Libor Market Model:

Specification and Calibration

Alex Ferris
April 27, 2012
ESE 499: Senior Design Project
Washington University in St. Louis

Supervisor:
Anatoliy Belaygorod, Ph.D.
Vice President of Quantitative Risk—R.G.A.
Adjunct Professor of Finance—Olin Business School
belaygorod@wustl.edu
Abstract

In this paper I implement and calibrate the Lognormal Forward-LIBOR Model (LFM) for the term

structure of interest-rates. This model is a subset of the LIBOR Market Model class of stochastic interest-

rate models and is characterized by the lognormal distribution of forward LIBOR rates under appropriate

numeraires. Specifically, I implemented the LFM under two different instantaneous volatility

formulations, Rebonato’s 6.21a (2002) and Brigo and Mercurio’s Formulation 7 (2006). The

implementations are then calibrated to market data for Caps and Swaptions. Finally, the results are

analyzed for their accuracy and their correspondence to financial theory and intuition. All

implementation and calibration is done in MATLAB.

Keywords: Calibration, Caps, LIBOR Market Model, LFM, Lognormal Forward-LIBOR Model, MATLAB,

Stochastic Interest-rate Model, Swaptions.

Acknowledgements

I would first like to thank Anatoliy Belaygorod for his help formulating and executing this project. His

knowledge, both theoretical and practical, was extremely helpful and he was able to point me to many

of the resources that I found to be invaluable. Additionally, I want to thank his Teaching Assistants for

FIN 551, Linda (Meng) Xu and Alan Yang, for their assistance. They were a sounding board for my ideas

and also showed me some great sources.

Page | 1
Contents
Abstract ......................................................................................................................................................... 1
Acknowledgements....................................................................................................................................... 1
Background ................................................................................................................................................... 3
Model Formulation ....................................................................................................................................... 7
Calibration ................................................................................................................................................... 11
Results ......................................................................................................................................................... 19
Analysis ....................................................................................................................................................... 25
References .................................................................................................................................................. 28

Page | 2
Background
Interest-rate models arise for many reasons, including economic forecasting, risk analysis, and

trading. As it is commonly believed that interest-rates are at least somewhat (if not entirely) random,

stochastic processes provide a natural framework upon which to build these models. The first models

were broadly known as Short-Rate Models because they modeled only a single interest, namely the

short-rate or the interest-rate applied to smallest available maturity. More intricate formulations of

these models are still used today, often with multiple stochastic variables acting as factors to underlie

the movement of the short-rate. Examples of these models include the famous Vasicek family of models

and Cox, Ingersoll and Ross (CIR) models. A strong limitation of these models is that you must make a

series of assumptions regarding the term structure of rates in order to extrapolate a curve across longer

maturities. This knowledge of the term structure is extremely important because a vast quantity of

derivative products rely on multiple interest-rates at multiple maturities. The desire to integrate more

market data into interest-rate models and the pressure to capture the evolution of the entire term

structure led to Market Models. Additionally, Market Models have a large degree of practical appeal for

trading desks and other practitioners because they allow for rigorous derivation of widely-used price

quoting formulas for Caps and Swaptions (which I will discuss below). These pricing formulas are Black’s

formula for Caps and Black’s formula for Swaptions, respectively. The formulas were developed by

applying the logic of the Black-Scholes-Merton Option pricing formula to interest-rate markets. 1 Before

Market Models, these formulas had no strong underlying theoretical framework. Market Models

effectively met the needs of market participants by satisfying their desire for a theoretical justification of

their pricing conventions and integrating a wider picture of the interest-rate market into their models.

1
(Brigo and Mercurio) Pg. 195

Page | 3
 While interest-rate models have many uses, the focus of this paper is their relevance to asset

pricing. The interest-rate derivative market is the largest security market in the world (by Notional

amount). As of 2011, the total, global OTC derivatives market was $708 trillion in notional (or contract)

amount, according the Bank of International Settlements (Figure 1). 2

Figure 1: 2011 Bank of International Settlements OTC Derivatives Summary

Of this market, interest-rate derivatives are an overwhelming majority, with a notional amount

outstanding of $554 trillion, or half of one quadrillion dollars (Figure 2).

Figure 2: Bank of International Settlements Interest-Rate Derivatives Summary

2
(Bank of International Settlements: Monetary and Economic Department) Pg. 1

Page | 4
 This paper examines the pricing of two of the most common and liquid interest-rate derivative

products, Caps and Swaptions. Both Caps and Swaptions are options on interest-rates. Conceptually, a

Cap is an insurance agreement that “caps” the floating interest-rate that the Cap buyer has elsewhere

agreed to pay at a series of fixed times in the future. It works by the seller of the Cap agreeing to pay the

buyer the difference between the strike rate of the Cap (fixed at its purchase) and the floating-rate, if

the floating-rate ever exceeds the strike. Therefore, the buyer is guaranteed to never have to pay more

than the strike rate in any given period. For a Cap, all periods are considered independent and exercising

(receiving payment from the seller) in one period does not impact the ability to exercise in another. For

example, if Company XYZ issues floating-rate bonds which pay LIBOR + 3% coupons quarterly, but they

never want to pay more than a total of 5%, they could buy a Cap with a strike rate of 2% from

Investment Bank ABC to achieve this protection. If LIBOR moved to 3% after the Cap was purchased, ABC

would pay XYZ 1%. Thus the net cash flow for XYZ would be – (3% + 3%) + 1% = -5%. If in the next

quarter LIBOR moved back below 2%, ABC would not owe anything that period.

In USD (U.S. Dollar) markets, Caps typically have three month payment periods. In the EUR

(Euro) market, payment periods are usually six months. This paper will focus on USD LIBOR markets.

Finally, Caps have two additional time parameters beyond payment frequency. These parameters are

Expiry and Term. Expiry is most often one period (three months) and is the length of time between the

purchase of the Cap and the actual beginning of the Caps’ protection. The Term (or Maturity) is the

length of the protection that the Cap offers. For example, a five year Cap purchased today will begin its

protection in three months, make its first payment (if necessary) in six months, and make its final

payment (again if necessary) five years from today. Each individual period in a Cap is referred to as a

Caplet. A final way of thinking about a Cap is as a basket of options on a series of Forward-Rate

Agreements (FRAs) with sequential Expiries and Maturities. This last intuition will be important for

drawing a parallel, but distinction between Caps and Swaptions.

Page | 5
 Swaptions, as their name implies, are options on interest-rate Swaps. Swaps, in turn, are

agreements to trade a floating-rate, to be determined in the future, for a fixed-rate determined at

contract initiation. This kind of contract allows someone who has a floating-rate obligation to fix their

risk today, by converting it to a fixed-rate. A feature of Swaps is that they are worth $0 at initiation. This

is achieved by adjusting the fixed-rate to make the expected payoff of the Swap zero. Swaptions, as

options on Swaps, have an inherent premium, because they allow the purchaser to enter into a Swap at

a later date, according to their own interests. Similarly, Caps also always have a positive value. The

Expiry of a Swaption (assuming European exercise) is the time at which the Swaption will be exercised,

therefore entering the purchaser into a Swap, or it expires. European exercise means that the Swaption

can only be exercised at Expiry, and not at any other time (before or after). The Tenor of a Swaption is

the length of the underlying Swap that would be created if the Swaption is exercised. The Tenor is

measured in time after the Expiry date. If, for example, Company FGH purchased a 1x5 (Expiry x Tenor)

Swaption today with strike rate 5%, in one year they would have the right to enter into a five year Swap

at the 5% fixed-rate set today. In one year, this Swap may be in-, at-, or out-of -the-money, determining

whether or not the Swaption will be exercised. If one year later a comparable Swap has a rate of 7%, the

Swaption would be exercised because it would have a positive value. On the other hand, if the Swap

rate in one year is 4%, the Swaption would not be exercised because a better deal can be found in the

market. Like a Cap is basket of options on FRAs, Swaptions can be thought of as a single option on a

basket of FRAs.

This last point is the key reason that Swaptions are harder to price than Caps. As Caps are a

series of independent options, the correlations of payoffs are not relevant and the value can simply be

calculated as the sum of the value of each option. Swaptions are a single option whose value is

correlated to all of the relevant forward rates, requiring that their joint movements be considered when

pricing. The issues of correlation and pricing will be discussed in more specific terms later.

Page | 6
Model Formulation

The cornerstone of the Lognormal Forward-LIBOR Model is the following Stochastic Differential

Equation (SDE) (Equation 1):

𝑑𝐹𝑘 (𝑡) = 𝜎𝑘 (𝑡)𝐹𝑘 (𝑡)𝑑𝑍𝑘 (𝑡) .

Equation 1: Dynamics of the 𝐹𝑘 Forward Rate under 𝑇𝑘 forward Measure

This equation shows that a given forward rate,𝐹𝑘 , is a martingale under the measure 𝑇𝑘 which is

associated with the numeraire of the price of a zero coupon bond maturing at time 𝑇𝑘 . It can be shown

using Ito’s Lemma and Ito’s Isometry that this SDE results in lognormally distributed forward rates. This

property is extremely useful in that normal distributions are very nice to deal with. Additionally, this

distribution function of forward rates is precisely what allows for the recovery of Black’s formula for

Caplets from the LFM. While this equation is very convenient, it relies on the appropriate choice of

measure. Therefore, 𝐹𝑖 for i ≠k, does not have driftless SDE, as a change of measure maintains the

diffusion term, but introduces a new drift. Brigo and Mercurio describe the procedure for modifying the

drift of a process under a change of measure in Chapter 2. The process relies on the calculation of the

quadratic covariation of the original stochastic process and the process of the quotient of old and new

numeraires. The equations below define the full dynamics of all forward-rates 𝐹𝑘 in the LFM under a

given 𝑇𝑖 forward measure (Equation 2).

𝑘
𝜌𝑘,𝑗 𝜏𝑗 𝜎𝑗 (𝑡)𝐹𝑗 (𝑡)
𝑖 < 𝑘, 𝑡 ≤ 𝑇𝑖 : 𝑑𝐹𝑘 (𝑡) = 𝜎𝑘 (𝑡)𝐹𝑘 (𝑡) � 𝑑𝑡 + 𝜎𝑘 (𝑡)𝐹𝑘 (𝑡)𝑑𝑍𝑘 (𝑡)
1 + 𝜏𝑗 𝐹𝑗 (𝑡)
𝑗=𝑖+1
𝑖 = 𝑘, 𝑡 ≤ 𝑇𝑘−1 : 𝑑𝐹𝑘 (𝑡) = 𝜎𝑘 (𝑡)𝐹𝑘 (𝑡)𝑑𝑍𝑘 (𝑡)
𝑖
𝜌𝑘,𝑗 𝜏𝑗 𝜎𝑗 (𝑡)𝐹𝑗 (𝑡)
𝑖 > 𝑘, 𝑡 ≤ 𝑇𝑘−1 : 𝑑𝐹𝑘 (𝑡) = −𝜎𝑘 (𝑡)𝐹𝑘 (𝑡) � 𝑑𝑡 + 𝜎𝑘 (𝑡)𝐹𝑘 (𝑡)𝑑𝑍𝑘 (𝑡)
1 + 𝜏𝑗 𝐹𝑗 (𝑡)
𝑗=𝑘+1

3
Equation 2: LFM Forward-Measure Dynamics of Forward-Rates

3
(Brigo and Mercurio) Pg. 16

Page | 7
In the above equations 𝜌𝑘,𝑗 is the correlation between two forward-rates, 𝜎𝑗 (𝑡) is the instantaneous

volatility of a given forward-rate, 𝜏𝑗 is time between adjacent forward-rate maturities, 𝑍𝑘 is a Brownian

motion under the 𝑇𝑖 forward measure.

Now that the forward-rate dynamics have been specified, one can begin pricing payoffs using

this model. The first security whose price will be considered is a Cap. Generally, the price of a Cap can be

calculated with the following equation (Equation 3).

𝐶𝑎𝑝 𝑃𝑟𝑖𝑐𝑒 = � 𝑁𝜏𝑖 𝐷(0, 𝑇𝑖 )(𝐹(𝑇𝑖−1 , 𝑇𝑖−1 , 𝑇𝑖 ) − 𝐾)+

𝑖=𝛼+1

Equation 3: General Cap Pricing Formula

Here the ”+” denotes the max(0,∗) operator, where * is the argument inside parentheses. To explain

the other notation, 𝐷(0, 𝑇𝑖 ) is the stochastic discount factor for the time period beginning at time zero

(today) and ending at 𝑇𝑖 ; tau is the year fraction length of the period between 𝑇𝑖−1 and 𝑇𝑖 ; 𝐾 is the strike

rate of the Cap; 𝛼 is the number of periods in the expiry of the Cap, minus one (if the Cap has a three

month expiry, then 𝛼 is zero); 𝛽 is the number of periods between the expiry and the maturity of the

Cap (if the Cap has a three month expiry and a two year maturity, then 𝛽 would be seven); and 𝑁 is the

notional amount of the Cap. The key fact to consider when pricing Caps is the independence of Caplets.

This feature allows all Caplets to be modeled separately, and then summed. Modeling Caplets

independently allows for a different choice of measure for each Caplet. To elaborate, the price of each

Caplet only depends on a single forward-rate. Therefore, for each Caplet, choose the measure 𝑇𝑘 that

allows the relevant forward-rate 𝐹𝑘 to be lognormally distributed. This lognormal distribution allows for

a closed-form solution for the price of each Caplet, and thus for the Cap by summing over the Caplets.

This closed-form solution for Caplet prices is exactly Black’s formula (Equation 4). 4

𝐶𝑎𝑝𝑙𝑒𝑡 𝑃𝑟𝑖𝑐𝑒 𝐿𝐹𝑀 (0, 𝑇𝑖−1 , 𝑇𝑖 , 𝐾) = 𝐶𝑎𝑝𝑙𝑒𝑡 𝑃𝑟𝑖𝑐𝑒 𝐵𝑙𝑎𝑐𝑘 (0, 𝑇𝑖−1 , 𝑇𝑖 , 𝐾)

4
(Brigo and Mercurio) Pg. 222

Page | 8
 = 𝑁 𝑃(0, 𝑇𝑖 )𝜏𝑖 𝐵𝐿(𝐾, 𝐹𝑖 (0), 𝑣𝑖 )

𝐵𝐿(𝐾, 𝐹𝑖 (0), 𝑣𝑖 ) = 𝐸 𝑖 [(𝐹𝑖 (𝑇𝑖−1 ) − 𝐾 )+ ]

= 𝐹𝑖 (0)Φ�𝑑1 (𝐾, 𝐹𝑖 (0), 𝑣𝑖 )� − 𝐾Φ�𝑑2 (𝐾, 𝐹𝑖 (0), 𝑣𝑖 )�

𝐹 𝑣2
ln �𝐾 � + 2
𝑑1 (𝐾, 𝐹𝑖 (0), 𝑣𝑖 ) =
𝑣

𝐹 𝑣2
ln �𝐾 � − 2
𝑑2 (𝐾, 𝐹𝑖 (0), 𝑣𝑖 ) =
𝑣

𝑣𝑖2 = 𝑇𝑖−1 𝑣𝑇2𝑖−1 −𝐶𝑎𝑝𝑙𝑒𝑡

𝑇𝑖−1
1
𝑣𝑇2𝑖−1 −𝐶𝑎𝑝𝑙𝑒𝑡 ∶= � 𝜎𝑖 (𝑡)2 𝑑𝑡
𝑇𝑖−1 0

Equation 4: Black’s Formula for Caplets

Unlike Caps, Swaptions are very difficult to price under the LFM. As mentioned above, the heart

of the issue is the importance of the correlation of forward-rates to pricing of Swaptions. This difficulty

manifests itself in the general formula for pricing Swaptions (Equation 5). 5

𝛽 +

𝑆𝑤𝑎𝑝𝑡𝑖𝑜𝑛 𝑃𝑟𝑖𝑐𝑒 = 𝑁 𝐷(0, 𝑇𝛼 ) � � 𝜏𝑖 𝑃(𝑇𝛼 , 𝑇𝑖 )(𝐹(𝑇𝛼 , 𝑇𝑖−1 , 𝑇𝑖 ) − 𝐾)�

𝑖=𝛼+1

Equation 5: General Swaption Pricing Formula

The notation in the above formula is the same as in the Cap formula (Equation 3) and the 𝑃(𝑇𝛼 , 𝑇𝑖 ) term

refers to the price of a zero coupon bond at time 𝑇𝛼 which matures at time 𝑇𝑖 . As the max function is

outside of the summation, this price cannot be subdivided in the way that Caps were. Therefore, the

entire summation must be evaluated together and the correlations of forward-rates in each iteration

must be considered.

5
(Brigo and Mercurio) Pg. 19

Page | 9
 There is no exact, analytical solution for Swaption prices under the LFM. There are two options

for pricing Swaptions in the LFM model, simulation and approximation. Simulation can be initially

attractive, but the process can be slow and influenced by Monte Carlo error. Especially in a trading

environment, speed is highly valued, even at the expense of small variation from true value. A true value

late, is worth infinitely less than a close value when needed. Approximation thus becomes the typically

preferred method for pricing Swaptions. Riccardo Rebonato, a patriarch of fixed-income derivative

mathematics and quantitative finance, developed a high-quality, closed-form, and theoretically

appealing approximation for pricing Swaptions in the LFM. This is a good time to mention a market

quoting convention that will be important throughout the rest of this paper. When one refers to the

“price” of a Cap or Swaption, what is actually quoted is the Black implied volatility. This implied volatility

is determined by inverting Black’s formula for Caps and Swaptions, respectively, with volatility as the

unknown and the price a trader is willing to buy or sell a security as an input, to infer a volatility. This is

the market convention for a variety of reasons, most significantly the ubiquity of Black’s formulas.

Rebonato’s approximation for the Swaption price in the LFM actually approximates the Black Swaption

volatility, not the dollar price of the Swaption. This volatility can then be compared to market quotes.

Rebonato’s Formula, as the approximation is known, is as follows (Equation 6). 6

𝛽
𝐿𝐹𝑀 2
𝑤𝑖 (0)𝑤𝑗 (0)𝐹𝑖 (0)𝐹𝑗 (0)𝜌𝑖,𝑗 𝑇𝛼
(𝑣𝛼,𝛽 ) = � � 𝜎𝑖 (𝑡)𝜎𝑗 (𝑡)𝑑𝑡
𝑆𝛼,𝛽 (0)2 0
𝑖,𝑗= 𝛼+1

𝜏𝑖 𝐹𝑃(𝑡, 𝑇𝛼 , 𝑇𝑖 )
𝑤𝑖 (𝑡) =
∑𝛽𝑘=𝛼+1 𝜏𝑘 𝐹𝑃(𝑡, 𝑇𝛼 , 𝑇𝑘 )

𝑃(𝑡, 𝑆)
𝐹𝑃(𝑡, 𝑇, 𝑆) =
𝑃(𝑡, 𝑇)

Equation 6: Rebonato’s Formula

6
(Brigo and Mercurio) Pg. 283

Page | 10
 In this formula, 𝑆𝛼,𝛽 (0) is the swap rate for a swap beginning at period 𝛼 and maturing at period

𝛽. At this point, the central aspect of the LFM formulation that remains is the specification of the

instantaneous volatility function,𝜎, and the instantaneous correlation matrix, 𝜌. The main instantaneous

volatility used in this paper is Brigo and Mercurio’s Formulation 7, after Rebonato’s time-homogeneous

formulation (Equation 7). 78

𝜎𝑖 (𝑡) = Φ𝑖 𝜓(𝑇𝑖−1 − 𝑡; 𝑎, 𝑏, 𝑐, 𝑑) ∶= Φ𝑖 �[𝑎(𝑇𝑖−1 − 𝑡) + 𝑑]𝑒 −𝑏(𝑇𝑖−1−𝑡) + 𝑐�

Equation 7: Formulation 7 Instantaneous Volatility

Above, Φ𝑖 , are time-invariant constants and a, b, c, and d are fit parameters. For instantaneous

correlation, another time-homogeneous specification is used. This form was also formulated by

Rebonato (Equation 8). 9

𝜌𝑖,𝑗 = 𝑒 −𝛽�𝑇𝑖 − 𝑇𝑗�

Equation 8: Instantaneous Correlation Matrix

Here 𝛽 is a parameter which can be manipulated for model calibration. After the formulation is finished,

the next step is calibration of the model to market quotes.

Calibration

The first step for calibrating a model to market data is deciding what fits to prioritize. This

decision impacts the formulation of the model, the data needed for calibration, and the steps taken to

optimize the solution. An additional tradeoff when fitting parametric models is between the optimality

of fit to the data and the economic and theoretical intuition behind the derived values. To illustrate this

point, the calibration procedure may find that a 𝛽 of 50 produces the best fit to market data, given other

parameters. This value is immediately suspect as it would imply that forward-rates as little as one period

7
(Brigo and Mercurio) Pg. 212
8
(Rebonato) Pg. 167
9
(Rebonato) Pg. 177

Page | 11
apart in maturity have essentially zero correlation, which makes little economic sense and fails to satisfy

the market-based intuition of many practitioners.

Rebonato describes two possible goals for calibration under the LFM. 10 The first goal is to

capture the current Caplet market and calibrate the model to evolve with a similar term structure into

the future. The second goal would be to again fit Caplet quotes, but calibrate the model’s parameters to

capture the Swaption market as well, at the cost of desirable evolution over time. Both of these forms of

calibration are considered in this paper. Before the details of calibration are discussed, another

dimension of market quoting must be considered.

This market convention is to quote a single, constant, and annualized implied volatility for a Cap,

𝑣𝒯𝑗−𝐶𝑎𝑝 , even though each underlying Caplet has its own, distinct implied volatility. This chart from

Figure 3: Bloomberg SWPM Showing a Four Year Cap Broken Down Into Caplets

10
(Rebonato) Pg. 226

Page | 12
Bloomberg (Figure 3) illustrates the cash flow and implied volatility structure of a Cap broken down into

Caplets.

The convention is convenient because it allows for a single number to convey the price of an entire Cap,

with maturity 𝒯𝑗 (Equation 9). The problem with this method is an apparent contradiction for the implied

volatility of Caplets. Specifically, a five year Cap and a three year Cap (given the same strike) would have

identical Caplets for the first three years, but they would have different implied volatilities. However,

this apparent contradiction is not actually a problem. When a trader constructs a quote, she solves a

non-linear equation for the constant implied volatility 𝑣𝒯𝑗−𝐶𝑎𝑝 that, when assigned to all Caplets, prices

the Cap the same as the sum of the Caplets with individual volatilities 𝑣𝑇𝑖−1 −𝐶𝑎𝑝𝑙𝑒𝑡 (Equation 10). 11

𝐶𝑎𝑝 𝑃𝑟𝑖𝑐𝑒 𝑀𝐾𝑇 (0, 𝒯𝑗 , 𝐾) = � 𝑁 𝑃(0, 𝑇𝑖 )𝜏𝑖 𝐵𝐿 �𝐾, 𝐹𝑖 (0), �𝑇𝑖−1 𝑣𝒯𝑗−𝐶𝑎𝑝 �
𝑖=1

Equation 9: Cap Price for a Given Market Quoted Implied Volatility

� 𝑁 𝑃(0, 𝑇𝑖 )𝜏𝑖 𝐵𝐿 �𝐾, 𝐹𝑖 (0), �𝑇𝑖−1 𝑣𝒯𝑗−𝐶𝑎𝑝 �

𝑖=1

= � 𝑁 𝑃(0, 𝑇𝑖 )𝜏𝑖 𝐵𝐿�𝐾, 𝐹𝑖 (0), �𝑇𝑖−1 𝑣𝑇𝑖−1−𝐶𝑎𝑝𝑙𝑒𝑡 �

𝑖=1

Equation 10: Equation for Inferring Cap Implied Volatility from Caplet Implied Volatilities
The volatilities in Equation 10 are scaled by �𝑇𝑖−1 because their squares are annualized for quoting and

the original value must be recovered for pricing.

11
(Brigo and Mercurio) Pg. 225-226

Page | 13
 Therefore, before a model can be fit to market quote data, that data must be modified to

recover the Caplet implied volatilities necessary for calibration for the Cap volatilities outlined above.

This process involves a bootstrapping procedure across increasing Cap maturities and Cap strikes.

Another difficulty that presents itself in market data is the phenomenon known as skew, or

volatility smile. Skew is the presence of a non-flat implied volatility across strike rates, namely that as

strike rates move further into or out-of-the-money the implied volatility increases. On an intuitive level,

this structure points to the fat-tailed underlying distribution of rate movements. In essence, a normal

distribution applied to larger moves requires a higher standard deviation, according to the market. Thus,

a normal distribution that uses the standard deviation of small price movements underprices the

likelihood of large jumps in interest-rates (there is a rich literature of stochastic volatility [notably SABR]

models and jump-diffusion models that capture this smile in a more systematic fashion). While this

paper does not use these more complex volatility schemes, the calibration procedure incorporates the

volatility smile present in market quotes.

The bootstrapping of Caplet implied volatilities from market quotes requires two steps. The first

step is to bootstrap all maturities for each quoted strike rate, and then inter/extrapolating a smile. The

second step is the adjustment to at-the-money (ATM) quotes. These quotes are treated separately

because they have non-standard strike rates that change at each maturity, thus additional knowledge of

the volatility smile is needed to be able to successfully bootstrap ATM Caplets. The reason for taking the

time to use ATM Caps is that they tend to be the most quoted and most liquid Caps, therefore they have

the freshest and most accurate data for the market.

Page | 14
 The market, when accessed through a Bloomberg Terminal (the standard market data service)

gives the following information for Caps at a given point in time (Figure 4).

Figure 4: Bloomberg VCUB Cap Market

The final obstacle to Caplet bootstrapping is an assumption about the term structure of implied volatility

of Caplets. As the Cap quote density is typically in one year (or greater) intervals, there is not sufficient

information to uniquely determine quarterly Caplet implied volatilities from market information alone. A

common, and typically robust, solution is to assume that volatility is piece-wise constant in-between

quoted Cap maturities. 12 This is the assumption which will be used in this paper.

To implement the first part of bootstrapping stated above, the following procedure is used 13:

1. Iterate over all available strike rates K.

12
(Levin) Pg. 8
13
(Levin) Pg. 8

Page | 15
 2. For each K, bootstrap Caplet implied volatilities, starting with the first Cap maturity, using the

assumption of piece-wise constant intervals and Equation 11 below.

3. After a Caplet volatility is found for all Caplet maturities at all quoted strikes, interpolate the

volatility smile across strikes, this time iterating over all Caplet maturities.

4. Now a mesh of Caplet implied volatilities, consistent with market quotes, has been built for all

strikes and maturities.

𝐶𝑎𝑝 𝑃𝑟𝑖𝑐𝑒 𝑀𝐾𝑇 �0, 𝒯𝑗+1 , 𝐾� − 𝐶𝑎𝑝 𝑃𝑟𝑖𝑐𝑒 𝑀𝐾𝑇 �0, 𝒯𝑗 , 𝐾�

𝒯𝑗+1 − 𝜏𝑗

= � 𝐶𝑎𝑝𝑙𝑒𝑡 𝑃𝑟𝑖𝑐𝑒(0, 𝑇𝑖 , 𝑇𝑖+1 , 𝐾)

𝑇𝑖 = 𝒯𝑗 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝜏𝑖

Equation 11: Equation for Bootstrapping Caplet Implied Volatilities

Equation 11 utilizes the additivity of Caplets to equate the difference between the prices of two adjacent

Caps to the sum of the intermediate Caplets.

The second part of bootstrapping, using ATM data, uses these steps 14:

1. Iterate over ATM Cap maturities.

2. For each maturity, find the term structure in the mesh of Caplet volatilities previously derived

that corresponds to the strike rate of the particular ATM Cap.

3. Use this synthetic term structure to price a Cap with a maturity one quoting interval less than

that of the ATM Cap being bootstrapped.

4. Finally, use Equation 11 to calculate an implied Caplet volatility, again piece-wise constant, which

correctly prices the difference between the ATM quoted Cap and the synthetic Cap priced in

step 3.

14
(Levin) Pg. 9

Page | 16
The results of part one are displayed below in Figure 5.

Figure 5: Market Cap Implied Volatility Surface

The volatility skew is the curvature in the strike dimension of the graph.

Now that the necessary market data has been extracted from quotes, the model can now be

calibrated. The procedure general for calibration is the following: using the model parameters as input

variables, minimize the objective function equal to the sum of squared errors between model and

market prices. Under this problem specification, a variety of optimization schemes can be used to

determine the model parameters. In this project, MATLAB’s fmincon constrained optimization with an

active-set algorithm was used.

Under the specific instantaneous volatility and correlation specifications made in this paper thus

far, the following model parameters were used in the optimization: a, b, c, d, and 𝛽. Additionally, the

primary focus of the project was the second of the two calibration goals stated at the beginning of this

section. Therefore, the optimization was done to minimize error between model and market Swaption

prices. A convenient feature of the instantaneous volatility Formulation 7 is its incorporation of the

Page | 17
time-invariant constants Φ𝑖 . These constants allow the model to exactly fit the initial term structure of

Caps by construction, regardless of the choice of model parameters. The tradeoff, as already

highlighted, is that the evolution of the term structure over time tends to be less desirable.

A second formulation of the LFM was also implemented, using the first calibration goal to

provide a comparison with the results of the first formulation. This formulation adds a time-variant term

(Equation 12). 15

3
𝑡𝜋𝑖
𝜖(𝑡) = �� 𝜖𝑖 sin � + 𝜖𝑖+1 � � 𝑒 −𝜀7𝑡
𝑀𝑎𝑡
𝑖=1

Equation 12: Time-Variant Instantaneous Volatility Component

In this equation, 𝑀𝑎𝑡, is the maturity in years of the longer tenor Cap.

Then the instantaneous volatility has the form of Equation 13.

𝜎𝑖 (𝑡) = Φ𝑖 𝜓(𝑇𝑖−1 − 𝑡; 𝑎, 𝑏, 𝑐, 𝑑)𝜖(𝑡) ∶

3
−𝑏(𝑇𝑖−1 −𝑡)
𝑡𝜋𝑖
= Φ𝑖 �[𝑎(𝑇𝑖−1 − 𝑡) + 𝑑]𝑒 + 𝑐� �� 𝜖𝑖 sin � + 𝜖𝑖+1 � � 𝑒 −𝜀7 𝑡
𝑀𝑎𝑡
𝑖=1

Equation 13: Rebonato’s 6.21a Instantaneous Volatility Formulation

This formulation adds five additional, time-varying parameters (𝜖1 , 𝜖2 , 𝜖3 , 𝜖4 , 𝜖7 ) for optimization of the

model. The choice of sinusoidal functions for the time-variant term was motivated by Fourier analysis.

Additionally, the number of frequencies is limited to balance having additional fit flexibility and having a

too loosely specified model which will incorporate too much market noise. 16 This formulation is then

optimized with respect to Cap prices alone, without considering Swaptions. A close reader would

question the purpose of this optimization, as the model is already guaranteed to exactly fit current

market data. The motivation is to optimize the evolution of the term structure. Under the previous

15
(Rebonato) Pg. 233
16
(Rebonato) Pg. 169

Page | 18
formulation, theΦ𝑖 constants are made whatever is necessary to fit Cap prices after the model is fit to

Swaptions. As these Φ𝑖 largely determine the model’s transition over time, the resulting term structure

can be highly erratic. In the present formulation however, the majority of the calibration “work” is done

by the time-homogenous and time-variant portions of the instantaneous volatility formula. Therefore

the Φ𝑖 ′𝑠 that arise from this formulation tend to be close to one, with much smaller variation. These

values result in a much more behaved evolution. 17 The second formulation utilized the same

instantaneous correlation formula, but fixed 𝛽 at 0.1 (see the Results for a justification of this choice).

The following section will give the calibration results of these two model formulations.

Results

For the code used to arrive at the following results, please consult the MATLAB Code Appendix.

As predicted, both formulations were able to precisely capture the current market Cap price, as

measured by Black implied volatility. Figure 6 shows the market and model Cap implied volatilities for the

Formulation 7. Figure 8 shows the same results for Rebonato’s 6.21a formulation.

17
(Rebonato) Pg. 234

Page | 19
 Figure 6: Formulation 7 Cap Implied Volatility Term Structure

Figure 8: Rebonato’s 6.21a Cap Implied Volatility Term Structure

The optimal model parameters for each formulation are summarized in Table 1.

Page | 20
 Parameter Formulation 7 Rebonato 6.21a

a 12.2690 -20

b 1.7798 6.3973

c 0.8290 1.3830

d 7.7659 0.6914

𝜷 0.108 0.1 (Set)

𝝐𝟏 - -0.3534

𝝐𝟐 - 2.1037

𝝐𝟑 - 1.4645

𝝐𝟒 - 3.8375

𝝐𝟕 - 0.1068

Table 1: Summary of Optimal Model Parameters

The above optimal values were found using the following constraints, shown in Table 2. 18

Formulation 7 Rebonato 6.21a

𝑑+𝑐 >0 𝑑+𝑐 >0

𝑐, 𝑑 > 0 𝑐, 𝑑 > 0

−20 < 𝑎, 𝑏, 𝑐, 𝑑 < 20 −20 < 𝑎, 𝑏, 𝑐, 𝑑 < 20

0.00005 < 𝛽 < 0.108 -

- −∞ < 𝜖1 , 𝜖2 , 𝜖3 , 𝜖4 < ∞

- 0.00000000001 < 𝜖7 < ∞

Table 2: Model Optimization Constraints

18
(Rebonato) Pg. 168 for the first two of these constraints.

Page | 21
 The first two constraints come from Rebonato as recommended bounds for maintaining the

shape of the instantaneous volatility function. The others are made to keep the optimization algorithm

in the general region of desirable solutions. The objective functions are both non-linear, fairly “ugly”

functions to optimize, therefore it is necessary to provide some guidance to the algorithm. The limits of

𝛽 are implicitly recommended by Rebonato who states that the correlation of forward-rates should be

kept reasonably high to maintain resemblance to the financial intuition underlying the model. 19 As a

general rule, 𝛽 has historically been set ≈ 0.1, thus I chose that value as the constant under my second

formulation. 20 The resulting correlation structures are very similar for both formulations, therefore

(Figure 9), the surface for Rebonato’s 6.21a, is shown to illustrate both.

Figure 9: Model Instantaneous Forward-Rate Correlation Surface (Rebonato’s 6.21a Formulation)

19
(Rebonato) Pg. 222
20
(Rebonato) Pg. 243

Page | 22
 For each formulation, the calibrated evolution of Caplet implied volatilities, as shown by their

Φ𝑖 ′𝑠 is depicted in (Figure 10) for Formulation 7 and 6.21a, respectively.

Figure 10: Model 𝚽 Values for Formulation 7 (Top) and Rebonato 6.21a (Bottom)

As expected, the Φ𝑖 values for formula 6.21a are both much closer to one (mean = 1.0002) and exhibit

less percentage variation (with all values within 20% of the mean). Formulation 7 resulted in Φ𝑖 ′𝑠 within

40% of the mean (0.1513).

The Formulation 7 of the model was also fit to Swaption data. For this calibration, I used the

three month expiry edge (across all quoted maturities) of the Swaption volatility surface. Figure 11 shows

the model and market implied volatility curve, along with a chart of the error.

Page | 23
 Figure 11: Formulation 7 Model Swaption Implied Volatility Term Structure

Finally, by relaxing the constraint on 𝛽 to an upper limit of 20, better Swaption calibration was achieved.

This result is summarized in Table 3 and Figure 12 below.

Parameter Formulation 7

a 20.000

b 0

c 0.0002

d 8.3989

𝜷 19.999

Table 3: Formulation 7 Relaxed Optimization Parameters

Page | 24
 Figure 12: Formulation 7 Model Swaption Implied Volatility Term Structure (Relaxed 𝛽Constraint)

Analysis

While the process of model calibration is very much a science, consisting of a formulaic process,

it is equally an art. In practice, models are not typically used to price products as liquid and relatively

simple as Swaptions and Caps. Rather, models aid in the analysis of the risk and valuation of exotic

derivatives and other one-off contracts for which market data is not available. Therefore, the decision of

how to calibrate a model, and what it should be calibrated to, is largely dependent on what the end

purpose of the model will be. For example, if the exotic derivative more closely resembles a Cap than a

Swaption, it would be more important for the model to fit the Cap market than the Swaption price data.

Other aspects of the art of calibration include the inconsistent quality of market data. For at-the-money

Cap and Swaptions, and those in popular maturities, price data is usually high quality because they tend

to be quoted often and by many traders. Therefore, judging which market points to consider is another

important step for refining model fit. The paper does not endeavor to incorporate these market

nuances, but they are worth noting for completeness.

Page | 25
 Another aspect of optimizing calibration is the difficulty of finding globally optimal parameters in

an acceptable value range. This paper used the fmincon MATLAB optimization framework with an

active-set algorithm. This method was recommended by a practitioner as typically producing the best

results. Something that deserves further exploration if this model were to be implemented for trading

purposes is the use of other optimization algorithms that may produce better results for these non-

linear, not “nice” objective functions. These optimizations also tend to be sensitive to the choice of seed

values. This project’s optimization used potential optimal parameter values listed by Brigo and Mercurio

as starting points for fmincon. 21

The Lognormal Forward-LIBOR Model, and the LIBOR Market Model generally, appears

promising based on the results of this implementation. The unique combination of theoretical rigor and

correspondence to market practices makes the LFM an attractive framework for pricing interest-rate

derivative products. The model performed very well with respect to Caps and satisfactory results were

achieved for Swaptions. When the model was calibrated strictly to Caps, the resulting Φ𝑖 values being

close to one indicated that the model was capturing the term structure of Cap implied volatilities well

and required minimal scaling. The calibration process is much more complex for Swaption than it is for

Caps, and is expected to be less accurate. The Rebonato approximation discussed above was used to

calculate Swaption implied volatilities that were compared with the market. This approximation has

been proven to be fairly accurate, but it still presents a source of error. Additionally, with only five

parameters, exact calibration to the Swaption market is impossible. That said, the model showed

promising results which indicated that it could be refined using the considerations described above to

be viable for pricing. The relaxation of the 𝛽 constraint beyond economically sensible bounds illustrates

the tension between the science and art of calibration; it is less theoretically sound, yet produces a

better fit to the data. Additionally, model extensions, namely stochastic volatility, and the addition of

21
(Brigo and Mercurio) Pg. 320

Page | 26
new market developments, such as OIS (Overnight Indexed Swap) discounting and CVA (Credit Value

Adjustment), show important steps for improving the fit of the LFM to market data going forward.

Page | 27
References

Bank of International Settlements: Monetary and Economic Department. OTC derivatives market activity
in the first half of 2011. Basel, Switzerland: Bank of International Settlements, 2011.

Belaygorod, Anatoliy. "FIN 552 Lecture Notes and Course Materials." 2011.

Brigo, Damiano and Fabio Mercurio. Interest Rate Models - Theory and Practice. 2nd. Berlin: Springer
Finance, 2006.

Levin, Kirill. "Bloomberg Volatility Cube." n.d.

Rebonato, Riccardo. Modern Pricing of Interest-Rate Derivatives. Princeton, New Jersey: Princeton
University Press, 2002.

Page | 28