. T )dW (t). so that we will present the general setup rst. Thus. The method is based on the pricing PDE and uses Finite Dierences techniques to compute the prices. T ))t∈[t0 . but apply dierent optimization techniques to solve it.1 The Cheyette Interest Rate Models Derivation of the Model The Cheyette interest rate model is a specialization of the general HJM framework. This technique will guide us to the representation of the class of Cheyette models. but it can only be applied to really easy volatility parametrizations in low-dimensional HJM models so far. Q). (1) It is assumed that W(t) is an m-dimensional (Q-)Brownian motion and the drift µ = (µ(t.On the Calibration of the Cheyette Interest Rate Model 5 [And00] suggests another approach to simplify the calibration especially to swaptions. He approximates the dynamic of the swap rate. It can be classied as a forward-rate model.1 The Heath-Jarrow-Morton framework The Heath-Jarrow-Morton approach yields a general framework for evaluating interest-rate derivatives.

u)|ds du < ∞ Q − a. The implied forward rate under the drift restriction is given by t f (t. This condition can. T ) = σ(t. T )dW (s). t0 T t0 T∗ t0 T∗ t0 2 σj (s.Beyna and Wystup . T )   s T σ(s. j = 1. T ) = f (0. If and only if the drift of the forward rate has the structure µ(t. . T ∗ ] and t ∈ [0.s.s. ∀t0 ≤ T ≤ T ∗ ... T ) + t0 σ(s.1 (Forward rate drift restriction) . be fullled by assuming a restriction on the forward rate drift µ(t. for instance. (2) (3) (4) (5) |f (t0 . T )|ds < ∞ Q − a.. Then the forward rate is given by t t |µ(s. T )ds + t0 σ(s. T ]. v) dv − λ(t)   ∀T ∈ [0.. T ) = f (0. T )   t T whereas λ(t) notes the measure exists . ∀t0 ≤ T ≤ T ∗ . f (t. (6) The construction of the forward rate structure is based on the assumption of a complete market. t0 . T )ds < ∞ Q − a. s)|ds < ∞ Q − a. v)dv − λ(s)  ds +  t σ(s. T ) + t0 µ(s.s. Condition 3.s.. m.   u  t0 The notations use the xed maturities T ≥ t as well as the maximum time horizon T ∗ ≥ T . Market Price of Risk. T ).On the Calibration of the Cheyette Interest Rate Model 6 T |µ(s. The economic concept of a complete market can be translated to the existence of a unique martingale measure. T )dW (s). then an equivalent martingale σ(t.

so f (t. T ) and achieves an exogenous model of the yield curve with Markovian dynamics. then the forward rate and the short rate are normally distributed. general HJM approach as a model having essential theoretical advantages and follow [KL07]. so that the HJM framework contains a really wide class of models.2. T ) is a deterministic function. T ) . σ(s. We will follow the ansatz of O. we will neglect the market price of risk and set λ(t) ≡ 0. One straightforward approach is given by setting σ(t. one can state that the term structure of volatility determines the forward rate at all times. T )dW (s).1. As already suggested in the literature.On the Calibration of the Cheyette Interest Rate Model that the forward rate reduces to t 7 In the following. T ) = const · f (t.2 The Cheyette Model The class of Cheyette interest rate models forms a subset of the general class of HJM models.3. we would assume the forward rate to be lognormally distributed. (ii) By specifying the volatility structure in this general class of yield curve models. we sum up the 3. T ) = f (0. 3.Beyna and Wystup . (7) t0 The derivation of the forward rate dynamics is only based on the forward rate drift restriction. T )   s T σ(s. Remark Remark 3. As a consequence. Cheyette . (iii) To make the general HJM model consistent with market practice. T ) + t0 If the volatility σ(t. (i) The yield curve dynamics are completely determined once the structure of (forward rate) volatility is specied and the initial market yield curve data is included. v)dv  ds +  t σ(s. In order to close the derivation of the general HJM setup. which is dominated by Black's formula and assumption on the forward rate dynamics. one recovers short-rate models or the Libor Market Model. one can choose a specic volatility structure σ(t.

As already mentioned. a humped shape can not be generated. Empirical research [Che94] has shown that already two summands deliver suciently accurate results. (ii) The dynamic of the forward rate will be determined by the short rate and the cumulative quadratic variation. the volatility and the number of summands N in equation (9) controls the accuracy. The dynamic of the forward rate can be reformulated as follows. we can extend the volatility term structure to a nite sum of separable functions and allow N N σ(t. (10) . T ) = i=1 αi (T ) βi (t) .g. The models implied by (8) do not cover all possible forward rate structures. Formally. (8) α(t) Following this idea. The volatility function is assumed to be separable into time and maturity dependent factors given by the structure β(t) σ(t. The theorem of Stone-Weierstrass states that the approximation is uniform. As a consequence of the reduction to the specic volatility term structure (9) the space becomes nite. it is an approximation of a continuous function. the class of Cheyette Models forms a subset of the class of HJM Models. T ) + j=1 αj (T ) αj (t) N xj (t) + i=1 Ai (T ) − Ai (t) Vij (t) αi (t) . Even this abstract denition implies the following changes: (i) In the general HJM framework the state space of volatility functions is not nite. αi (t) (9) The choice of the volatility structure aects the characteristics of the covered models. Consequently. f (t. where the risk-neutral HJM dynamics can be represented through a system of Markovian SDEs. In order to increase the complexity of the class.On the Calibration of the Cheyette Interest Rate Model 8 [Che94] and use a separable volatility term structure.4. we can interpret the class of the Cheyette models as an approximation to the set of HJM Models. one arrives at a subclass of yield-curve models. T ) = i=1 σ (i) (t. The extent of the set of Cheyette models is determined by the set of supported volatility functions. e.Beyna and Wystup . T ) = f (0. if we assume the mentioned volatility structure: N Remark 3. T ) = α(T ) .

. Due to this improvement. N . 2 The dynamics of the short rate and the quadratic variation are given by Markov processes as N dxi (t) = xi (t)∂t (log αi (t)) + k=1 Vik (t) dt + βi (t)dW (t) (14) (15) d Vij (t) = βi (t)βj (t) + Vij (t)∂t (log(αi (t)αj (t))). αi (s)αj (s) The dynamic of the forward rate is determined by the state variables xi (t) and Vij (t) for i. it can be transformed to a PDE. . N : t Ak (t) = 0 t αk (s)ds. The rst variable can be interpreted as a yield curve factor and the second as a convexity adjustment term ensuring that the model is arbitrage-free. j = 1. Several well known numerical methods like Finite Elements or Finite Dierences to solve it. the simulation of the forward rate becomes more ecient. dt The Cheyette Model and all features were achieved just by assuming a special structure for the volatility and on the basis of the assumptions of the general HJM framework.. the forward rate is determined by N (N + 3) state variables.Beyna and Wystup .. αi (t) βi (s) dW (s) + αi (s) t (11) xi (t) = 0 0 αi (t)βi (s) αi (s) t N k=1 Ak (t) − Ak (s) βk (s) ds. can thus be applied ... . j = 1. The representation of the forward rate (10) delivers several advantages concerning the numerical tractability: (i) The dierential equations describing the dynamic of the state variables (14) and (15) can be solved independently. The stochastic variable xi describes the short rate and the non-stochastic variable Vij states the cumulative quadratic variation.. Summarizing. (ii) The SDE ( 14) is a Markov Process and by using the Feynman-Kac Theorem.On the Calibration of the Cheyette Interest Rate Model 9 The presentation uses the following notations for i. αk (s) (12) (13) Vij (t) = Vji (t) = 0 αi (t)αj (t) βi (s)βj (s)ds.

functions σ (i) (t. T ) = i=1 σ (t. which are obviously separable functions. N In the Cheyette Model. 3. implying The implied model is called exponential Cheyette Model . Thus the parametrization β(t) and The most intuitive and easiest choice is assuming reduces to σ (i) (t. the volatility function is assumed to be the nite sum N σ(t.Beyna and Wystup . but it f (t. an exponential structure of αi (t) is promoted. e. T ) = i=1 (i) αi (T ) βi (t). T ) = c ≡ const. the Cheyette Model is dened as a one factor model. The additional factors are given by several independent Brownian motions and the forward rate is given by M At this stage.2 Volatility Parametrization of separable functions given by (9). T ) = f (0. This representation can be achieved by choosing αi (t) = 1 βi (t) = c.  (18) This suggested structure again includes two degrees of freedom. One summand should form a constant shift of the structure and consequently we choose σ (1) (t. namely the deterministic κ(u) to be constant. T ) = βi (t) exp − κ(u). T ) denotes a one factor forward rate dened by (10). The value of the constant does not aect it (17) just adds a constant shift. T ) = βi (t) exp (−(T − t)κi ) . The shape of the volatility structure is determined by the other summands taken into account. T ) + i=1 ˜ f i (t.On the Calibration of the Cheyette Interest Rate Model 10 can easily be generalized to a multi-factor model. (16) ˜ where f i (t. αi (t) Empirical results [Che94] propose a volatility function with two or three summands to achieve a reasonable structure including a permanent humped shape.  T t κi (u)du . (19) . and the shape of the volatility function. T ). In the academic literature.g [Che94].

β(t) to be a linear or a quadratic (20) (21) β(t) = at + b or β(t) = at2 + bt + c. 4 . Especially the shapes in direction of the remaining lifetime T are more exible. and 10 years. we we implement. nodes for 1. Assuming κ to be constant implies no exibility of the volatility structure with respect to the remaining lifetime. a piecewise constant function κ(t) (red surface). The results are improved by assuming function of time. only β(t) remains as a time-dependant parameter. The plot demonstrates that the second volatility structure (κ(t) piecewise constant) supports more exibility and movements of the parametrization. on the other hand. on the one hand. but equal in both structures.e. The rst structure ( κ(t) = const. i. values of the parameters are determined by the calibration of the model in section install some nodes dividing the domain of κ(u). 5. on the other hand it increases the number of degrees of freedom for the volatility parametrization. This motivates us to allow time horizon. β(t) cannot control the volatility with respect to the remaining lifetime (T − t). Therefore. As it is a function of time and not of maturity. 7. The calibration is equivalent to a (global) minimization and the In addition to the structure of κ(u) one can inuence the range and exibility of the volatility structure by determining β(t). Instead of one parameter ( κ constant). Unfortunately this leads to a time-independent scaling factor and from our experience with various tests the range of the resulting volatility function is not wide enough for most of the applications. 10] representing the current time t and the remaining lifetime T . 3. Figure 1 and Figure 2 show possible volatility surfaces in the domain [0. The κ(u) is assumed to be constant within each time interval. we have to determine six values ( κ piecewise constant with ve nodes). for example. The simplest choice is assuming β(t) to be constant.Beyna and Wystup . The coecients of the linear function (20) are chosen arbitrarily. On the one side this extension improves the lack of exibility with respect to the remaining lifetime. disadvantage in uctuating markets. Figure 1 compares volatility surfaces for a linear function β(t) = at + b combined with.) just supports constant changes with respect to the remaining lifetime and thus it appears . This factor can be addressed by α(t). 10] × [0. Depending on the simulated boost of free parameters increases the dimension of the optimization problem as well. This might be a κ(u) to be piecewise constant. a constant κ(t) = κ (blue surface) and.On the Calibration of the Cheyette Interest Rate Model 11 Assuming this structure.

the more complex model takes the quotient of remaining lifetime and current time linearly into account: T (3) σT (t.Beyna and Wystup .07. rather rigid. κ4 = 0.1. Figure 2 shows possible structures and to emphasize the dierences we x the time to maturity T = 5. κ3 = −0. The parametrization with two summands consists of a linear function β(t) and a piecewise constant κ(t) in the exponential part.05. The constant shift in both structures is given by c = 0.3.2 and the constant κ is chosen as κ = 0. The market often implies large movements in this section in comparison to volatilities for longer lifetimes. One crucial point within the volatility function is the behavior for short lifetimes. As demonstrated. t 2 compares volatility structures taking two and three summands of This implies more exibility especially for small t and bears some advantages in extreme markets. the seFigure cond model is superior concerning this aspect.1. separable functions into account. κ5 = 0.06. κ2 = −0. b = 0.06. The coecients of β(t) take the values a = 0. The coecients of the piecewise constant κ are given by κ1 = 0.On the Calibration of the Cheyette Interest Rate Model 12 Figure 1: Representation of possible volatility surfaces of the exponential Cheyette model with β(t) as a linear function (20) and a constant κ (blue) and a piecewise constant κ (red). Thus the plot depicts a slice . T ) = c . Additionally.0.03.

A possible volatility surface based on this structure is exemplarily shown in Figure 3.0. β(t) = a ln(t)2 + b ln(t) + c. The plot shows a slice plane of the surface at T = 5. we tested several dierent choices of βi (t) and one interesting alternative takes time logarithmically into account. The coecients are chosen as in Figure 1 and the additional scaling factor of the quotient for the remaining lifetime equals c = 0.Beyna and Wystup .0003. 3. plane of the volatility surface. this results. i.3 Pricing Formulas The interest rate models are based on several assumptions.On the Calibration of the Cheyette Interest Rate Model 13 Figure 2: Representation of possible volatility structures consisting of two and three summands: constant shift.e. exponential form with linear β and piecewise constant κ and quotient for remaining lifetime. structure especially bears several advantages and leads to signicantly (22) better If derivatives with extremely varying lifetime are used simultaneously. Furthermore. hence each model implies dierent formulas for pricing plain derivatives and exotic products. .

Henrard [Hen03] gives semi-explicit formulas for pricing options on bonds and swaptions in the general one factor HJM framework. The model structure and the coecients are the same as in Figure 1 .Beyna and Wystup . we derive pricing formulas for these type of interest rate models. Theorem 3. The calibration of the models incorporates the pricing formulas for at-the-money caps. ti )N (λ + αi ) − P (0. PV = i=1 ci P (0. (23) λ is the unique solution to .On the Calibration of the Cheyette Interest Rate Model 14 Figure 3: Possible volatility surface with two summands and logarithmical time scaling.5 The present value (PV) for (receiver) at-the-money swaptions in the Cheyette model is given by the semi-implicit formula n (Swaption pricing formula) . oors and European swaptions. HJM and especially the Cheyette models are based on the same assumptions and dier only in the volatility function. t0 )N (λ + α0 ). These formulas are based on The models in the class of a general structure of the volatility and if we choose the volatility according to the assumptions in the Cheyette Model (9).

Henrard [Hen03] introduces explicit forThe pricing formula for caps and mulas for options on bonds. Similar to the pricing of swaptions.On the Calibration of the Cheyette Interest Rate Model 15 n ci P (0. these values and the resulting present value depend on φ ∈ Rd as well..) standard normal cumulative distribution function with mean 0 and standard deviation 1 • θ < t0 • • σ(s. t1 . . cn = 1 + strike discount factor at time 0 with maturity ti • P (0. Furthermore we use the notations: • N (.Beyna and Wystup . tn coupon payment ci at date ti : ci = strike ∀i < n. x)dx ds 2 maturity of the option. one can divide a cap/oor into a portfolio of European put/call options on zero coupon bonds with adjusted payo and strike. x)dx − s ti = 0  θ θ whereby σ(t.. The realisation of volatility functions depends on the choice of the degrees of freedom in the parametrization summarized in the vector φ ∈ Rd . T ) denotes the volatility at time t with maturity T . As the values of αi and κ are determined by the volatility. Applying this decomposition one can reduce the pricing of caps/oors to the pricing of several options on zero bonds. ti )e− 2 αi −αi λ = P (0. ti ) oors can be created by dividing the payo structure into smaller units. x)dx ds 2 σ(s.. the dierence is about 2 business days and thus we approximate θ ≈ t0 payment dates (in years) in the swap t0 ≈ θ. Following Hull [Hul05]. . t0 )e− 2 α0 −α0 λ i=1 1 2 1 2 (24) and αi is dened by θ 2 αi = 0    s ti θ σ(s.

. t0 )N (−λ − α0 ) − σT (s.Beyna and Wystup .1 The Calibration Problem Formulation Dealing with term structure models.6 (Pricing Let θ ≤ t0 ≤ t1 ≤ . ≤ tn . The aimed exibility can be achieved by . The existence and the uniqueness of λ for the general case is shown in [Hen03]. ci P (0. t0 )N (λ + α0 ) i=1 where λ is the (unique) solution of n i=1 1 2 ci P (0. x)dx ds. the possibility to represent the current state of the This feature can be obtained by creating nancial market is one of the most important characteristics. n 2 ci P (0. 4 4. ln α1 c1 P (0. ti )N (λ + αi ) − KP (0. i=1 In case of a zero coupon bond the presented formula reduces in case of a zero coupon bond to a single payment and thus λ is given explicitly by λ= −1 KP (0. ti ) exp(− αi − αi λ) = KP (0.. t1 ) 2 Based on these pricing formulas we can formulate the minimization problem for the calibration. t0 ) 2 and the αi 's are the positive numbers such that θ 2 αi = 0   θ ti The price of a European put is given by KP (0.On the Calibration of the Cheyette Interest Rate Model 16 Theorem 3. exible models that could generate any shape of the term structure. Consider a bond which pays ci at times ti (1 ≤ i ≤ n). t0 ) 1 − α1 . ti )N (−λ − αi ). At time 0. the price of a European call on the bond with expiry θ and strike price K to be paid in t0 is n Formula for Options on Bonds) .

Each factor represents a predened shape. The quality of the approximation of the market can be measured by comparing characteristic numbers of standardized nancial instruments quoted in this market fraction. The formulation of the calibration problem is based on the degrees of freedom within the specic structure.g. As a consequence the major processes have the Markov property that forms an advantage for the subsequent numerical simulation. because the level of values is standardized as opposed to the prices. we can therefore compare the dierences in the implied Black-Scholes volatilities. e. The comThe calibration of the parison of the implied volatility should be preferred. Changing the point of view slightly. the calibration problem is given by Problem 4. Instead of comparing the price dierences as presented. The prices of the derivatives as quoted in the market (based on the Black-Scholes model) are compared to the prices computed in the model depending on coecients φ ∈ Rd . so the complexity of the model appears proportional to this number. The number of free model parameters depends on the number of included factors. They can thus be inverted locally with respect to the volatility. we calibrate a model to a specic market section. parallel shift or twist. we can look for a term structure model minimizing these differences for a specic market section. A comparison is thus more meaningful. The market values of the derivatives are assumed to be computed in the Black-Scholes model and the corresponding pricing formulas for swaptions and caps/oors are one-to-one. prices or volatility. for swaptions) depending on the degrees of freedom. Furthermore.g.Beyna and Wystup . As a consequence the models become more complex and the mathematical tractability decreases.g.1.e. i. let σi denotes the implied Black˜ .On the Calibration of the Cheyette Interest Rate Model 17 term structure models taking several (stochastic) factors into account. Cheyette model is performed on the basis of several at-the-money European swaptions (receiver) with dierent maturities up to ten years and caps/oors (at the money). the model should rebuild a reasonable fraction of the nancial market at best and at the same time stay as simple as possible. Following this ansatz. The set of free parameters in the model is limited to the coecients representing the parametrization of the volatility. Cheyette models have a special volatility term structure that is parameterized by a sum of separable functions. e. Assume Φ ⊂ Rd to be the set of free parameters in the model and further P V : Φ ⊂ Rd → R to be a pricing formula (e. A common method to measure the accuracy of the approximation is based on the dierences in pricing between the observed market prices and the model implied prices. In sum.

T ) ≥ 0 ∀t > 0. E(φ) = i=0 ωi |σ(P Vi (φ)) − σi |2 . I φ∈Φ inf E(φ). One even has to perform a one-dimensional Newton optimization to compute these values. 4.On the Calibration of the Cheyette Interest Rate Model 18 Scholes volatility as observed in the market and σi (P Vi (φ)) the implied BlackScholes volatility representing the price P Vi (φ) in the Cheyette model with the parameter set φ ∈ Φ.Beyna and Wystup . like a constant volatility. Furthermore.2 Constraints rameters. T ≥ t Instead. the function does not hold .3 Characterization of the optimization space space is determined by the realizations of the function E : Φ ⊂ R → R. the d The optimization function is generally neither convex nor concave. for example swaptions and caps might be weighted dierently. Obviously E is non-linear and the formula for the (Black-Scholes) implied volatility based on the present values in the Cheyette Model is not given explicitly. the corresponding minimization problem appears mostly high dimensional. Often it contains more than eight free coecients. Only in very simple cases. the function is monotone. In addition to the non-linearity. the function appears to be monotone in this variable. To sum up. which complicates the optimization problem dramatically and many methods are no longer applicable. In the calibration process we are looking for the global minimum of the function E : Rd → R. because it is linked to the choice of volatility parametrizations. The optimization does not include any explicit constraints on the single paσ(t. 4. The search for the global minimum strongly depends on the characteristics of the function E . the reasonable constraint (26) should be fullled by the interaction of all coecients in the parametrization of the volatility. ˜ (25) The integer I denotes the number of all derivatives taken into account for the calibration and ωi ∈ R denotes weights that control the inuence of any instrument. The intractability of the function is even intensied by the fact that the function is not monotone for all coecients. The combination of variables destroys this feature. Only if we x all coecients but one.

Thus there are ve free parameters. 0. we will show some examples of possible spaces for a implied volatility between the market values and the model implied values.Beyna and Wystup . Figure 4 and Figure 5 show the dierences in implied (Black-Scholes) volatility for swaptions between the observed market values and the model implied values for a given parameter set κ(u)du . the computation of a single function value mization takes about methods are based on (partial) derivatives. Several opti- In the following. T ) = c + (at + b) exp −  T As the space represents the dierences in the t The deterministic function κ(u) is assumed to be piecewise constant and the domain is divided into two disjoint intervals. This structure is rather crude.1 sec1 on average and is rather time consuming. The sample takes 10 swap- In order to demonstrate the structure of the error surface. if one of the coecients is far from the minimum. the volatility structure is determined by a. used a Windows based PC with Intel Core 2 Duo CPU @ 1. 4. cients in each plot and present the resulting error surface. 5 and 6 years. which are not applicable in this case. The dierence quotient often vanishes. As the space in the example is ve-dimensional. As the coecient c ∈ R just adds a constant shift. 3. Several plateaus with the same function value appear and hence the dierence quotient must vanish. but it shows most of the characteristics of the minimization surface. especially the non-explicit formula for the implied volatility. given state of the market. The use of dierence quotients does not produce stable results either. The Cheyette Model used is determined by the volatility parametrization σ(t. κ1 and κ2 . The presentation is based on the market data observed at the end of June 2009 and it just takes (at-the-money) swaptions into account. because small changes might not inuence the value of the function. Due to the complex structure of the function.On the Calibration of the Cheyette Interest Rate Model 19 many helpful features to simplify the optimization problem.  (27) tions into account and the error is given as the sum of the squared dierences. b.66 GHz and 3. we vary two coefThe remaining 1 We RAM. φ ∈ Φ.25 GB . we have to use an appropriate projection for the graphical presentation. We take 10 swaptions with option lifetime of 3 months and 1 year each combined with swap lifetimes of 2. the space depends on the used state of the market and changes dynamically.

On the Calibration of the Cheyette Interest Rate Model 20 coecients are held constant in each graph. The surface contains several holes.1] for possible values is suggested by empirical tests. 0. Each hole in the surface implies .1. due to the range of possible values of the z -axis. Figure surface with respect to the variables a and b in the calibration process of the model implied by equation (27) 4: Minimization Figure 5: Zoom into the minimization surface given in Figure 4 Figure 4 and Figure 5 display the movement of the error surface with respect to the coecients a and b.Beyna and Wystup . The interval [−0.

This is caused by the numerical computation of the implied (Black-Scholes) volatility. a hole will be created in the used plot. the error values increase rapidly. because a calculated Hence. As the range of the z -axis is . Consequently.On the Calibration of the Cheyette Interest Rate Model 21 Figure 6: Minimization surface with respect to the variables κ1 and κ2 in the calibration process of the model implied by equation (27) Figure 7: Zoom into the minimization surface given in Figure 6 an error value exceeding the maximal range. there are several jumps and the surface is not continuous. At the border and in the middle of the analyzed interval for the coecients. value exceeding the limit of 100% volatility is forced to take a high technical In Figure 4 there are two separate regions possibly containing the (global) minimum. value.Beyna and Wystup . if one value implied by the model exceeds a predened level.

4 Quality Check order to apply a standardized technique. The surface appears rather regular and contains no But the pictured surfaces only include two degrees of freedom for illustration. that have to be treated in the optimization process. Several methods that can handle these functions exist. but none of them guarantees the convergence to a global optimum. Each of the presented surfaces only depends on two parameters. Figure 6 and Figure 7 display the movement of the error surface with respect to the coecients κ1 and κ2 of the piecewise constant function κ(u). The whole surface is given as a combination of these structures. The core conditions are given by: • Price Bias ≤ 30%: At most 30% of the relative price dierences exceed the limit of 30% • Volatility Bias ≤ 30%: At most 30% of the relative volatility dierences exceed the limit of 30% • < σ: The mean of all price residuals does not exceed the standard deviation σ with mean 0 Mean price residual . we zoomed in and displayed the same surface in Figure 5 for a smaller range of values. we come up with several conditions to the prices and the implied Black-Scholes volatility to evaluate the calibration. We focus on the price and volatility residuals between the Cheyette and the Black-Scholes model. This gure demonstrates that there are valleys in the surface and that the error reacts sensitively to small changes in these coecients. In general. The remaining coecients are xed and chosen to minimize the error for the given data set. The implemented check of the calibration consists of twelve conditions subdivided into four core and eight secondary ones.Beyna and Wystup . In The design and the denition of the quality check is comparable to the criteria used in the pricing software `FINCAD Analytics Suite 2009'. Thus the resulthills. The evaluation of the calibration results leaves much freedom to the user. there are several local minima.On the Calibration of the Cheyette Interest Rate Model 22 rather big. The structure of the surface suggests that the choice of κ(u) does not ing surface becomes more turbulent and irregular. Consequently. 4 . change the values dramatically. emerge. including (local) minima. jumps and hills. κ(u) takes at least ve coecients into account and thus several valleys.

g. If all conditions . e. Newton algo- rithm with and without step size adjustment. Simplex and Powell algorithm and Downhill 3.core and secondary . If at most one secondary condition does not hold. dierent classes: Developing the calibration method. we have implemented several minimizaMainly. we can classify the tested algorithms into three 1.1 Optimization Methods Overview of methods tion algorithms. 5 5. netic Optimization. passed and failed.On the Calibration of the Cheyette Interest Rate Model • Mean volatility residual 23 The mean of all volatility residuals does not exceed the standard deviation σ with mean 0 < σ: Furthermore we identied several secondary conditions given by: • Convergence of the minimization algorithm • Minimum does not lie on the boundary • All volatility residuals are bounded by 3σ • All price residuals are bounded by 3σ • 90% of the volatility residuals are less than 2σ • 90% of the price residuals are less than 2σ • • Mean of squared volatility residuals is limited by 1 Maximal volatility residual does not exceed 5% The quality check delivers three possible outcomes: good.are fullled. The calibration algorithm fails. e. if one or more of the core conditions or more than one of the additional conditions are violated.Beyna and Wystup . non-linear optimization methods with derivatives. non-linear optimization methods without derivatives. e.g. stochastic optimization methods. the calibration is assumed to be `good'. 2.g. the calibration is interpreted to be still valid and thus the algorithm will deliver `passed'. Simulated Annealing and Ge- .

The resulting Jacobi. Due to the vanishing partial derivatives.2. presented in [Bre73].1 Powell Algorithm Description The Powell algorithm is a powerful tool for minimizing non-linear functions and is.1. The derivatives cannot be computed explicitly. among others.1 5. The derivatives in direction of the variables κ (in case of the exponential model) vanish especially often. This method makes use of the rst and second derivatives with respect to all directions. because changes in one variable do not have a huge eect on the price. the iterations steps are computed. Convergence cannot be achieved by using a standard step size adjustment.2. so that one has to use dierence quotients.2. Consequently.2 Assessment of methods In the following we present the analysis of the optimization methods consisting of method descriptions followed by the results of the application to the calibration problem. The parametrization of the volatility implies that the partial derivatives vanish sometimes.2.On the Calibration of the Cheyette Interest Rate Model 24 5.2 5. The advantage of the Powell algorithm is based on the fact that no derivatives are necessary.Beyna and Wystup . Based on these matrices. In order to reach an optimum in the . the Newton algorithm does not converge. the Jacobi.2.1.and Hessian matrix are assumed to have full rank and furthermore the Hessian matrix has to be invertible. 5.2.and Hessian matrix including the rst and second partial derivatives of the function to be minimized.1 Newton Algorithm Description The Newton algorithm is a local optimization method developed by Isaac Newton in 1669.2 Results The Newton algorithm as presented in [Sto03] is based on the Jacobi. 5. Even if we vary the initial value. 5. The algorithm will terminate in a position where the rst derivatives vanish and thus the algorithm in general detects only local optima. The multidimensional (n dimensions) minimization problem can be divided into several (n + 1) one-dimensional problems.and Hessian matrix have lower ranks and the Hessian matrix is no longer invertible. we cannot generate a stable algorithm.

because the function appears monotone. the algorithm seems to improve and produces better results. In contrast to the calibration taking the whole swaption matrix into account. the reduction to a fraction of the matrix increases the accuracy of the solution. Unfortunately. We tested the algorithm for the calibration to the whole swaption matrix and for the calibration to a fraction of the matrix. we need to nd the global optimum in each of the onedimensional optimization problems. the dierence in the implied Black-Scholes volatility is less than 0.5% on average. We reinitialize the basis of directions to search after processing n iterations with the purpose to improve the Powell algorithm and decrease the linear dependency. If we reduce the dimension of the problem by decreasing the number of coecients.On the Calibration of the Cheyette Interest Rate Model 25 superior algorithm. First. new directions for the optimization are created by combining dierent one-dimensional optima. The corresponding minimization space does not stay monotone any longer.2. The most important diculty of the minimization is the dierent sensitiv- .2 Results The performance of the Powell algorithm depends strongly on the choice of the initial values. In this case. The calibration to the whole swaption matrix reaches the global optimum. we receive suciently accurate results. but if the algorithm starts far away of the global optimum. 5. We have tested the algorithm with a known global optimum and the algorithm reached it quickly. This means. Based on these results.Beyna and Wystup . but several waves will appear complicating the minimization.2. the monotonicity is lost and the one-dimensional problems becomes more complicated. This method costs a lot of time. but we need to nd a global minimum of the one-dimensional problem in order to reach the global minimum with the superior algorithm. The new directions eect more than one coecient of the volatility function. the algorithm is sensitive to the initial values. the convergence is not guaranteed. so that the algorithm does not nd the global optimum if we start somewhere in the solution space. if we start close to it. If we calibrate the model to a set of swaptions with the same xed option lifetime and variable swap lifetime up to 10 years. As a result. the one-dimensional minimization is performed separately along a line in direction of each volatility coecient. the one-dimensional problem is easy. Due to this fact. But this behavior cannot be observed in each case. we have to use a brute force algorithm testing more or less all possible values in a predened interval. This kind of calibration is more robust concerning the initial values.

Based on function valuations. The second termination criteria is only feasible.Beyna and Wystup .On the Calibration of the Cheyette Interest Rate Model 26 ity of the coecients to the minimization function. These steps are called reections . the algorithm expands the simplex automatically to take larger steps. In combination one single step in the Powell algorithm is time consuming and thus this algorithm is not ecient for the calibration of the Cheyette Model. In case of a `valley' the method contracts itself in the transverse direction and tries to escape the valley. but it is robust and the results are reliable. the algorithm takes a series of steps. the eect of the coecients of the piecewise constant function κ(t) is relatively small. the constant and the coecients of the linear function (in the case of the example presented in (27)) have a strong eect on the minimization function.1 Downhill Simplex Algorithm Description The Downhill Simplex method is an optimization method requiring only function valuations and no derivatives. the algorithm stops if a cycle is identied and second. As the global optimum (along the line) is required only naive algorithms are available taking lots of time. There are many possible conditions and we selected two: First. if the function value is suciently close to zero. Most of the steps move the point of the simplex with and are constructed to conserve the volume of the simplex. The setup of the model carries undesired features destroying the advantages of the Powell algorithm for high dimensions. It might happen that one of the above criteria is violated by a single the highest value to the opposite face of the simplex to a lower point. 5. The simplex method takes an n-dimensional simplex consisting of (n + 1) points as initial value. The method continues as long as one of the terminal conditions is fullled. While. It is not very ecient in terms of the number of function valuations. If it is possible to perform the reections.2. Thus one has to use dierent discretizations or step sizes for the coecients to receive suciently accurate results and minimize iterations. The idea of the Powell algorithm includes the minimization along one-dimensional lines and the directions are composed by several coecients. This is even intensied by the characteristics of the one-dimensional minimization problem possessing several waves. The discretization and the step size must thus be adapted to all coecients which implies a large number of iterations and consequently large computation times.3 5.3. for example. The method is due to Nelder and Mead and was rst published in 1965 [Mea65].2. because we know that all values are positive and zero would imply a perfect t.

8 β 0. Numerous publications mention only three parameters for the DownhillSimplex algorithm and in this case they do not distinguish between parameter β and σ. we applied the calibration to dierent market data. It is therefore a good idea to restart the minimization routine at a point where it claims to have found a minimum.2 Results We applied the Downhill Simplex algorithm to the calibration problem taking ten (at-the-money) swaptions into account. The choice of these values inuences the speed of the method and might change the computed optimum.2 0. α 1.2.8 0.6 γ σ 1. Based on this data the three parameter sets deliver the best (set 1). July 09 and Jan.On the Calibration of the Cheyette Interest Rate Model 27 abnormality in the function. controls the expansion.2 0.Beyna and Wystup . 08.9 0.3. The plotted values are the minimal function values according to equation (25). 5.5 2. Figure 8 shows the results for the time series and the dierent parameter sets. The downhill simplex algorithm includes four parameters that control the movement of the simplex.5 Set Set Set 1 2 3 Table 1: Sets of analyzed parameters in the Downhill Simplex algorithm In addition to changes in the parameters of the algorithm. We have used data from the end of Jan. Jan. July 08. controls the contraction. the worst (set 3) and average (set 2) results as shown in Table 1 .6 0. .0 1. • α • β • γ • σ controls the reection. we tested three dierent parameter sets and analyzed the inuence on the solutions. controls the compression. The parameter sets were chosen on the basis of a standardized calibration test with market data of 2006.7 0.3 1. Furthermore. 09. 10.

hold the best standards in all cases. But even the calibration results in 2009 fulll the highest standards of the quality check. parameter set 1 is obviously dominant. one should prefer this calibration algorithm and the assumed model can build up these movements. The time series demonstrates that the calibration results are good in 2008. July 09 and the corresponding quality of the calibration results. in Jan.4 · 10−4 08 and 6. Starting in 2009. Based on the outcomes. . The The computed minima of the time series vary between 2. the range enlarges in 2009. one can state that parameter set 1 is superior to the other ones. but the values are quite similar. according to section 4. Next to this. While the performance for the three sets only dier slightly in 2008. the changes of the market data during the crisis inuence the behavior of the calibration with respect to the used parameter set. In 2008. when the results worsen. the level of swaption volatilities increases dramatically due to the nancial crisis.On the Calibration of the Cheyette Interest Rate Model 28 Figure 8: Time series of minimal values of the calibration to swaptions based on the Downhill Simplex algorithm and multiple parameter sets. After that.Beyna and Wystup . The but nevertheless the quality of the market reproduction decreases. the minimal function values worsen and imply that the proposed model does not t the market data as well as in 2008. During that time. Consequently.6 · 10 −3 in uncertainty in the market changed the structure of the volatility matrix.4 . set 1 does not deliver the best results.

9 shows that . The dierences between the coecients of two minima can be measured by the vector norm d |x (1) −x |= i=1 (2) |xi − x(2)i |2 (1) (28) and the dierences between two minimal values by |y (1) − y (2) |2 . The presented eects are all based on the same initial values for the minimization algorithm. 9: Histogram of changes in the minimal values for dierent initial values in the Downhill Simplex algorithm.On the Calibration of the Cheyette Interest Rate Model 29 setting. Additionally. We tested 136 dierent initial values for the calibration and monitored the changes in the minimal values and in the coecients of the minimum. although it does not always produce the best results.Beyna and Wystup . Figure the dierences are limited by 1 · 10 The histogram of changes in the minimal value in Figure −4 for 90 of 136 cases. A signicant criterion for reasonable optimization is the stability with respect to changes in the starting points. Figure 9 and Figure 10 show the histograms of dierences in the minimal value by changing the initial points and the corresponding changes in the position given by the coecients.

01. Summarizing all results. the algorithm produces adequate results in a reasonable time and is robust with respect to changes in the initial values.1 seconds.4.07. ulated Annealing Description The Simulated Annealing method is a (global) optimization method deliver- . 500 an appropriate minimum.2. In addition to the quality and the stability of the algorithm. 85% of the cases do not exceed the squared dierence of 2 · 10−4 . Figure 10 shows that 111 of 136 (82%) tests with varied initial values produce dierences in the coecients of less than 0. 5. there are no cases producing dierences of at most larger than 0.On the Calibration of the Cheyette Interest Rate Model 30 Figure 10: Histogram of changes in the coecients of minima for dierent initial values in the Downhill Simplex algorithm. the algorithm converges in 5 min. To be more precise.2.4 5. The used algorithm converged for every set of market data and arbitrary initial values.Beyna and Wystup . Furthermore.1 Sim Assuming the mentioned time for a function valuation of 0. at the latest to iterations in average. it has taken 3. the convergence of the method is one of the most important evaluation criteria. These results imply more than that the Downhill Simplex algorithm applied to the calibration is stable with respect to changes in the initial values.

2. especially the way of freezing and crystallizing liquids. The Boltzmann probability distribution controls the likelihood of exploring the space and allowing temporally worse movements. Unfortunately it does not guarantee the convergence to the global optimum. this algorithm reduces exactly to the Downhill Results Simulated Annealing method to the same market data. 2008 and ending in Jan. According to the results in section 5.Beyna and Wystup . 5. the thermal mobility of atoms decreases for cooling liquids and ends in crystals forming the state of minimum energy for this system. In nature.3. the algorithm makes use of the Downhill Simplex method. this generator should be robust and stay ecient even in narrow valleys. The modication allows uphill movements whose maximal length depends on a synthetic temperature schedule. As previously described.2. We used the same set of swaptions as described in section In analogy to the tests of the Downhill Simplex algorithm. The probability is linked to the `temperature' and the algorithm converges in dependence to a predened annealing schedule. Empirical tests have shown that the temperature should be reduced suciently slowly to achieve the best results [Pre02]. The method of Simulated Annealing pics up this idea and allows more movements even uphill (worse solution) for `high temperature'. Consequently we based the implementation of the Simulated Annealing method on a modied Downhill Simplex algorithm. We applied the calibration method to a series of market data starting in Jan.3. 2010. The minimal values of the calibration in the time series vary between . Following the ideas of [Pre02] In addition to the annealing schedule. there must be a procedure for it- T controlled in the annealing T → 0.On the Calibration of the Cheyette Interest Rate Model 31 ing remarkable results if a desired global extremum is hidden among many local optima. Thus it is possible that the algorithm leaves local minima or valleys and increases the probability of nding the global extremum. but especially combinatorial minimizations like the `Traveling Salesman Problem' were solved eciently by this technique.2 we used parameter set 1.4. The method of Simulated Annealing is inspired by thermodynamics.2. In the limit Simplex method.2 erating the solution. Figure 11 presents the results of these calibrations to at-the-money swaptions in form of the minimal values according to equation (25) and compares them to the corresponding outcomes of the Downhill Simplex algorithm. we applied the 5.

July 09 and 4. As usual. . Therefore. The level of minimal values measuring the quality of the calibration is signicantly lower in comparison to the results of the Downhill Simplex algorithm. The histogram of changes in the minimal value in Figure 12 shows that 136 of 171 (80%) cases end up in minimal values whose squared dierences do not exceed 2 · 10−4 .1 · 10−5 in In addition to the level of minimal values. we applied 171 dierent initial values.3 · 10−4 in Jan. market data. And the check of the calibration quality according to 4. 1.Beyna and Wystup .4 shows that it fullls the highest standards in any case.. In total. we analyzed the sensibility of the optimization method with respect to the initial values.) and compared the position of the minima according to the vector norm ( 28). 10. The dierences in the minimal values and in the coecients of the minima are presented in Figure 12 and Figure 13 . we varied the starting points in a xed scenario (parameter.. the range is much tighter although the level of corresponding swaption volatilities enlarges in times of nancial crisis.On the Calibration of the Cheyette Interest Rate Model 32 Figure 11: Time series of minimal values of the calibration to swaptions based on the Simulated Annealing algorithm and comparison to the Downhill Simplex algorithm. Figure 13 shows that130 of 171 (76%) of initial values .

The number of these alternative minima is small (27 of 171) and the minimal values do not enlarge.5·10−3 with respect to the vector norm. We tested several types like linear and exponential cooling schemes and some adaptive schemes.e.2. end up in a minimum whose coecients do not vary more than 2. In addition. The setup of the calibration problem does not imply a unique (global) minimum and it might happen that there are several (local) minima with values close to the global one. that the linear cooling scheme delivers the best results. Due to the operating mode. one can change the starting temperature to . only the coecients of the minimum vary slightly. the Simulated Annealing method explores the optimization space much further. Thus it is not surprising. the algorithm does not appear as stable as the Downhill Simplex algorithm concerning the position of the minimal values.On the Calibration of the Cheyette Interest Rate Model 33 Figure 12: Histogram of changes in the minimal values for dierent initial values in the Simulated Annealing algorithm. but the minimal values are signicantly better than the outcomes of the Downhill Simplex Algorithm. at least 0. if the cooling factor is chosen suciently small. nevertheless. The Simulated Annealing method and its behavior depend on the assumed annealing schedule. that the algorithm detects other and better minima.Beyna and Wystup . The test generated furthermore some outcomes with quite large dierences i. But. It turned out. In the test.

Furthermore.2.1 Geneti Summarizing c Optimization Description Genetic Optimization is a stochastic optimization method which incorporates probabilistic elements [Pol08] and uses some analogies The method of . Assuming a xed cooling factor. we used this setup for all calibrations based on Simulated Annealing. The initial temperature and the cooling factor in the the results. The tests have shown that an initial temperature of 5 is suciently large.5 5. determine the maximal probability of accepting worse results and possibly used scheme determine the number of iterations in the algorithm. one should assign the initial temperature as small as possible by ending up in a reasonable minimum. one can state that the Simulated Annealing algorithm delivers good results and it is stable with respect to changes in the initial values. leaving local minima. 5.Beyna and Wystup .5. one can choose a small starting temperature which implies a small number of iterations.On the Calibration of the Cheyette Interest Rate Model 34 Figure 13: Histogram of changes in the coecients of minima for dierent initial values in the Simulated Annealing algorithm with respect to the norm dened in equation (28).2. Consequently.

we need a sucient large set of candidates to reach adequate results with high probability. cross and mutation. The corresponding binary codes are both cut in the middle (after half of the number of bits) and one reaches the representation x = x1 ∪ x2 and y = y1 ∪ y2 . Based on a randomly chosen set of potential solutions. As a last step. If we discretize each interval with a step size of 10−3 . As a second step. As the valuation of the function to minimize is expensive. The Genetic Optimization method searches the (global) optimum just in a discrete and nite space.e. The application is based on the binary code of the possible solutions. the algorithm tries to nd a global optimum by creating new solutions. The algorithm stops. if a solution is found. This is done by combining and adjusting the previous solutions stochastically. This eect is even intensied by the fact.Beyna and Wystup . the cross operator swaps the ends of both binary codes and creates two new solutions x = x1 ∪ y2 and y = y1 ∪ x2 . The mutation operator is applied to a single solution respectively its binary code. The discretization is incorporated by the choice of the binary code specially the number of used bits. Consider for example a 9-dimensional optimization problem with coecients that have values in an interval of length 1. cluded in the creation of the initial set of solutions and in the continuous adjustment. in the optimization The space. that sometimes solutions are reevaluated at subsequent steps. First the cross operator selects a pair of solutions x and y with a predened probability. the selection operator identies the best/worst solutions so far and duplicates the best by deleting the worst at the same time. that is suciently close to the optimum or the set of solution does not vary any longer. the resulting discrete solution space includes 1024 possible values. Each bit is changed with a given probability and consequently new solutions are created stochastically. Therefore each coecient is encoded to a binary representation consisting of single bits and all these codes are combined to a string representing one possible solution. the optimization may take long time to nd the optimum. So. a string of 0/1. The discrete space is created by discretization As the genetic optimization method is based on some stochastic part in- selection operator implies a concentration of the set of solutions somewhere . The iterations within the algorithm are determined by three operators: selection. the detection of the global minimum is not guaranteed. since they could be generated several times. i. The convergence is controlled by the set of initial values and typically we have to use a large set of numbers to explore the space of solutions.On the Calibration of the Cheyette Interest Rate Model 35 to the evolution theory in biology developed by Charles Darwin.

In contrast the number of candidates and the discretization are strongly problem-specic. the optima of the discrete and the continuous space correspond in the limit as well. there are 1024 possible solutions and one should at least use 10. the number of candidates should be chosen suciently high. then the exploration is limited because of the selection operator. ing. One should implement equidistant discretizations in each dimension to simplify the valuation. The elements of the initial set are the origin of the exploration and even if the evolution is smart. one should at least use 10. then the space will not be well explored with high probability. The minimization function for the calibration shows strong sensitivity with respect to changes in some coecients like the constant shift. The discretization spaces are equivalent. As the coecients of the volatility parametrization eect the minimization function dierently. The choice of the probabilities is mainly independent of the problem. If the number is chosen too small.On the Calibration of the Cheyette Interest Rate Model 36 controls the quality of the approximation and in the limit `grid size → 0' these of the continuous one and hence it is an approximation. If the grid size is chosen too large. These values aect the exploration of the optimization space by the algorithm. the probabilities of cross and mutation and the discretization of the domain of each coecient. The Unfortunately the valuation of the minimization function is time consum- .2. The behavior of the algorithm is mainly determined by some specications like the number of candidates. This operator forces a concentration somewhere in the space by duplication of the best solutions. Thus. Thus the grid size has to be chosen quite small to represent all possible values accurately. The range and the grid size of the discretization depends on the possible values of the coecients and on the sensibility with respect to small changes.5. 5. 000 candidates. Within the mentioned example with 1024 possible values. the constructed discrete space is not a good approximation to the continuous one and consequently the results would not contain any conclusions. If there are not many solutions outside this area left. 000 candidates to cover a signicant part. Numerical tests have shown. that the algorithm needs in average at least 500 iterations to identify the global optimum. The number of candidates depends on the dimension of the problem and on the used discretization.2 Results The Genetic Optimization method does not guarantee the convergence to the global minimum. the choice of the grid size should be done separately for each component. Within the mentioned example.Beyna and Wystup .

Unfortunately the algorithm sometimes runs far away from the initial values which destroys this method again. But in combination with a previous Simulated Annealing minimization. Genetic Optimization algorithm is really auspicious. The Powell algorithm oers a good opportunity of dimensional reduction. because it incorporates the possibility to leave local minima with a predened probability. Summarizing. Unfortunately the minimization function does not oer any benecial features. this optimization method can only be applied reasonably to low dimensional calibration problems. In the successful case the results are not superior to the corresponding ones produced by the Downhill Simplex algorithm.On the Calibration of the Cheyette Interest Rate Model 37 function valuation takes in average 0. In order to decrease the computation time. the optimization method is not practicable. Furthermore this methods is robust with respect to changes in the initial values and thus it can reasonably be applied to the calibration problem.3 Conclusions The presentation of the results produced by the dierent minimization algorithms has shown that the Downhill Simplex algorithm delivers reliable and reasonable results. 5. One possible extension not discussed so far is the combination of both. Thus the probability of detecting the global minimum is higher in comparison to the Downhill Simplex algorithm which is originally a local minimization algorithm. Numerical tests have shown that this combination improves the results slightly. because the Simulated Annealing algorithm is based on the Downhill Simplex method quent application just enlarges the number of iterations. one has to decrease the dimension or to coarsen the grid. In combination of a function whose valuation takes quite long. thus each one dimensional The . Unfortunately the space in this setup is often of high dimension. Afterwards one starts the Downhill Simplex algorithm with the results of the Simulated Annealing method as initial values. and in the limit T emperature → 0 both methods are identical. The subse- The Newton algorithm does not produce stable results as it makes use of derivations. it just has to work locally in a normal area. because it explores the space extensively. Nevertheless the Simulated Annealing algorithm seems to produce slightly better results.1 seconds and this implies a total computation time of about 140 hours.Beyna and Wystup . First one applies the Simulated Annealing algorithm with a high initial temperature to explore the space. but this implies worse results and with high probability inadequate ones. This is not really surprising. but there was no substantial gain measurable.

we have applied the calibra- Summarizing. tion to caps and In addition to the calibration to swaptions. Again we have used the market data of Jan. while the surface of July 09 appears quite calm. The extrapolation of these results might imply some negative values. Consequently the Powell algorithm does not work convincingly in the setup of the calibration of the Cheyette Model. Using the volatility parametrization (27) and calibrate the model to 16 swaptions observed at Jan. 08 shows several waves. The minimal values are given by 6. but in this case the volatility surface of Jan. the minimization methods Downhill Simplex and Simulated Annealing deliver the best and reliable results for the calibration of the Cheyette model. 2008 (red) and July 2009 (green) we receive two volatility surfaces presented in Figure 14 . The calibration to caps/oors is evaluated as `passed' in both cases and the minimal value is given by 3. This eect has happened. The results of the calibration as a comparison between the Cheyette model and the Black-Scholes model in terms of implied volatility is given in Table 2 and Table 3 . The implied volatility surfaces have dierent shapes again. oors as well.5 · 10−3 (Jan. 08 (red) and July 09 (green) and the implied surfaces are displayed in Figure 15 .8 · 10−5 (Jan.Beyna and Wystup .On the Calibration of the Cheyette Interest Rate Model 38 optimization is hard to solve globally.5 · 10−3 (July 09). This fact destroys the produced advantages of decomposing the dimensions. 08) and 1. This example demonstrates that the appearance of the volatility surface changes over time and the model should support maximal exibility by the choice of the volatility parametrization. because the calibration is performed on a market fraction. The results of the second calibration violates one secondary condition and thus it is evaluated as `passed'. 08) and 8. 08 appears quite regular and calm.1 · 10−4 (July 09). The implied surfaces have dierent shapes and includes some negative values as well. The calibration to market data observed at the end of Jan. .4 and consequently it is evaluated as `good'. This appearance diers from the structure of the volatility implied by the calibration to swaptions and underlines the need of exible models. The volatility surface of Jan. In addition to the analysis of the behavior and the accuracy of the algorithm we present the results of the calibration in form of the volatility surface. 2008 fullls all conditions on the quality dened in section 4.

0 5.00 % 0.40 % -0. Option Swap Cheyette Black-Scholes Residual lifetime lifetime volatility volatility 0.0 15.25 3.0 3.10 % 0.0 3.10 % 0.71 % 31.0021 4.25 4.0 2.0 31.45 % 41.0294 0.69 % 27.0180 1.35 % 29.0 6.0 1.44 % 33.34 % 17.0138 4.0 18.0 24.0 28.0352 0.0 17.70 % 0.30 % -0.30 % 0.0267 4.0021 4.0 33.Beyna and Wystup .07 % 19.20 % 0.On the Calibration of the Cheyette Interest Rate Model 39 Figure 14: Volatility surface implied by the calibration with the Simulated Annealing method to swaptions with market data observed at the end of Jan.40 % 0.00 % -0.0229 0.0046 1. 2008 and July 2009.0534 1.71 % 17.0 35.50 % -0.0 5.0 38.30 % 25.25 5.0 26.0 4.70 % -0.0 24.0189 Table 2: Results of the calibration to swaptions with market data observed at the end of January 2008.0079 1.62 % 36.70 % 29.0 18. .99 % 33.0236 4.50 % 0.0 28.58 % 17.0 4.0 2.0 6.90 % 0.90 % 0.0 18.0 22.25 6.93 % 18.0119 4.0044 0.03 % 26.99 % 17.50 % -0.0 39.0229 1.25 2. The comparison between the Cheyette model and Black-Scholes model is done in terms of implied Black-Scholes volatilities.

70 % -0.0 1.90 % 0.40 % -0.40 % -0.0 4.0130 0.0 6.0035 1.0029 1.0 15.0 5.40 % 0.0 15.27 % 15.0 15.80 % 0.0029 0.50 % -0.10 % -0.20 % 12. Annealing method to caps/oors with market data observed at the end of Jan.95 % 15.11 % 14.0 15.0031 0.25 3.0 14.0 14.0014 Table 3: Results of the calibration to swaptions with market data observed at the end of July 2009.0077 4.0029 4.0 15.60 % 0.75 % 15.0004 4.63 % 15.07 % 16.09 % 16.0032 4.51 % 15.10 % 14.0 4.25 4.0 16.0033 1.25 5. 2008 and July 2009 Figure 15: Volatility surface implied by the calibration with the Simulated .0017 4.0 14.0 14.0 2.0039 0.60 % -0.0 2.0 14.30 % -0.50 % 0.0 14.0 16.0011 1.40 % -0.90 % 16.27 % 15.11 % 14.0 3.0 5.80 % -0.76 % 15.50 % 0.25 6.Beyna and Wystup .25 2.On the Calibration of the Cheyette Interest Rate Model 40 Option Swap Black-Scholes Cheyette Residual lifetime lifetime volatility volatility 0.0050 4.0002 1. The comparison between the Cheyette model and Black-Scholes model is done in terms of implied Black-Scholes volatilities.0 6.0 16.0 3.20 % -0.10 % 13.52 % 16.0 15.0 16.

3 19.4 20.2 21.4 14.5 The 24.0 17.9 15.2 25.0 18.3 16.1 14.5 14.7 16.8 17.6 21.3 2 19.2 15.2 21.4 14.6 17.5 24.2 16.7 16.8 15.1 14.2 20.4 16.3 18.3 22.1 15.On the Calibration of the Cheyette Interest Rate Model 41 The calibration is based on implied volatilities quoted at the market for swaptions and caps/oors.8 3 17.3 19.1 15.0 15.5 20.5 19.3 16.0 16.3 15.5 24.7 18.4 18.6 1.3 18.3 20.4 17.6 16.3 20.6 17.6 23.4 18.6 14.3 21.5 13.6 16.3 16.1 16.0 22. Swaption volatility matrix observed at the end of July 2008.1 4 16 15. .1 18.Beyna and Wystup . In the following.7 16.8 15.8 16.6 13.9 14.0 13.6 14.2 4 17.6 21.3 19.9 24. we will present the used market data published by `Moosmüller & Knauf AG'.2 16.9 19.9 24.6 16.9 17. Option Lifetime 1 12 2 12 3 12 6 12 9 12 1 20.4 20.3 21.4 16.8 17.2 12.7 21.0 14.2 15.0 14.2 17.9 19.1 14.6 18.4 15.6 15.7 22.6 17.7 23.7 21.2 18.7 15.2 17.7 21.8 Table 5: volatilities are quoted in percentage and the lifetimes are quoted in years.4 19.8 17.0 20.3 14.0 20.2 14.0 15.3 22.8 16.1 23.5 15.9 17.6 16.9 20.1 22.9 20.6 The 21.4 15.6 18.4 13.2 2 18.8 21.9 19.4 15.5 19.0 17.9 21.2 15.9 16.5 23.7 12.9 16.9 12.3 13.1 17.5 22.8 1.3 16.2 17.7 21.6 17.1 23.4 16.5 20.6 16.0 18.7 21.4 19.1 13.5 23.4 20.7 15.3 18.0 17.0 22.9 13.8 18.5 15.9 21.0 13.7 19.3 19.4 26.0 19.9 20.4 14.1 16.9 3 17.6 17.1 20.1 19.0 16.4 16.1 15.7 15.6 14.0 26.7 13.0 14.2 17.7 15. Swap Lifetime 1 2 3 4 5 6 7 8 9 10 Table 4: Swaption volatility matrix observed at the end of January 2008.4 15.9 21.4 16.4 21.5 23.8 17.7 18.0 18.5 15.7 17.1 volatilities are quoted in percentage and the lifetimes are quoted in years.0 18.5 15.7 15.9 17.9 18.4 19.4 19.8 20.1 16. Swap Lifetime 1 2 3 4 5 6 7 8 9 10 A Appendix Option Lifetime 1 12 2 12 3 12 6 12 9 12 1 21.

1 27.7 18.9 26.6 22.2 29.1 20.6 42.6 27.9 24.9 26.6 27.1 20.7 29.1 17.9 18.8 26.6 31.4 34.1 28.3 29.4 24.5 24.5 Table 7: volatilities are quoted in percentage and the lifetimes are quoted in years.0 40.7 34.0 2 26.2 20.0 47.5 39.7 18.1 20.6 26.4 29.5 32.4 30. Swap Lifetime 1 2 3 4 5 6 7 8 9 10 Table 6: Swaption volatility matrix observed at the end of January 2009.8 20.3 20.9 27.2 29.0 27.8 27.0 23.4 53.3 22.9 31.6 3 21.2 34.5 41.6 28.5 17.0 42.1 45.1 The 54.3 52.4 27.0 31.6 32.5 31.7 18.0 31.0 39.5 4 18.0 38. Swaption volatility matrix observed at the end of July 2009.9 38.5 18.5 27.4 36.9 24.0 48.3 17.1 43.2 35.8 19.5 26.8 38.0 29.9 53.6 28.5 32.3 35.4 24.0 32.0 17.7 18.2 26.8 59.2 20.0 17.9 23.1 44.2 22.1 22.4 20.2 20.4 50.5 21.8 30.Beyna and Wystup .1 34.4 51.1 The 72.7 28.1 47.3 49.6 18.5 18.8 3 22.2 33.2 20.9 37.8 18.1 26.7 24.2 25.3 22.4 22.9 26.6 45.0 19.3 22.8 23.8 23.4 38.8 18.5 44.8 40.7 33.6 37.4 28.0 28.9 20.5 45.6 19.1 24.7 18.9 26.6 44.4 18.2 65.8 28.6 18.5 2 27.5 43.5 50.9 43.7 22.1 28.7 4 19.4 26.9 22.0 18.9 20.7 22.1 29.5 25 24.3 21.8 31.9 25.1 33.5 1.5 29.0 17.7 19.0 23.3 27.7 24.3 17.On the Calibration of the Cheyette Interest Rate Model 42 Swap Lifetime 1 2 3 4 5 6 7 8 9 10 Option Lifetime 1 12 2 12 3 12 6 12 9 12 1 40.1 46.8 58.2 volatilities are quoted in percentage and the lifetimes are quoted in years.3 45.8 27.2 22.2 20.6 24.7 28.3 30.9 31.2 22.0 52.7 59.8 23.3 26.4 33.6 25.5 63.4 45.6 32.0 41.4 1.2 17.5 25.5 25.7 29.3 33.8 48.6 55.5 28.2 33.1 20.9 40.8 27. . Option Lifetime 1 12 2 12 3 12 6 12 9 12 1 42.

1 22.3 20.7 volatilities are quoted in percentage and the lifetimes are quoted in years.5 24.6 19.6 21.5 17.2 20.0 39.9 38.3 17.4 14.3 23.2 24.3 19.1 33.7 28. 2009 July 2009 Jan.0 24. The volatilities are quoted in percentage and the lifetimes are quoted in years.6 19.0 22 20.6 43.4 17.5 63.5 59.6 29. Table 9: Cap (ATM) volatility matrix observed at dierent times.5 18.4 20.5 20.1 20.9 23.3 17.8 16.4 31.3 43.1 22.7 26.2 28.3 18.4 29.9 17.7 21.7 28.5 28.8 38.6 23.On the Calibration of the Cheyette Interest Rate Model 43 Swap Lifetime 1 2 3 4 5 6 7 8 9 10 Option Lifetime 1 12 2 12 3 12 6 12 9 12 1 47.5 21.0 32.7 25.8 24.8 19.4 16.6 40.7 1.7 14.0 2 34.5 2 3 4 5 6 7 8 9 10 17.2 20.7 21.3 18.8 19.2 15.8 22.0 37.6 51.1 26.8 23.Beyna and Wystup .0 18.1 21.0 15.5 21.5 48.1 Table 8: Swaption volatility matrix observed at the end of January 2010.8 22.7 46.7 15.5 20.5 23.2 52.8 22.7 21.8 24.0 22.2 25.9 19.1 18.5 23.3 33.5 39.3 24.8 18.4 17.1 22.6 26.9 19.1 19.7 19.8 17.0 18.1 16.4 25.4 19. 2010 Lifetime 1 1.6 32.9 39.8 18.0 22.1 22. Dates Cap Jan.0 17.3 33.2 18.0 29.3 28.8 20.8 25.1 51.4 20.1 19.0 17.4 18.1 18.5 20.3 3 27.2 4 23.9 33.7 16. .8 49.2 19.9 20.9 27.0 33.3 The 45.6 17.9 21.5 19. 2008 July 2008 Jan.0 28.3 17.9 47.8 33.2 14.8 26.9 20.6 16.6 18.5 26.7 38.1 20.3 38.2 16.7 21.0 20.2 21.5 20.4 24.8 29.0 21.3 21.3 45.

Farshid Jamshidian. Alan Brace. Marek Musiela. Bond Pricing and the Term Structure of Interest Rates: A New Methodology for Contingent Claims Valuation. Robert Jarrow. 1997. Working Paper Gen Re Securities. 2007. . BARRA Inc. John C. Riccardo Poli. 1997. Books on Demand. Finance and Stochstics . Jan. EconWPA. 2 edition. 2000. Oren Cheyette. 1992. Algorithms for the Minimization without Derivatives. Springer. 2005. A Simplex Method for Function Minimization. 5 edition. Futures. 1973. 2008. Jesper Andreasen. October 2003. Roger Mead. Back to the future. Turbo Charging the Cheyette Model. Econometrica. 7:308313.On the Calibration of the Cheyette Interest Rate Model 44 References [And02] [And05] [Bre73] [Che94] [Hen03] [Hul05] [Jam97] [KL07] [Mea65] [Mer05] [Mor92] [Mus97] [Pol08] [And00] Jesper Andreasen. 60(1):77105. LIBOR and Swap Market Models and Measures. rst edition. and Other Derivatives . John Nelder Ashworth. working paper . 7:127147. Richard P. Andrew Morton. December 2002. Brent. 2005. Mathematical Finance . The Market Model of Interest Rate Dynamics. Interest Rate Models: Theory and Practice. Markov Representation of the Heath-JarrowMorton Model. David Heath. Volatile Volatilities. Dariusz Gatarek. September 2005. Lulu Press. Leif Andersen. May 1994.Beyna and Wystup . Explicit bond option and swaption formula in HeathJarrow-Morton one factor model. 1:293330. Peter Kohl-Landgraf. Finance. Financial Times. Hull. Risk Magazine. 1965. A Field Guide to Genetic Programming . Damiano Brigo. Marc Henrard. Fabio Mercurio. Risk Magazine. Jesper Andreasen. Prentice Hall. Computer Journal . Options. PDE Valuation of Interest Rate Derivatives: From Theory To Implementation .

Beyna and Wystup . The Review of Financial Studies . Oldrich icek. 2003. Closed Form Solutions for Term Structure Derivatives with Lognormal Interest Rates. Springer Series in Computational Mathematics. 3(4):573592. Alan White. [Zag02] Rudi Zagst. 5:177188. Josef Stoer. Pricing interest-rate derivative securities. [Sto03] Florian Jarre. 1977. [San97] Kristian R. Journal of Financial Economics . 1990. Interest-Rate Management. [Vas77] Vas C a a sa Structure. Klaus Sandmann. Optimierung. Numerical Recipes in C++: The Art of Scientic 2002. 52:409430. An Equilibrium h r cteri tion of the Term Computing. Journal of Finance. 2002. second . Springer.On the Calibration of the Cheyette Interest Rate Model 45 [Pre02] William H. [Whi90] John Hull. edition. 1997. Press. Miltersen. Springer.

2010 2010 2010 2010 2010 2010 139. Rainald Ratingmodell zur Quantifizierung des Ausfallrisikos von LBO-Finanzierungen Bannier. Christian The economic function of credit rating agencies . / Schalast. Jens Modellierung des Kreditrisikos im Portfoliofall Cremers. Carsten Social Capital. Carsten Rethinking Evolution. André The Value-Added of Investable Hedge Fund Indices Herrmann-Pillath. Sebastian G. 2010 2010 2010 2010 2010 2010 2009 2009 2009 2009 131. Sebastian T. Michael / Cremers. Heinz / Walzner. Alexander Words or Deeds – What Matters? Experience of Decentralization in Russian Security Agencies Kostka. / Hach. Dieter G. 142. Christian / Winker. Christina E. . Thomas / Schmaltz. Michael Funktionsweise und Replikationstil europäischer Exchange Traded Funds auf Aktienindices Libman. Carsten / Libman. Yassin / Rauch. Thomas / Rogalski. Xiaofan State and market integration in China: A spatial econometrics approach to ‘local protectionism’ Lang. Chinese Style: Individualism. 146. Carsten Dimensionen des Wissens: Ein kognitiv-evolutionärer Ansatz auf der Grundlage von F. 128. Genia / Zhou. 135. 144 143. Thomas / Kahlert. 124. Christian Market Liquidity: An Introduction for Practitioners Herrmann-Pillath. 2009 2009 2009 2009 2009 2009 46 Frankfurt School of Finance & Management Working Paper No. Heinz / Hentze. 129. 138. 137. and Taxes: Contradictions of Different Aspects of Decentralization Herrmann-Pillath. 126. Carsten Entropy. Sabrina Are SMEs large firms en miniature? Evidence from a growth analysis Heidorn. André Determinanten von Banken-Spreads während der Finanzmarktkrise Bannier. Jianghua Chinese firms entering China's low-income market: Gaining competitive advantage by partnering governments Herrmann-Pillath. 130. / Hirsch. 140. Relational Collectivism and the Cultural Embeddedness of the Institutions-Performance Link Schäffler. 125. Marc It’s the Market Power. Author/Title Libman. Alexander Constitutions. Heinz / Walzner. 132. 141. Christina E. Christina / Feess.What does the watchlist tell us? Year 2010 2010 2010 145. Stupid! – Stock Return Patterns in International Bank M&A Herrmann-Pillath.A. von Hayeks Theorie der „Sensory Order“ Hankir. Christian / Umber. Carsten The Evolutionary Approach to Entropy: Reconciling Georgescu-Roegen’s Natural Philosophy with the Maximum Entropy Framework Heidorn. Christoph M&A im Bereich Erneuerbarer Energien Birkmeyer. 136. / Behr. 148. / Metz. 133. Regulations. Thomas / Kaiser. Thomas / Löw. Entropy and Economics: A triadic conceptual framework for the Maximum Entropy Principle as applied to the growth of knowledge Heidorn.FRANKFURT SCHOOL / HFB – WORKING PAPER SERIES No. Jens Modellierung des Kreditrisikos im Einwertpapierfall Heidorn. 147. Christian / Schmaltz. Patrick / Güttler. Jörg / Heidorn. Christian Interne Transferpreise für Liquidität Bannier. 127. Function and Evolution: Naturalizing Peircian Semiosis Bannier. Andre Rating opaque borrowers: why are unsolicited ratings lower? Herrmann-Pillath. / Voinea. Christina E. Carsten Outline of a Darwinian Theory of Money Cremers. Alexander / Yu. 134. Eberhard When high-powered incentive contracts reduce performance: Choking under pressure as a screening device Herrmann-Pillath. Dennis Implied Correlations of iTraxx Tranches during the Financial Crisis Fritz-Morgenthal.

Carsten Diversity Management und diversi-tätsbasiertes Controlling: Von der „Diversity Scorecard“ zur „Open Balanced Scorecard Hölscher. Kay-Michael / Scholl. Herrmann-Pillath. Sebastian The Genesis of the Black-Scholes Option Pricing Formula Heimer. Christina E. Luise / Haug. der Türkei und in den GUSStaaten Heidorn.) / Schanz. Adalbert / Vogel. Peter Die Rolle des Internets als Informationsbeschaffungsmedium in Banken Herrmann-Pillath. Claus / Siepmann. Michael Das IRB-Modell des Kreditrisikos im Vergleich zum Modell einer logarithmisch normalverteilten Verlustfunktion Heidorn. Frankfurt School of Finance & Management Working Paper No. Christoph: Staatsfonds – „neue“ Akteure an den Finanzmärkten? Schalast. 103. Stephan Einführung in das Kapitalstrukturmanagement bei Banken Rossbach. 101. 110. Thomas / Hölscher. 120. Is there a hold-up benefit in heterogeneous multiple bank financing? Roßbach. Carsten The Naturalistic Turn in Economics: Implications for the Theory of Finance Schalast. Johannes Konstruktion einer Anleihe mit hypothekarischer Besicherung Schalast. 2008 2008 2008 2008 99. 118. 114. 102. Andreas Analyse von Steueramnestiedaten Heimer. Heinz / Vetter. Thomas / Pleißner. Michael / Schweinberger. Michael H. Luise / Werner. 107. 111. Christoph / Benita. Carsten Kulturelle Hybridisierung und Wirtschaftstransformation in China Schalast. Stefan Die Auswirkungen der Subprime-Krise auf den deutschen LBO-Markt für Small. Barten Private Equity und Familienunternehmen – eine Untersuchung unter besonderer Berücksichtigung deutscher Maschinen. Christoph / Bolder. 47 ./ Müsch. Sven Erfolgsfaktoren von Private Equity Fonds Bannier. Naturalism and Language Schalast. 97. Dieter G. Mathias Determinanten Europäischer CMBS Spreads. 117. Carsten Moralische Gefühle als Grundlage einer wohlstandschaffenden Wettbewerbsordnung: Ein neuer Ansatz zur erforschung von Sozialkapital und seine Anwendung auf China Heidorn. 2009 2009 2009 2009 2009 2009 2009 2009 2009 109.123. Karl / Moormann. Luise / Clasen. Christoph (Hrgs. Peter / Gießamer. 2009 2008 2008 2008 2008 2008 2008 2008 106. Carsten A Neurolinguistic Approach to Performativity in Economics Winkler. Dirk Ein eLearning-System zur Unterstützung der Wissensvermittlung von Web-Entwicklern in Sicherheitsthemen Herrmann-Pillath. Thomas / Arend. Markus / Radünz. Matthias Ralf Access to Finance and Venture Capital for Industrial SMEs 2009 2009 2009 2009 2009 121.und Anlagenbauunternehmen Bannier. Equity Gap? – Which Equity Gap? On the Financing Structure of Germany’s Mittelstand Herrmann-Pillath.) / Schanz. 122. 115. 2008 100. Stephanie / Weber. Ein empirisches Modell zur Bestimmung der Risikoaufschläge von Commercial Mortgage-Backed Securities (CMBS) Schalast. Christoph (Hrsg. Christina E. Thorsten Transaktionen und Servicing in der Finanzkrise: Berichte und Referate des Frankfurt School NPL Forums 2008 Werner. Carsten Neuroeconomics. 105. 98. Christoph Empirische Analyse der Drawdowns von Dach-Hedgefonds Herrmann-Pillath. 116.-Entscheidung des US District Court for the Southern District of New York Hölscher. / Roder. Ursula Finanzierungsstrukturen und makroökonomische Stabilität in den Ländern Südosteuropas. 104. Thomas / Rupprecht. 113. 108. Christoph / Alram. Wolfgang Aktionärsschutz in der AG falsch verstanden? Die Leica-Entscheidung des LG Frankfurt am Main Bannier. Jürgen Efficiency and Profitability of European Banks – How Important Is Operational Efficiency? Herrmann-Pillath.und MidCaps Cremers. 119. 112. Kay-Michael Schaeffler KG/Continental AG im Lichte der CSX Corp. / Grote. Christina E. Thomas / Kaiser.

82. Thomas / Schmaltz. 85. Thomas Loss Given Default Modelle zur Schätzung von Recovery Rates Almer. 92. 88. Christina E. Clemens J. Thomas Commodities in Asset Management Cremers. Wolfgang / Schmaltz. Rainer / Safran. Bernd / Sprengetter. Christian / Kaiser. Heinz / Löhr. Böttger. 76. Heterogeneous Multiple Bank Financing: Does it Reduce Inefficient Credit-Renegotiation Incidences? Cremers. 48 Frankfurt School of Finance & Management Working Paper No. Christian The Economics of Rating Watchlists: Evidence from Rating Changes Demidova-Menzel.Wann sind stabile Ratings optimal und welche Anforderungen sind an optimale Berichtsregeln zu stellen? Bannier. Thomas / Hoppe. 70. Ingo 10 Jahre deutsche Buyouts Bannier. 77. 94. 75. Christian The Dynamics of Short. Barbara Monetary Analysis: A VAR Perspective Heidorn. Christoph Work-Out und Servicing von notleidenden Krediten – Berichte und Referate des HfB-NPL Servicing Forums 2006 Abrar. Robert Der deutsche NPL Markt 2007: Aktuelle Entwicklungen. Christoph Distressed Debt in Germany: What´s Next? Possible Innovative Exit Strategies Belke. Thorsten How the ECB and the US Fed set interest rates Heidorn./ Hirsch. Hans Otto / Ockens. Jens Risikosteuerung mit Kreditderivaten unter besonderer Berücksichtigung von Credit Default Swaps Cremers. Christoph / Morgenschweis. Patrick Handlungsalternativen einer Genossenschaftsbank im Investmentprozess unter Berücksichtigung der Risikotragfähigkeit Gerdesmeier. Stefan / Löchel. Christoph / Schanz. Florian The Strategic Value of Investments in Chinese Banks by Foreign Financial Insitutions Schalast. Andrea Portfoliooptimierung mit Hedgefonds unter Berücksichtigung höherer Momente der Verteilung Jobe. Christoph / Stralkowski. / Ockens. 73. 80.and Long-Term CDS-spreads of Banks Barthel. . Methoden und Status quo Löchel. 83. Anja / Heidorn. 72. 86. Marc / Guthoff. 84. 95. Robert / Schalast. Ansgar / Polleit. Andreas Produktivität und Effizienz in Banken – Terminologie. Christina “Smoothing“ versus “Timeliness“ . Dieter G. 71. Heinz / Traughber. Dieter G. Verkauf und Bewertung – Berichte und Referate des NPL Forums 2007 Schalast. Erich / Wollersheim. Nadeshda / Heidorn. Kay-Michael Wertpapierprospekte: Markteinführungspublizität nach EU-Prospektverordnung und Wertpapierprospektgesetz 2005 Dickler. Dieter / Roffia. Dennis Determinants of banks' engagement in loan securitization Bannier. Robert A. 79. Thomas / Kaiser. Jutta Kulturunterschiede bei Mergers & Acquisitions: Entwicklung eines Konzeptes zur Durchführung einer Cultural Due Diligence Heidorn.96. 87. Thomas Gold in the Investment Portfolio Hölscher. / Schalast. Klaas / Safran. Christina E. Thomas / Heidorn. Christoph Fusionskontrolle in dynamischen Netzsektoren am Beispiel des Breitbandkabelsektors Schalast. Andreas Deskription und Bewertung strukturierter Produkte unter besonderer Berücksichtigung verschiedener Marktszenarien Demidova-Menzel. Heterogenität von Hedgefondsindizes Baumann. Klaas / Stachuletz. Horst The Endogeneity Approach of the Theory of Optimum Currency Areas . 81. Kamyar / Schalast. Nadeshda / Heidorn. Luise / Rosenthal. 2007 2007 2007 2007 2007 2007 2007 2006 2006 2006 2006 2006 2006 2006 78. Christina / Hänsel. 74.What does it mean for ASEAN + 3? 2008 2008 2008 2008 2008 2008 93. Thomas / Kunze. 90. 91. Horst / Pecher. 2008 2008 2007 2007 2007 2007 2007 89. Christian Liquiditätsmodellierung von Kreditzusagen (Term Facilities and Revolver) Burger. / Muschiol. Heinz / Walzner. Johannes Leistungsmessung der Internen Revision Bannier.

Thomas / Siragusano. 46. Norbert Internationale Cash Flow-Rechnungen aus Eigner. Rauno / Kühn. Gernot M. Luise Financing the Embedded Value of Life Insurance Portfolios Schalast. Christian / Kaiser.und Kapitalmarktrechts I : Non-Performing-Loans / Faule Kredite .69. Lars Investitionen in Collateralized Debt Obligations Kahlert. 49. 67.S. 56. Christoph (Hrsg. Christoph / Schanz.Handel. Andreas / Hesdahl. Italien und den Niederlanden Gerdesmeier. Heidorn. 64. Horst / Polleit. Michael / Schlink. Christian Temperaturderivate zur strategischen Absicherung von Beschaffungs. Thomas / Trautmann. 60. Christoph / Thöne. Thomas / König. 58. Holger / Seeger. 61. Norbert Bilanzierung von Unternehmenszusammenschlüssen nach US-GAAP 2005 2005 2005 2005 2005 2005 2005 62. 45. Notes on convexity and quanto adjustments for interest rates and related options Hess. macroeconomic releases Cremers. Jörg / Schalast. Norbert Wertsicherungsstrategien für das Asset Management Löchel. 65. 48. Thomas Investitionen und Emissionen von Convertible Bonds (Wandelanleihen) Chevalier. Work-Out. Thorsten The Slowdown in German Bank Lending – Revisited Heidorn. Christian / Schalast. Dieter / Roffia.„Staying Public“ oder „Going Private“? . Thomas / Meyer. Thomas / Krieger. 43. Christoph Modernisierung der Wasserwirtschaft im Spannungsfeld von Umweltschutz und Wettbewerb – Braucht Deutschland eine Rechtsgrundlage für die Vergabe von Wasserversorgungskonzessionen? – Bayer. 63. Dieter / Polleit. Marcus / Cremers. 47. 2003 2003 2003 Frankfurt School of Finance & Management Working Paper No. / Seeger. Thorsten A case for money in the ECB monetary policy strategy Richard. Wolfgang M. Daniel / Schalast. Pierre / Heidorn. Perham / Hölscher. Dieter Determinants of the relative price impact of unanticipated Information in U. Möglichkeiten der Strukturierung von Hedgefondsportfolios Belke. Heinz / Kluß. Wolfram / Schmidt. Thorsten (How) Do Stock Market Returns React to Monetary Policy ? An ARDL Cointegration Analysis for Germany Daynes. Erfolgsabhängige Vergütungsmodelle deutscher Publikumsfonds Heidorn. Ansgar / Polleit. Heinz / Kluß. Markus Incentive Fees. Thorsten Measures of excess liquidity Becker. Barbara The Relevance of real-time data in estimating reaction functions for the euro area Barthel. 54. 49 .und Kapitalmarktsrechts II: Distressed Debt . 55. Tindaro Die Anwendbarkeit der Behavioral Finance im Devisenmarkt Schütze. / Harding. Dieter G. Norbert / König. Dietmar / Binder. Thomas / Hoppe. 51.Nutzenanalyse der Börsennotiz Heun.Investing in Deutschland Gerdesmeier. Thomas / Köhler. Erich / Gierig.Implementation of a currency crisis model for Uganda Heimer. Alexandra Niederschlagsderivate Heidorn. 66. 57. Outsourcing und Securitisation Polleit. 59. 2005 2005 2004 2004 2004 2004 2004 2004 2004 2004 2004 2004 2004 2003 2003 2003 2003 53.und Absatzrisiken Becker. Alexander Performanceeffekte nach Directors´Dealings in Deutschland. 52. Kay-Michael Unternehmen im Prime Standard . 68. Mirko / Heidorn.) Wider die Verschleuderung von Unternehmen durch Pfandversteigerung Gerhold. Thomas Aktuelle Rechtsfragen des Bank. 44. Christoph Aktuelle Rechtsfragen des Bank.und Gläubigersicht Boenkost. Torsten Early Warning Systems of Financial Crises . Bernd / Pietrowiak. 50. Thomas Auswirkungen des Basel II Akkords auf österreichische KMU Heidorn. Ilmhart-Wolfram Unterschiedliche Ansätze zur Messung des Humankapitals Anders. Ralf / Schalast. Gernot M.

Daniel / Cremers. Thomas / Graf Waldstein. Thomas / Rütze. 37. Ludger / Schmidt. 32. Frank Economic Value Added zur Prognose der Performance europäischer Aktien Cremers. Jürgen Fit für den Euro. Thomas / Kantwill. 25. Sophie / Thiele. Merle Gründung einer deutschen Strombörse für Elektrizitätsderivate Deister. Michael Portfoliooptimierung mit Hedge Fonds unter besonderer Berücksichtigung der Risikokomponente Heidorn. 40. 21. 27. 23. 22. Thomas / Jaster. Peter / Seiler.Eine Alternative zur vorherrschenden Kapitalmarkttheorie? Heidorn. Heinz / Robé. 24.42. Thomas / Schmidt. Sven Einführung in die fundamentale Aktienanalyse Seeger. Norbert International Accounting Standards (IAS) Moormann. Beiträge von Studierenden des Studiengangs BBA 012 unter Begleitung von Prof. Henner / Seeger. Birgit Die Eigenmittelkonzeption des § 10 KWG Cremers. Horst Die Auswirkungen des Übergangs zum Kapitaldeckungsverfahren in der Rentenversicherung auf die Kapitalmärkte Heidorn. Ulrich Event Risk Covenants Biswas. 36. Peter Behavioral Finance . Horst Recent Trends in U. Rita / Löchel. Dr. 29. 28. 31. Heinz Value at Risk-Konzepte für Marktrisiken Chevalier. Wolfgang Modeling Default Dependence with Threshold Models Balthasar. 30. Heinz Optionspreisbestimmung Cremers. 50 Frankfurt School of Finance & Management Working Paper No. Pierre / Heidorn. Oliver / Willeitner. Hans-Dieter / Siebrecht. 26. Andreas / Heidorn.Eine Untersuchung am europäischen Aktienmarkt Cremers.S. Jürgen / Stehling. Horst Die ökonomischen Dimensionen der ‚New Economy‘ Frank. 17. 16. . Thomas / Klein. Sven / Heidorn. Norbert Bilanzierung von Finanzderivaten nach HGB. Jürgen Grenzen des Outsourcing: Eine Exploration am Beispiel von Direktbanken Heidorn. Thomas Entscheidungsorientierte Mindestmargenkalkulation Wolf. 34. Jens Eine empirische Analyse der Spreadunterschiede von Festsatzanleihen zu Floatern im Euroraum und deren Zusammenhang zum Preis eines Credit Default Swaps Böttcher. Frank Strategic Positioning of E-Commerce Business Models in the Portfolio of Corporate Banking Sokolovsky. Thomas / Weier. Heinz Konvergenz der binomialen Optionspreismodelle gegen das Modell von Black/Scholes/Merton Löchel. Günter Georg / Löchel. EstG. Philipp Hybrides Kernkapital für Kreditinstitute Heidorn. Zbynek / Strohhecker. Jürgen Terminologie und Glossar der Bankinformatik Heidorn. 33. Thomas Bewertung von Kreditprodukten und Credit Default Swaps Heidorn. Axel / Moormann. 2003 2002 2002 2003 2002 2001 2001 2001 2001 2001 2001 2001 2001 2001 2000 2000 2000 2000 2000 2000 2000 2000 2000 1999 1999 1999 1999 38. 19. Simulationsbasierte Euro-Maßnahmenplanung für Dresdner-Bank-Geschäftsstellen Roßbach. IAS und US-GAAP Moormann.HGB-Bilanzierung im Wettbewerb mit den internationalen Standards nach IAS und US-GAAP Overbeck. Stefan Neue Möglichkeiten durch die Namensaktie Böger. Daniel / Ehrlicher. 15. 39. 35. Dirk Beta als Risikomaß . 20. 18. Heinz / Schmidt. Norbert Seeger Rechnungslegung im Umbruch . and German Banking: Convergence or Divergence? Eberle. Thomas CatBonds 2003 41.

19. 03. Uwe FX Basket Options Weber. 09. 22. Wolfgang Latin hypercube sampling with dependence and applications in finance Hakala. Jürgen Lean Reporting und Führungsinformationssysteme bei deutschen Finanzdienstleistern 1999 1999 1999 1999 1998 1998 1998 1998 1998 1998 1997 1997 1996 1995 FRANKFURT SCHOOL / HFB – WORKING PAPER SERIES CENTRE FOR PRACTICAL QUANTITATIVE FINANCE No. Wolfgang LIBOR in Arrears Jahresbericht 1997 Ecker. 17. Dimitri / Wystup. 14. Credit dynamics in a first passage time model with jumps Reiswich. Jürgen Terminologie und Glossar der Bankinformatik Löchel. Thomas / Moormann. 15. Thomas Kreditderivate Heidorn. Jürgen Die Bank als Betreiberin einer elektronischen Shopping-Mall Jahresbericht 1996 Cremers. 04. Natalie / Schlögl. Lutz / Schmidt. Carlos / Wystup. Thomas / Schmidt. Robert Potential PCA Interpretation Problems for Volatility Smile Dynamics Keller-Ressel. Jürgen / Wystup. Horst The EMU and the Theory of Optimum Currency Areas Löchel. Fiodar Forward-Start Options in the Barndorff-Nielsen-Shephard Model Griebsch. Uwe On the Valuation of Fader and Discrete Barrier Options in Heston’s Stochastic Volatility Model Veiga. 18. Jochum. Jürgen Die Umstellung auf die Stückaktie für deutsche Aktiengesellschaften Moormann. Martin / Kilin. Manuel L. 08. Wolfgang M. Jürgen Stand und Perspektiven der Informationsverarbeitung in Banken Heidorn. 05. 16. Uwe Vergleich von Anlagestrategien bei Riesterrenten ohne Berücksichtigung von Gebühren. 23. Eduard Hoshin Kanri / Management by Policy (MbP) Heidorn. Carlos / Wystup. Uwe Unifying Exotic Option Closed Formulas Packham. Uwe Closed Formula for Options with Discrete Dividends and its Derivatives Packham. 01. Credit gap risk in a first passage time model with jumps Packham. 21. Andreas / Wystup. Dimitri / Tompkins. 06. Willi Interpolation of Discount Factors Moormann. Susanne / Wystup. 11. Thomas / Hund. 10. Lutz / Schmidt. 51 . Natalie / Schmidt. Horst Die Geldpolitik im Währungsraum des Euro Heidorn. Uwe FX Volatility Smile Construction Reiswich. 20. 07. 13. Wolfgang M. 02.14. Eine Simulationsstudie zur Verteilung der Renditen Year 2010 2009 2009 2009 2009 2008 2008 2008 2008 2008 2008 Frankfurt School of Finance & Management Working Paper No. Natalie / Schlögl. 13. Thomas Kreditrisiko (CreditMetrics) Moormann. / Veiga. Author/Title Esquível. 12. Heinz / Schwarz.

Susanne/ Kühn. 11. Uwe Efficient Computation of Option Price Sensitivities for Options of American Style 2008 2008 2008 2008 2008 2007 2007 2007 2006 2005 2004 2004 HFB – SONDERARBEITSBERICHTE DER HFB .BUSINESS SCHOOL OF FINANCE & MANAGEMENT No. Cross currency swap valuation Wallner. Interest Rate Convexity and the Volatility Smile Becker. 10.frankfurt-school. Wolfram / Schmidt. Uwe Was kostet eine Garantie? Ein statistischer Vergleich der Rendite von langfristigen Anlagen Schmidt. Christoph / Wystup. Uwe Foreign Exchange Symmetries Becker.html CPQF: http://www. Germany Phone: +49 (0) 69 154 008 – 734 Fax: +49 (0) 69 154 008 – 728 eMail: m. 02.fs. Christoph/ Wystup.--) Year 2003 Printed edition: € 25.frankfurt-school.12.00 + € 2. 01. . Wolfgang Default Swaps and Hedging Credit Baskets Kilin.de/content/de/research/quantitative_Finance/research_publications. Andreas / Wystup. 06.de/content/de/research/Publications/list_of_publication0. Wolfgang M.50 shipping Download: Working Paper: http://www. Uwe On the Cost of Delayed Currency Fixing Boenkost. Fiodor Accelerating the Calibration of Stochastic Volatility Models Griebsch.de Further information about Frankfurt School of Finance & Management may be obtained at: http://www. 09. Uwe Instalment Options: A Closed-Form Solution and the Limiting Case Boenkost. Eine Simulationsstudie zur Verteilung der Renditen Wystup. Author/Title Nicole Kahmer / Jürgen Moormann Studie zur Ausrichtung von Banken an Kundenprozessen am Beispiel des Internet (Preis: € 120. Uwe Riesterrente im Vergleich. Wolfram / Schmidt. Uwe Foreign Exchange Quanto Options Wystup. Christian / Wystup.html Order address / contact Frankfurt School of Finance & Management Sonnemannstr. Uwe Vanna-Volga Pricing Wystup. Wolfgang M.de 52 Frankfurt School of Finance & Management Working Paper No. 9 – 11 D – 60314 Frankfurt/M. 04. 03. 08. 05. 07. Weber.biemer@fs. 01. Christoph / Wystup.

Sign up to vote on this title
UsefulNot useful